The phenology mismatch hypothesis: are declines of migrant birds linked to uneven global climate change?

Summary

1. Migrant bird populations are declining and have been linked to anthropogenic climate change. The phenology mismatch hypothesis predicts that migrant birds, which experience a greater rate of warming in their breeding grounds compared to their wintering grounds, are more likely to be in decline, because their migration will occur later and they may then miss the early stages of the breeding season. Population trends will also be negatively correlated with distance, because the chances of phenology mismatch increase with number of staging sites.

2. Population trends from the Palaearctic (1990–2000) and Nearctic (1980–2006) were collated for 193 spatially separate migrant bird populations, along with temperature trends for the wintering and breeding areas. An index of phenology mismatch was calculated as the difference between wintering and breeding temperature trends.

3. In the Nearctic, phenology mismatch was correlated with population declines as predicted, but in the Palaearctic, distance was more important. This suggests that differential global climate change may be responsible for contributing to some migrant species’ declines, but its effects may be more important in the Nearctic.

4. Differences in geography and so average migration distance, migrant species composition and history of anthropogenic change in the two areas may account for the differences in the strength of the importance of phenology mismatch on migrant declines in the Nearctic and Palaearctic.

Introduction

Global warming has been shown to have major effects on bird populations associated with changes in phenology of life-history events, particularly breeding (Crick et al. 1997; Winkel & Hudde 1997; Visser, Holleman & Gienapp 2006) and migration (Tryjanowski, Kuzniak & Sparks 2002; Cotton 2003; Gordo et al. 2005; Jonzen et al. 2006). Warming affects the phenology of different species in different ways, which can negatively impact upon events, such as migration and breeding, which were previously synchronized with the phenology of resources. For example, the shift in resource peak date, due to warming temperatures, may differ from the shift in breeding and migration dates of birds, which rely upon these resources being abundant during these periods (Cresswell & McCleery 2003; Visser, Both & Lambrechts 2004). Migrant birds are probably particularly vulnerable to such mismatches in resource availability as climate change does not act equally over the globe (IPCC 2007) and so migration may be mistimed with resource availability at any number of staging and/or breeding areas. Furthermore, changes in the timing of migration have been linked to climate change both on breeding (Tryjanowski et al. 2002) and wintering grounds (Gordo et al. 2005) as well as to climatic conditions en route (Ahola et al. 2004), all of which raise the chances of a mismatch occurring. Migrant birds have also been found to generally be in decline relative to non-migrants, with long-distance migrants declining the most (Sherry & Holmes 1996; Sanderson et al. 2006).

A mismatch between resource availability and arrival dates, on breeding grounds, has been observed in Central-European Pied flycatchers, Ficedula hypoleuca (Pallas). Arrival dates have shown no significant advance whereas food-peak date has advanced due to warming spring temperatures leading to a 90% decline in abundance for some populations, which arrive after the food peak (Both et al. 2006). Furthermore, of 100 European migrant bird species, those that have advanced their migration date have significantly more positive population trends than those that have not changed their migration strategy in line with mismatches in arrival and food-peak dates (Moller, Rubolini & Lehikoinen 2008). Such mismatches have also been shown to occur in American wood warbler species, where wintering and staging areas have not warmed as significantly as the breeding areas leading to arrival after the advancing resource peak (Strode 2003).

Here we test the generality of whether differential global climate change is correlated with population trends in migrant birds, testing two principle hypotheses. Firstly, the ‘phenology mismatch’ hypothesis (Fig. 1) states that a species, which experiences increased warming at the wintering grounds with respect to the breeding grounds, should on average show relatively positive population trends. This is because migrants depart earlier and so a greater proportion of the population arrive in time to exploit the optimum timing of spring resources. Conversely, a species which experiences a slower change in temperature in their wintering areas compared to the breeding areas is more likely to show a relatively negative population trend because they are likely to depart later, and so arrive after the optimum timing of spring resources.

Figure 1.  Phenology mismatch hypothesis. Left: The predicted directions of population trend with Phenology Index. Right: Possible climate change scenarios which would give rise to phenology indices of that value. Temperature is related to the timing of resources becoming available. Thus if there is a mismatch in the temperature trends, it is likely there will be a mismatch in resource availability and migration phenology leading to population consequences.

Secondly, the ‘distance’ hypothesis predicts that long-distance migrants are more likely to be in decline than short-distance migrants. In most cases, long-distance migrants will have to make more stops at staging sites than short-distance migrants, increasing the chances of phenology mismatch during migration, and also the chance that any one staging site will have been detrimentally affected by anthropogenic changes. As a consequence, they will arrive later at the breeding grounds, and/or suffer greater population reductions en route.

We tested these hypotheses using generalized linear models fitted to the population trends of 193 separate populations of migrants in the Palaearctic and Nearctic, with a measure of phenology mismatch and distance as potential predictors. Phenology mismatch was calculated as the difference between wintering and breeding temperature. We also consider the mass of the bird species as a key explanatory variable in all models because the maximum distance of each stage in the migration, and so the number of staging sites for a given distance, is likely to be related to mass (Weber & Houston 1997). Mass is also likely to influence a species’ ability to survive on or capitalize on stored energy reserves on arrival at the breeding ground (Drent 2006). Therefore, we might predict that larger species are more robust to phenology mismatch because they potentially have more options in choice of staging sites during migration, and more resilience in terms of energy stores on arrival at the breeding grounds. We also include continent as a key explanatory variable in all models because differences in geography between the two continents associated with migration routes and restrictions on breeding and wintering latitudes are likely to also affect any phenology mismatch.

