Estimates of Population Size

Population size estimates of European bird species were obtained from websites and persons responsible for the Pan-European Common Bird Monitoring Scheme in every European country. We obtained data for only 12 countries, either because in some countries this scheme has started very recently or has not started yet, or because information was unavailable upon request. Population indices from some countries or regions could not be used in this study owing to incomplete information (e.g. Wallonia), or because bird censuses were done only in one type of habitat (e.g. Latvia). Countries with available and suitable information on avian population size, the source of this information, number of bird species per country, mean number of years with data in every country (not all species were always surveyed the same number of years in a country), and last year with information on population size are shown in Table 1. Canada goose Branta canadensis and common pheasant Phasianus colchicus were excluded because they are introduced in Europe [41,42]. In Sweden, information on the willow warbler only includes the subspecies Phylloscopus trochilus acredula and information on the chiffchaff only includes the subspecies Phylloscopus collybita abietinus. The search for information was finished by 8 June 2009.

The estimates of population size that we obtained were always standardized to a value of one in a particular year and the rest of the years indicated a value relative to the reference year. For example, if the population size estimate in a year was two for a particular species, it meant that the population size of that species was twice the value in the reference year. Therefore, these population size estimates are always relative values and are usually called population indices (for an example of the use of population indices, see [43]). These indices were calculated in the same way in all European countries following recommendations made by the European Bird Census Council (EBCC; see http://www.ebcc.info). Specifically, population indices were calculated using the software TRIM (Trends and Indices for Monitoring Data). More information about the programme TRIM can be found at the website http://www.cbs.nl/en-GB/menu/themas/natuur-milieu/methoden/trim/default.htm. For most countries, the reference year was the first year of census, although that was not the case for Sweden, Germany and France, where the reference year was 1998, 1999 and 2001, respectively. In the case of Spain, the reference value was 0 instead of 1, and then population indices were transformed ((X+100)/100, where X was the population index), to make values comparable with those from other countries. For some countries (Czech Republic, Finland, France, and Poland) only graphs showing population indices, but not population indices themselves were available, and in these cases population indices were estimated from graphs using the program ImageJ (http://rsbweb.nih.gov/ij/). In order to assess the accuracy of population indices calculated from the graphs, we estimated population indices from graphs in two countries, Spain and Sweden, for which both population indices and graphs were available. Twenty five species were randomly chosen per country, but only considering species with data for the maximum number of years available for the country, i.e., 11 years in Spain and 34 years in Sweden. Population indices estimated from the graphs or directly shown at the websites were highly repeatable ([44]; Spain: r ≥ 0.998, F_10,11 ≥ 842.00, P < 0.00001 for 25 species; Sweden: r ≥ 0.994, F_33,34 ≥ 326.32, P < 0.00001 for 25 species).

The present study considered populations at the country level, a geographical scale considerably larger than the traditional ecological concept of a population, and thus we used countries as sample units. This should be appropriate for a continental-scale study, as was the case here, if two requirements are fulfilled. First, countries should be relatively small in relation to the breeding range of the birds, as happened to be the case here (mean (SE) country area = 0.263 x 10⁶ (0.053 x 10⁶) km², n = 12 countries, Table 1; mean (SE) Western Palearctic breeding distribution area = 18.121 x 10⁶ (0.331 x 10⁶) km², n = 73 bird species [32], and unpublished data). Second, censuses must be representative of the whole country, covering different regions and habitats within the country in a balanced way. This is the procedure recommended by the EBCC and followed by most countries. Other studies investigating gradients at a continental scale and population indices also used countries and/or very large regions as sample units (e. g. [45]).

Information on population indices is freely available in different websites (see Table 1) or can be provided upon request.

Population Latitude and Marginality

In order to estimate the marginality of bird populations, i.e., whether populations were central or marginal within the breeding distribution range, we first calculated a latitude index for each of the 12 European countries included in the study (Table 1). This latitude index was estimated as the latitude of the mid-point between the northernmost and the southernmost mainland points of every country. This information was obtained using maps in atlases and online with the software Google Earth (http://earth.google.com). Islands were not considered because they generally add a lot of distance in terms of latitudinal degrees, but very small area in relation to the entire country. The exception was Denmark, because a large part of that country is made up by islands, and the southernmost point of the country was located at Falster. The latitude index for each country was considered the latitude for all bird populations in that particular country regardless of the actual distribution of every species within a country. Once the latitude of the bird populations had been estimated, we calculated the distance (in degrees) from the bird population latitude to the northernmost and southernmost limits of the breeding distribution range of the species. The smaller of these two distances (L) was considered the smallest distance to the distribution limits of the species. In the few cases in which the country latitude index was more southern than the southernmost limit of the species range (herring gull Larus argentatus in France), or more northern than the northernmost limit of the species range (reed warbler Acrocephalus scirpaceus and golden oriole Oriolus oriolus in Finland, yellow-legged gull Larus michahellis, great spotted cuckoo Clamator glandarius, cattle egret Bubulcus ibis, southern grey shrike Lanius meridionalis, black-eared wheatear Oenanthe hispanica, subalpine warbler Sylvia cantillans, and Sardinian warbler Sylvia melanocephala in France, and woodlark Lullula arborea and red kite Milvus milvus in Sweden), the distance between bird population latitude and southernmost or northernmost limits of the distribution range of the species was considered to be zero. Northernmost and southernmost limits of the breeding distribution range of every species were obtained from published maps [46]. We also calculated the latitude of the mid-point between the northernmost and the southernmost limits of the distribution range of the species, and the distance (C) in degrees between this latitude and the bird population latitude. Marginality of a population was estimated as log₁₀(C+1) - log₁₀(L+1), with positive values representing marginal populations and negative values central populations. Finally, these values were transformed to ensure that marginality estimates ranged from 0 (central population) to 1 (marginal population). The transformation consisted of adding the absolute value of the most negative number and dividing by the largest value resulting from the previous addition.