Finally, we consider some other potential alternative variables that may be driving any relationships between population declines and phenology mismatch and distance: longitudinal span and latitude of breeding and wintering ranges, breeding and wintering habitat, arrival time and foraging guild (Walther et al. 2002; Rubolini et al. 2007). We tested latitude and longitude, as well as breeding and wintering habitat, because any effects of phenology mismatch or distance may be affected by latitudinal and longitudinal differences in temperature or habitat variation. Temperature variation within a year is much lower, and it is warmer in lower latitudes (arctic vs. temperate vs. tropical climates) so the effects of any temperature change on energy budgets and resource availability are likely to differ with breeding latitude (Sanz 1998; Visser & Both 2005). We also considered habitat because species wintering and/or breeding in different habitats may show different population changes regardless of their phenology indexes, because the anthropogenic changes may be different in different habitats (Sanderson et al. 2006). We tested arrival time because any apparent effects of phenology mismatch or distance may arise because the strength of effects of any temperature change on the phenology of resources needed by a bird population are likely to vary within a season (Visser et al. 2004; Jonzen, Hedenstrom & Lundberg 2007). Finally, we tested foraging guild because any apparent effects of phenology mismatch and distance may arise because of latitudinal and longitudinal variation in the distribution of different guilds of migrants, and the likelihood of differing relative effects of phenology changes on different trophic levels (Durant et al. 2007; Parmesan 2007).

We predict that if the phenology mismatch hypothesis is causing global population declines that:

• 1
   Populations with positive values of Phenology Index (acceleration of temperature change on the wintering ground relative to the breeding ground) will show positive trends. We should therefore observe significant positive parameter estimates for Phenology Index in a model of population trend with Phenology Index.
• 2
   Populations with longer migration distances (controlling for mass) will show negative trends. We should therefore observe significant negative parameter estimates for distance in a model of population trend with migration distance.
• 3
   Populations with longer migration distances (controlling for mass) will show stronger negative trends where there is a negative value of Phenology Index (i.e. there are potentially more staging grounds where peak resources may become too early to exploit relative to an earlier stage). We should therefore observe a significant interaction between distance and Phenology Index in the model.
• 4
   The strength of effects of all variables may be influenced by the differences in geography and species composition between the largely independent migratory systems of the Palaearctic and the Nearctic. We should therefore expect that there will be significant interactions between Phenology Index and also distance with continent in the model.

Materials and methods

Data collation and preparation – population trends

Any geographically distinct breeding and wintering migrant bird species with data on population trends from three or more countries or regions were considered for the analysis. Species with significant overlap of breeding and wintering areas were excluded, due to the possibility of including resident populations. Furthermore, reintroduced species were also excluded due to the possible inflation of trend and population size estimates.

Birdlife International’s (2004)‘Birds in Europe’ and the American Breeding Bird survey (Sauer, Hines & Fallon 2008) give regional population trends and abundance data for European and North American breeding birds, respectively. ‘Birds in Europe’ gave maximum and minimum estimates of population trends and abundances, per country, for European breeding birds over the period 1990–2000. These were given in a similar format to the IUCN (2001) red list criteria thresholds of percentage change per country over a 10-year period. The upper and lower estimates were averaged in order to gain a single value for the population trend and abundance within that country. Data of the lowest quality were omitted from the analysis to avoid the inclusion of qualitative data (Sanderson et al. 2006). Data of Russian origin were also omitted from the analysis, because of uncertainty in distinguishing between population trends of European and Asian subspecies that winter in different areas. The American Breeding Bird survey provides trend estimates for each species by physiographic province in North America over the period 1980–2006. Data of the lowest credibility class were excluded from the analysis to improve precision. The percentage change per year was converted to percentage change per decade, so units in Europe and America were directly comparable. This was achieved by assuming a geometric growth/decline model in which the percentage change per decade is the cumulative percentage change over a 10-year period. Relative abundance estimates by physiographic province were calculated as the product of the number of routes the birds were observed on and the average abundance of birds per route (Sauer & Droege 1990).

To increase the resolution of the data and the number of data points available for analysis, species were split into spatially independent subspecies or populations where possible. This method assumes that any spatially isolated subpopulation in both wintering and breeding grounds will respond equally independently to the effects of a phenology mismatch or migration distance. This same logic applies to us considering species as independent points, rather than considering the confounding effects of phylogeny (e.g. the comparative method, see Bennett & Owens 2002). It seems unlikely that evolutionary history differentially affects a taxon’s ability to withstand the population effects of a phenology mismatch outwith the effects of body size, habitat and foraging considerations that we consider anyway (see below). In any case, we explore the effects of this potential pseudoreplication by limiting our sample size to single points from species (see below). Subspecies-/population-specific breeding and wintering ranges were identified from ‘Birds of the Western Palearctic’ (BWP) (Cramp 1998) and ‘The Birds of North America Online’. This enabled regional abundance and trend estimates to be assigned to an individual subspecies or geographically distinct population. In total 193 spatially separate subspecies or populations were identified comprising 134 species (Appendix S1, Supporting Information). Subspecies/population specific population trends were then calculated as the average of the regional trends, identified for each subspecies, weighted by regional abundance or relative abundance in North America.

Data collation and preparation – temperature trends

Monthly average temperature data are available from the NOAA Satellite and Information Service (NOAA 2008). For European and African countries, only temperature stations with full coverage over the period of 1990–2000 were utilized to remove the chance that large positive or negative trends, arising from short time series data, are included in the calculation of overall temperature trends. In North and South America, temperature stations corresponding to each physiographic province or country, respectively, were identified for the period 1980–2006 and were filtered in a similar fashion. For each country/region, a linear model was fitted to the average monthly temperatures data, using statistical package r version 2.6.0. The temperature trend for each month was taken to be the resultant trend from the linear model fitted to the monthly average temperatures, separately for each country or region. Therefore, for each country, there was a temperature trend assigned to each month expressed as change in °C per decade. Some countries in Africa and South America, however, had limited data for the desired period and as such temperature trends for the closest country were used as a surrogate. Larger well-covered countries (Brazil, Mexico and Argentina) were split into regions (North, Central and South for Brazil and Argentina, and North and South regions for Mexico) to increase resolution.

Calculation of variables

Months occupied in the breeding and winter seasons were identified from species profiles from ‘BWP’, ‘The Birds of North America Online’ and DeGraaf & Rappole (1995). The overall breeding season temperature change, for each species, was calculated by averaging the monthly temperature trends. Firstly, trends were averaged across breeding season months, for example if the species occupies its breeding grounds from April to August, the temperature trends for the months April, May, June, July and August were averaged to give an average temperature trend for the season. This was done for each separate region occupied by that species during the breeding season. These regional averages were then averaged across regions to obtain an average for the occupied range of each species/subspecies. This was done using regional abundance (relative abundance in North America) as a weighting factor. The winter temperature trend was similarly calculated, except in the final averaging stage, area of suitable habitat type was used as a weighting factor for averaging trends across regions, in place of abundance data, which was not available for the winter regions (see Table 1, for a detailed description of the specific methods). Phenology Index was then calculated as the arithmetic difference between the winter temperature trend and the breeding temperature trend. Furthermore, using information regarding the breeding and wintering ranges of each subspecies, average migration distance was calculated. This was calculated as the distance between the average central point of the winter and breeding ranges calculated by weighted average, of the central points of each occupied region (see Table 1). Other potential important explanatory variables such as mass, habitat and foraging types were obtained from species profiles in ‘BWP’, ‘The Birds of North America Online’, DeGraaf & Rappole (1995) and DeGraaf, Tilghman & Anderson (1985) (see Table 1).

Table 1.   Description of calculated variables, including a brief description of the calculation method

Qualitative classifications were based on DeGraaf et al. (1985) for American breeding birds and BWP (Cramp 1998) for European birds.
Breed_Hab	Breeding habitat type listed as one of the following; 1: Grassland, 2: Savanna, 3: Alpine, 4: Desert, 5: Forest, 6: Freshwater, 7: Marine, 8: Brackish In the case of several habitat types, the major habitat type was chosen
Wint_Hab	Wintering habitat type using the same classification as Breed_Hab
Breed_Temp	Temperature trend of breeding grounds in °C per decade. Calculated by averaging the monthly trends (equal weight) across the breeding months for the relevant regions for each subspecies. The overall trend for each subspecies was then found by weighted average of the regional trends, across the breeding range, using abundance (relative abundance in North America) as a weighting factor
Wint_Temp	Temperature trend of wintering grounds in °C per decade. Calculated by averaging the monthly trends (equal weight) across the winter months for the relevant regions. The overall trend for each subspecies was found by weighted average of the regional trends across the wintering range. As abundance data for wintering regions was not available, the area of suitable habitat of each region (based on wintering habitat type) was used as a weighting factor. Habitat type coverage by country was obtained from the ‘Earthtrends’ website (World Resources Institute 2008) for African and South American countries. As some North American species winter in south USA, habitat cover by area, for the relevant physiographic provinces, was used, obtained from the ‘Centre of Advanced Spatial Technologies (CAST)’ website CAST (2008). The habitat types used were ‘Forests’ for woodland species, ‘Shrublands, Savanna and Grasslands’ for savanna and grassland species and ‘Wetlands’ for freshwater and brackish species. For marine coastline species the length of coastline of each region occupied was used as the weighting factor. In the northern sub-Saharan region of Africa, where the habitat is split into bands from desert to savanna grasslands/forest, only non-desert area was considered
Phen_Index	Phenology Index: Wint_Temp − Breed_Temp
Breed_Lon	The central latitude and longitude of countries were obtained from ‘The University of Harvard’s Centre for International development Research Datasets’ webpage (Centre for International Development 2008). The positions of physiographic regions in North America were obtained from CAST (2008). Breed_Lon was the average longitude of the relevant breeding region, weighted by abundance
Breed_Lat	Average breeding latitude calculated in the same way as Breed_Lon
Wint_Lon	Average longitude of the winter grounds. This was calculated in the same way as Breed_Lon, except the weighting scheme was as described in Wint_Temp for averaging across winter regions
Wint_Lat	Average wintering latitude calculated in the same way as Wint_Lat
Ave_Dist	Straight line migration distance between the average wintering and breeding positions. This was calculated using the spherical law of cosines (Weisstein 2008), which gives distance in km between two points on the Earth. d = a cos(sin(lat1).sin(lat2) + cos(lat1).cos(lat2).cos(lon2 − lon1)).R, where d is a distance in km, lat1, Wint_Lat; lat2, Breed_Lat; lon1, Wint_Lon; lon2, Breed_Lon; and R, Radius of the earth, taken to be 6371 km
Max_Dist	The largest possible migration distance found by considering the central points of wintering and breeding countries/regions which had the largest spatial separation. The distance was calculated as Ave_Dist
Min_Dist	The smallest possible migration distance found by considering the central points of the wintering and breeding countries/regions which had the smallest spatial separation. The distance was calculated as Ave_Dist
Arrival	Arrival time at the breeding grounds.
Forage	Foraging behaviour listed as one of the following 1 – birds of prey, 2 – divers and dabblers, 3 – waders and larids, 4 – land-birds (open), 5 – land-birds (covered) and 6 – aerial feeders
Mass	Mass in grams
Cont	Two-way factor: European breeding birds and North American breeding birds
Span	Average longitudinal span in °longitude of winter and breeding grounds. Individual wintering and breeding spans were found by considering the longitudinal span between easternmost and westernmost regions

Methods – data analysis

The four main explanatory variables chosen as potentially important with regards to testing the phenology mismatch hypothesis were Phenology Index, Average Distance, Mass and Continent (Cont). Models incorporating all possible subsets and two-way interactions of these four variables (n = 113 models) were investigated. Adopting an Information Theoretic approach the best model was identified as the model with the lowest AIC value where predictors included in the model had significant parameter estimates. All models were generalized linear models, run using r version 2.6.0 (normal based errors model utilizing an identity link function) with population trend as the dependent variable.

The predictive performance of two-way interactions involving the main variables, particularly Phenology Index × Cont and Average Distance × Cont, were evaluated against the performance of the interaction of Cont with a dummy variable (random), with random values drawn from a population of the same mean and standard deviation as the variable being tested. If the interaction between Phenology Index × Cont and Average Distance × Cont are not significant predictors, then an interaction of random × Cont should give similar levels of predictive power (i.e. similar levels of AIC). All possible models (1115) incorporating the four independent variables and a random variable were run using the genmod procedure in sas version 9.1. The models were then ranked according to AIC weight. The sum of Akaike weights for each variable was then calculated across all possible models (effectively allowing the variables to be ranked according to their predictive power, with the random variable setting the lower limit below which variables could be considered not to be predictors – see Whittingham et al. 2005). To obtain confidence limits for our random variable, we then ran this procedure 25 times, each with a different set of values for the random variable. The significance of a variable as a predictor was then determined by matched pairs t-tests of each variable against the random variable (N = 25).

We tested whether the degree of pseudoreplication (because of using separate populations as our sampling unit) in our data set was affecting the results by replicating the final best model with a smaller subset of data where only one case was taken from a species, so resulting in 134 cases. Where several separate populations existed for a single species we selected the population where the population change was the highest.

Alternative explanatory variables

We considered whether alternative explanatory variables could account better for the observed patterns of population decline with phenology mismatch or distance. If the addition of a new variable improved the model, while removing the significant effects of phenology mismatch or distance, we could then conclude that any effect of mismatch or distance was caused by covariance with the new variable. We tested for the effects of longitudinal span of breeding and wintering ranges (combined into an average span = {breeding span + wintering span}/2 as breeding and wintering span are highly correlated), breeding and wintering latitude, breeding and wintering habitat, arrival time and foraging guild (Table 1).

We also considered whether any of the alternative explanatory variables could account better for the observed difference in the effects of phenology mismatch or distance on patterns of population decline by continent. If any one of the alternative variables was significantly different between the two continents, then an interaction term of the alternative variable with Average Distance and the alternative variable with Phenology Index were also included in a model. If the addition of the interaction term improved the model, while removing the significant effects of any interactions with continent, we could then conclude that any effect of continent was caused by differences in the alternative variable between the two continents. Span (t₁₉₁ = 3·4, P = 0·001), breeding latitude (t₁₉₁ = 8·4, P < 0·001), arrival time (z_113,80 = −5·5, P < 0·001) and foraging guild (z_113,80 = −6·8, P < 0·001) were significantly different compared across continents but not wintering latitude (t₁₉₁ = 1·2, P = 0·23), breeding habitat (z_113,80 = −0·07, P = 0·95) and wintering habitat (z_113,80 = −0·4, P = 0·71).

Results

The highest ranked model showed significant interactions with both Phenology Index × Cont and Average Distance × Cont (Table 2), indicating Phenology Index and Distance relationships were significant and that they were significantly different between the two continents. Phenology Index showed a significant positive relationship with population trend in the Nearctic (14·05 ± 12·21, P = 0·025) but a negative relationship in the Palaearctic (−2·7 ± 4·6, P = 0·25). Average Distance showed a positive relationship with population trend in the Nearctic [(1·07 ± 1·78) × 10⁻³, P = 0·24] but a negative relationship with population trend in the Palaearctic [(−1·68 ± 2·1) × 10⁻³, P = 0·12]. There was a trend for the effects of distance to be less for species with larger mass.

Table 2.   Models incorporating all possible subsets and two-way interactions of the four variables (n = 113 models) were investigated. The best model was identified as the model with the lowest AIC value where predictors included in the model had significant parameter estimates. All models were generalized linear models (normal based errors model utilizing an identity link function) with population trend (percentage change per decade) as the dependent variable. The following table lists the models with ΔAIC <2. N = 193 bird populations from 134 species

Modnum	Factors	d.f.	Log-likelihood	par	AIC	ΔAIC
AveDist = Average Distance, Cont = Continent, PhenIndex = Phenology Index, d.f. = degrees of freedom, par = number of parameters.
81	Mass, AveDist, Cont, PhenIndex, Mass × AveDist, AveDist × Cont, Cont × PhenIndex	185	−801·16	9	1620	0·00
6	AveDist, Mass, AveDist × Mass	189	−805·24	5	1621	0·16
71	Mass, AveDist, Cont, PhenIndex, AveDist × Cont, Cont × PhenIndex	186	−802·24	8	1621	0·16
100	Mass, AveDist, Cont, PhenIndex, Mass × AveDist, Mass × PhenIndex, AveDist × Cont, Cont × PhenIndex	184	−800·73	10	1622	1·15
5	AveDist, Mass	190	−806·79	4	1622	1·27
17	AveDist, Mass, Cont, AveDist × Mass	188	−804·87	6	1622	1·42
27	PhenIndex, AveDist, Mass, AveDist × Mass	188	−805·01	6	1622	1·71
1	Mass	191	−808·05	3	1622	1·79
87	Mass, AveDist, Cont, PhenIndex, Mass × Cont, AveDist × Cont, Cont × PhenIndex	185	−802·07	9	1622	1·82
102	Mass, AveDist, Cont, PhenIndex, Mass × AveDist, AveDist × Cont, AveDist × PhenIndex, Cont × PhenIndex	184	−801·13	10	1622	1·95
97	Mass, AveDist, Cont, PhenIndex, Mass × AveDist, Mass × Cont, AveDist × Cont, Cont × PhenIndex	184	−801·13	10	1622	1·95

The best model (Table 3) (identified as the model with the lowest AIC value where predictors included in the model had significant parameter estimates) was the third model having ΔAIC = 0·1615 from the top model, and so not discernibly different from the highest ranked model (see Table 2). Phenology Index again showed a significant positive relationship with population trend in the Nearctic (14·04 ± 12·25, P = 0·026) (Fig. 2a) but a non-significant relationship in the Palaearctic (−2·66 ± 4·6, P = 0·26) (Fig. 2b). Average Distance showed a non-significant relationship with population trend in the Nearctic [(0·89 ± 1·8) × 10⁻³, P = 0·32] (Fig. 2c) but a significant negative relationship with population trend in the Palaearctic [(−2·31 ± 1·9) × 10⁻³, P = 0·02] (Fig. 2d).

Table 3.   Best model based on AIC and significance of component variables. Dependent variable: Population Trend (percentage change per decade). Phenology Index in the Nearctic exhibits a significant positive relationship and Average Distance has a significant negative parameter estimate in the Palaearctic. The mass relationship indicates larger birds are declining less than smaller birds

Source	Model effects			Parameter estimate
	Wald χ^2	d.f.	Sig.	Estimate	SE	Wald χ^2	d.f.	Sig.
^aDue to the interaction term with Cont this parameter estimate represents correlations in the Nearctic and the χ^2 is for the parameter estimate. ^bDue to the interaction term with Cont this parameter estimate represents correlations in the Palaearctic. The Phenology Index parameter values for the Palaearctic are derived from summing the parameter estimates for Phen_Index and Cont × Phen_Index, and the Average Distance parameter values for the Palaearctic are derived from summing the parameter estimates for Ave_Dist and Cont × Ave_Dist. The χ^2 is for the parameter estimate.
Intercept	1·424	1	0·233	−15·747	6·2075	6·435	1	0·011
Cont	7·612	1	0·006	22·021	7·9819	7·612	1	0·006
Ave_Dist	1·161	1	0·281	0·001^a	0·0009	1·026^a	1	0·311
Phen_Index	3·007	1	0·083	14·037^a	6·1353	5·235^a	1	0·022
Mass	11·145	1	0·001	0·007	0·002	11·145	1	0·001
Cont × Phen_Index	6·48	1	0·011	−16·7^b	6·5606	6·48^b	1	0·011
Cont × Ave_Dist	5·902	1	0·015	−0·003^b	0·0013	5·902^b	1	0·015

Figure 2.  Partial residual plots of Phenology Index and Average Distance on population trend (percentage change per decade) of the best biological model (see Table 3), individually for the Nearctic (a,c) and Palaearctic (b,d) with the fitted relationships plus 95% confidence intervals. The relationships in (a) and (d) are significant.

The best model was then tested further to ascertain the robustness of the observed relationships. All possible two-way interactions were included in the model (Table 4). The results are broadly similar with Phenology Index × Cont remaining a significant predictor and Average Distance × Cont showing a trend. The predictive performance of the two-way interactions was significantly much greater than interactions with a dummy variable of random values. The sum of Akaike weights for all possible models of Phenology Index × Cont was significantly higher than for random × Cont (Phenology Index × Cont, 0·60 ± 0·02; random × Cont, 0·31 ± 0·04; matched pairs t-test, N = 25, t = 9·6, P < 0·001). The sum of Akaike weights for all possible models for the Average Distance × Cont interaction was significantly higher than for random × Cont (Average Distance × Cont, 0·53 ± 0·02; random × Cont, 0·31 ± 0·04; matched pairs t-test, N = 25, t = 7·5, P < 0·001). The predictive power of the interaction term Average Distance × Mass was also tested. This term is marginally significant, P = 0·095, in the model incorporating all four variables and two-way interactions (Table 4). This term acts as a surrogate for the number of staging sites. In general, larger birds have a shorter flight range and so a negative estimate for the interaction term between mass and distance would indicate that an increase in the number of stops during migration is related to declining bird populations. The sum of Akaike weights for models incorporating this factor was significantly higher than for random × Average Distance (Average Distance × Mass, 0·53 ± 0·02; random × Average Distance, 0·41 ± 0·05; matched pairs t-test, N = 25, t = 2·4, P < 0·023). There was no evidence for the effects of Phenology Index to be stronger for longer distance migrants (Table 4).

Table 4.   Model with the four main effects and all their two-way interactions. Dependent variable: Population Trend (percentage change per decade)

Source	Model effects			Parameter estimate
	Wald χ^2	d.f.	Sig.	Estimate	SE	Wald χ^2	d.f.	Sig.
^aDue to the interaction term with Cont this parameter estimate represents correlations in the Nearctic. The χ^2 is for the parameter estimate. ^bDue to the interaction term with Cont this parameter estimate represents correlations in the Palaearctic. The χ^2 is for the parameter estimate.
Intercept	2·73	1	0·099	−16·713	6·34	6·95	1	0·008
Cont	4·65	1	0·031	19·483	9·04	4·65	1	0·031
Mass	2·71	1	0·099	0·021^a	0·0184	1·31^a	1	0·253
Phen_Index	0·75	1	0·386	13·841^a	7·79	3·16^a	1	0·076
Ave_Dist	0·06	1	0·808	0·001^a	0·0009	1·43^a	1	0·232
Cont × Mass	0·08	1	0·773	0·017^b	0·0066	6·39^b	1	0·012
Cont × Phen_Index	5·44	1	0·02	−2·396^b	7·12	0·11^b	1	0·736
Cont × Ave_Dist	2·66	1	0·103	−0·001^b	0·0012	1·48^b	1	0·225
Mass × Phen_Index	0·86	1	0·353	−0·006	0·0065	0·86	1	0·353
Mass × Ave_Dist	2·79	1	0·095	−2·83 × 10^−6	1·69 × 10^−6	2·79	1	0·095
Phen_Index × Ave_Dist	0·02	1	0·885	0	0·0012	0·02	1	0·885

Our results were broadly replicated when we reduced pseudoreplication using only one case for each species (N = 134 rather than 193), with only the reduction in statistical significance and power expected, due to a smaller sample size (Table 5). However, the significant effect of mass in the model was not present when using the reduced data set. The effect of mass in the model is driven by several large bird species, that can be split into several populations and that have positive population trends (Table 3).

Table 5.   Model using a reduced data set (n = 134), where all possible pseudoreplicated data points have been removed by including only one population (largest population trend) for each species. Dependent variable: Population Trend (percentage change per decade)

Source	Model effects			Parameter estimate
	Wald χ^2	d.f.	Sig.	Estimate	SE	Wald χ^2	d.f.	Sig.
^aDue to the interaction term with Cont this parameter estimate represents correlations in the Nearctic. The χ^2 is for the parameter estimate. ^bDue to the interaction term with Cont this parameter estimate represents correlations in the Palaearctic. The Phenology Index parameter values for the Palaearctic are derived from summing the parameter estimates for Phen_Index and Cont × Phen_Index, and the Average Distance parameter values for the Palaearctic are derived from summing the parameter estimates for Ave_Dist and Cont × Ave_Dist. The χ^2 is for the parameter estimate.
(Intercept)	3·97	1	0·046	−18·261	6·2818	8·45	1	0·004
Cont	5·35	1	0·021	19·693	8·5163	5·35	1	0·021
Ave_Dist	0	1	0·998	0·001^a	0·0009	1·97^a	1	0·161
Phen_Index	3·86	1	0·049	14·902^a	6·342	5·52^a	1	0·019
Mass	1·09	1	0·297	0·002	0·0023	1·09	1	0·297
Cont × Phen_Index	5·81	1	0·016	−16·41^b	6·8099	5·81^b	1	0·016
Cont × Ave_Dist	3·48	1	0·062	−0·003^b	0·0014	3·48^b	1	0·062

Sensitivity analyses indicated that distance relationships were robust to variations in measures of migration distance. When Average Distance was substituted with either the maximum possible migration distance (ΔAIC = 0·26; Cont × Max Distance, P = 0·026; Cont × Phenology Index, P = 0·012), or the minimum distance (ΔAIC = 2·3; Cont × Min Distance, P = 0·042; Cont × Phenology Index, P = 0·008), the model results were similar.

When Phenology Index was substituted with its component wintering area temperature trend and breeding area temperature trend (ΔAIC = 1·4; Cont × Average Distance, P = 0·005; Cont × wintering temperature trend, P = 0·004; Cont × breeding temperature trend, P = 0·54), a similar but slightly poorer model resulted. Winter temperature trend was a strong predictor of population trends whereas breeding area temperature trend was not, although note that the temperature difference (Phenology Index) was a stronger predictor of population trend than temperature trends of either of the areas alone.

The effects of alternative explanatory variables, which may have been driving relationships, were assessed by inserting each variable into the best model identified above. The Phenology Index × Cont result remained significant in all models except for where arrival month was included as an interaction term (Table 6). Arrival time was also found to be significantly different between the two continents (P < 0·001) and so may be driving the difference in relationships, but Arrival × Phenology Index in place of the Cont × Phenology Index interaction in the best model was not a significant predictor (Arrival × Phenology Index, P = 0·39). The Average Distance × Cont result remained for most models but was lost when Foraging Guild was present in the model (Table 6). In this case, the interaction between Foraging Guild × Average Distance was significant but the model AIC was not substantially improved (Table 6). These results suggest that differences between the continents in Foraging Guild may be driving the interaction Average Distance × Cont. The interaction terms for Phenology Index and Average Distance remain for most models and so the relationships can be said to be reasonably robust. These additional models suggest that differences between the continents in the effect of Phenology Index and distance may be due to differences in arrival times and foraging guilds due to differences in species assortments between the two continents.

Table 6.   Other potentially explanatory variables were added to the best biological model, either as a main effect, or when the variable was significantly different between continents, included as an interaction term with Average Distance and Phenology Index

Potential explanatory variable	ΔAIC	As main effect χ^2	Variable ×  Ave_Dist χ^2	Variable ×  Phen_Index χ^2	Ave_Dist ×  Cont χ^2	Phen_Index ×  Cont χ^2
^aParameter estimate −8·6 × 10^−1. ^bParameter estimates relative to arrival in April, March: 1 × 10^−2*, May: 2·9 × 10^−4, June: 1·1 × 10^−3. ^cParameter estimates relative to foraging guild 1, guild 2: 0·01, guild 3: −6·5 × 10^−3*, guild 4: −2·9 × 10^−3, guild 5: −3 × 10^−3, guild 6: 8·7 × 10^−4. *P = 0·05, **P < 0·01, ***P < 0·001.
Span	0·1	–	0·01	3·3	3·8**	5·6**
Breeding latitude	−4·5	–	0·1	9·7***^a	5·4**	5·3**
Wintering latitude	2	0·04	–	–	5·9**	6·3**
Breeding habitat	9·4	0·6	–	–	5·2**	5·4**
Wintering habitat	2·5	5·6	–	–	4·1**	4·7**
Arrival	−0·4	–	4·8**^b	0·37	4·9**	2·5
Foraging guild	8·8	–	2·3*^c	1·4	0·003	4·9**

The testing of alternative variables gave two further additional significant effects on the action of Phenology Index and distance. The effects of a phenology mismatch were less pronounced at higher breeding latitudes: the parameter estimate for Phenology Index × Breeding Latitude was significantly negative (Table 6). There was also some indication that the detrimental effects of distance on population declines were less for early arriving birds: the parameter estimate for Arrival × Average Distance was significantly more positive for late arriving populations (Table 6). This relationship, however, may be driven by a few populations arriving exceptionally early in March (n = 9) as parameter estimates for populations arriving in May and June are not significantly different from those for April.

Discussion

Overall phenology mismatch and migration distance were found to be general predictors of population trend in migrant birds in the Nearctic and Palaearctic respectively. Differences in physical geography, anthropogenic change and effect of climate change will be different between the two continents; hence differences between continents perhaps might be expected. This result was robust to the effects of variation in mass, latitudinal and longitudinal variation in a species’ range, habitat, arrival time and the foraging guild of the species.

There was some evidence for an effect of breeding latitude and arrival time on the effects of phenology mismatch and distance respectively. The degree of population decline that arises with phenology mismatch was more pronounced for those species that breed at lower latitudes. This may have arisen because temperature variations are less at lower latitudes, so that any one degree change in temperature has a greater overall effect at lower latitudes: bird populations at higher latitudes may have already undergone selection for phenotypic plasticity in response to inter-annual variability in phenology (Charmantier et al. 2008). The degree of population decline that arises as migration distance increases is greater for those species that arrive later. This might arise because later arriving, longer distance migrants are more constrained in terms of resources used, and time available to regain them for successful breeding (Drent 2006).

There are, however, several caveats for these conclusions. The phenology mismatch hypothesis makes the assumption that climate change affects migration phenology and resource availability to the same degree, so that a mismatch in temperature change will cause a correlated mismatch of arrival and resource availability. However, responses to climate change will vary across taxa so that birds with a similar Phenology Index, but which forage on largely separate resources, may experience different mismatches dependent on their resources’ response to climate change (Harrington, Woiwod & Sparks 1999). This may in part be responsible for the relative weakness of observed phenology relationships.

We also predicted that a large positive Phenology Index would be beneficial. However, in some cases, a large positive Phenology Index may be detrimental, because if birds arrive long before the resource peak they may have to wait for resources to become available or migrate further. However, we found no evidence for a decline in population at higher positive phenology mismatch values (i.e. a negative quadratic relationship between Phenology Index and population change: Phenology Index squared term added to the best model in Table 3 without interactions, Nearctic, B = 12·9 ± 10·2, P = 0·21; Palaearctic B = 3·7 ± 1·9, P = 0·053).

Further caveats are that climate change affects more than phenology. For example, drought and desertification in the northern Sahel region of Africa are believed to be one of the primary reasons for decline of trans-Saharan migrants (Newton 2004), and in warmer drought years there will be fewer resources available regardless of phenology changes (Gordo et al. 2005). Shifting and expansion of breeding and wintering ranges in response to climate change (Pulido et al. 2001), as well as changes in migration routes (Bearhop et al. 2005), would also have significant confounding effects for the detection of climate driven phenology changes. Specifically, short-distance migrant wintering grounds have advanced northwards in response to milder winters (Gordo et al. 2005). This would mask any of the effects associated with differential climate change on population trends, as the primary change in conditions affecting population trend will largely be driven by milder winters on the breeding grounds rather than mismatch. However, our data selection criteria deliberately excluded such populations and so we would consider these effects to be minimal in our analysis. Controlling for the effects of changes in distribution of course depends on evidence being available, and this is almost entirely lacking for wintering migrants.

As expected from previous studies (Berthold et al. 1998; Sanderson et al. 2006), distance in the Palaearctic was negatively correlated to population trend. The distance hypothesis states that this is because the probability of experiencing a mismatch at any one site is increased with the number of sites used. If this is the case then we would expect larger birds – migration range is inversely proportional to mass (Weber & Houston 1997) and so larger birds need to make fewer migration stops – to show smaller population effects that were dependent on distance. We would also expect the effects of phenology mismatch to be greater for those birds that travelled the greatest distance. We, however, found only weak evidence for a significant negative interaction between Mass × Distance, and no evidence for an interaction between Phenology Index × Distance (Table 4) suggesting that population decline differences between long and short-distance migrants may not be due to mismatch at staging sites. Nevertheless, there is other evidence that staging area phenology is a vital component of migration phenology, because climactic conditions on staging grounds strongly influence the availability of resources during migration (Marra et al. 2005;Huppop & Winkel 2006).

Further factors could be driving declines in long-distance migrants. Arrival date amongst short-distance migrants has been shown to be more closely linked to temperature on breeding grounds than for long-distance migrants (Cotton 2003; Gordo et al. 2005). This may be due to similar temperature trends for breeding and wintering ranges of short-distance migrants as the two areas are less separated and are more likely to share similar climatic conditions. For long-distance migrants, however, winter climate may be completely uncorrelated with the climate on the breeding grounds. Hence, short-distance migrants may be responding to a cue which is closely linked to that which they need to respond to in order to arrive and breed successfully, whereas longer distance migrants are responding to a cue which may be uncorrelated with the timing of spring phenology events. Furthermore, long-distance migrants may respond to endogenous rhythms, associated with changes in photoperiod on the wintering grounds, to initiate migration and thus have limited abilities to respond to climate change (Both & Visser 2001; Gordo et al. 2005). However, it can also be argued that because migrants rely on sufficient resources being available prior to migration, which is directly affected by climactic conditions preceding migration, some additional flexibility in departure date in response to resource levels is likely to be present.

There are numerous resolution limitations in our analysis. For example, inadequate resolution of breeding and wintering areas and also specific migration dates for specific populations. Birds within a population tend to arrive and depart wintering and breeding grounds over an extended period of time with higher latitude populations arriving later than lower latitude populations (Cramp 1998). The Phenology Index and migration distance used in the analysis were average values for the entire inhabited region of that particular subspecies, and thus may not be representative for all populations of that subspecies [although distance and Phenology Index were calculated as weighted averages by abundance (see Table 1) to minimize this]. In addition, wintering areas had a significantly lower resolution than breeding areas, due to the lack of distribution data in South America and Africa. We also assumed that area of suitable habitat is representative of the species distribution in their wintering ranges (see Table 1). Problems arise when birds, in particular wetland populations, congregate at high densities in small areas so that a region with a small area of suitable habitat may contain a disproportionate number of birds. Furthermore, the habitat classifications (Forests, Wetlands, etc.) are quite coarse. Nevertheless, such errors in resolution are only likely to cause noise in the data. So the fact that robust relationships are observed at such relatively poor spatial resolution probably reinforces the strength of them.

Despite the inevitable shortcomings of the data, we provide evidence to support phenology mismatch driving migrant bird declines and also some evidence that distance of migration may lead to declines. We do not conclude that the phenology mismatch hypothesis is necessarily a general global relationship, because there were significant differences between continents with Phenology Index as a strong predictor in the Nearctic and distance as a predictor in the Palaearctic. However, as anthropogenic habitat destruction and further effects of climate change strongly affect population dynamics, such relationships may be being masked or weakened in the two areas to different degrees. Furthermore, it can be argued that as we detected such relationships, when there are so many confounding variables and deficiencies in the data to consider, the strength of the effects are likely to be considerable. However, an increase in resolution would strengthen the plausibility of these conclusions. Furthermore, trends in arrival and breeding dates could be included in the analysis to determine whether climate change and population changes are linked to actual changes in the timing of migration and breeding events. However, data of sufficient quality, resolution and over a long enough time period to detect significant changes are not available for many populations, limiting the analysis to a regional, rather than global approach.

Acknowledgements

We thank the many thousands of people who collected the huge amount of distribution, temperature and natural history data used here. We also thank Christiaan Both, Graeme Ruxton, Mark Whittingham and an anonymous referee for comments on an early draft. Will Cresswell was supported by the Royal Society.