Confounding Factors

One population parameter that might affect among-year variability in population size and, as a consequence, the results of our study is population density. One example of such an effect would be if populations at high density are close to carrying capacity. If that is the case, population size at high density is unlikely to vary much, and temporal variation in population size should be smaller than at low population density. However, populations at high density might vary to the same extent as populations at low density, the former mostly decreasing and the latter mostly increasing in size. In that case we would not expect a difference in temporal variation between populations with high and low density. Furthermore, in some populations with occasional or regular demographic explosions (e.g. snowy owl Nyctea scandiaca; [47]), variability in population size may be larger at high than at low density. Whatever happens in particular cases, density-dependence of population size variation is a general rule in a wide range of animal taxa [48-50]. Population indices do not provide information on absolute density (number of individuals per unit area), because they are relative values in relation to population size in a particular year. However, they give information about relative density within a population, because large population indices correspond to high density and small population indices to low density in that particular population. Therefore, within-population comparisons among years with (relatively) high and low density are possible. We estimated effects of density-dependence on among-year variability of population indices by comparing CV between years with high and low population indices (within populations). Specifically, for every bird population (i.e., every species in every country) we ranked population indices decreasingly and then divided the data in two halves, one with the largest population indices and the other with the smallest population indices. When number of study years was an odd number, the mid-point population index was included in both halves. In this way, the two subsets for a given population always included the same number of population indices. Then, CV of population indices was estimated for each of the two subsets, CV_high for large population indices and CV_low for small population indices. Since number of study years (n) used to calculate CV varied among countries and sometimes also among species within a country, CV values were transformed to obtain CV corrected for sample size (CV*) with the following transformation: CV* = CV (1 + 1/4n) [51]. Reduced Major Axis (RMA) regressions of CV*_high on CV*_low had in general slopes not significantly different from one after sequential Bonferroni correction ([52]; 12 tests; mean (SE) slope = 1.02 (0.05), n = 11, range from 0.79 to 1.25; -2.83 ≤ t ≤ 2.71, 56 ≤ df ≤ 168, and P ≥ 0.0078 for 11 countries, the exception being Poland: slope = 0.76, t₁₀₀ = -3.65, and P < 0.001) and intercepts not significantly different from zero after sequential Bonferroni correction (mean (SE) intercept = -0.37 (0.59), n = 12, range from -4.47 to 2.39; -2.08 ≤ t ≤ 1.68, 56 ≤ df ≤ 168, and P ≥ 0.044 for the 12 countries). CV*_high and CV*_low did not differ significantly from each other after sequential Bonferroni correction when they were compared within populations (paired t-test; -2.60 ≤ t ≤ 1.72, 57 ≤ df ≤ 169, and P ≥ 0.012 for the 12 tests). As density-dependence should be controlled statistically, we included in our analyses a variable that reflected density-dependence in every population calculated as log₁₀CV*_high - log₁₀CV*_low. Very large positive values would imply strong density-dependence with larger variation at high than at low densities. In contrast, very large negative values would indicate strong density-dependent effects with larger variation at low than at high densities. Values around zero would imply weak or no density-dependence.

Another parameter that might have an effect on population size is habitat fragmentation [53-55]. Fragmentation of land due to urbanisation, transport infrastructure and agriculture has been estimated for most European Union countries by the European Environment Agency (EEA). Methods used by EEA to calculate fragmentation indices and a map of Europe at a 10x10 km grid resolution showing levels of fragmentation can be found at the website http://www.eea.europa.eu/data-and-maps/figures/fragmentation-by-urbanisation-infrastructure-and-agriculture. In this map, the degree of fragmentation is shown by different colours representing six categories of fragmentation from minimal to extreme. We first calculated the percentage of the area of each country for every category of fragmentation by using the software Adobe Photoshop CS4 Extended, v. 11.0.2 (Adobe Systems Inc., San Jose, CA). We then assigned a value of fragmentation from one to six to the different degrees of fragmentation and calculated the mean fragmentation index weighted by area for every country (Table 1). Information on fragmentation was unavailable for Norway, but we assumed that it was very similar to the fragmentation index of the other countries of the Scandinavian Peninsula, i.e., Finland and Sweden. We performed all the statistical analyses three times using as fragmentation index for Norway the Finnish value, the Swedish value, or the mean of the Finnish and Swedish values. As results were qualitatively identical in the three cases, for brevity we only present results obtained using the mean of the Finnish and Swedish values.

At least two methodological factors might affect the estimation of population indices. First, we would expect that sampling effort affected population size estimates, with possible implications for among-year variability of estimates. Sampling effort was calculated as the number of fieldworkers in a country (this information can be found at the EBCC website, see Table 1) divided by area of the country (Table 1), i.e., fieldworkers per square kilometre. Second, census method might influence estimates of population fluctuations if some methods are more prone to errors than others. Actually, this seems to be the case, especially because point-counts show particularly large errors (B.-E. Sæther, personal communication). Observation errors have been suggested to significantly affect not only estimates of variation in population size [50], but also the apparent influence of density-dependence on this variation [56]. According to the EBCC website, most countries only used point-counts in their censuses, but methodology differed in four countries (Table 1). To control for different methodologies in different countries, we included in the analyses the variable “census method” with two categories, countries exclusively using point-counts (eight countries) and countries using other methods (four countries).

The present study implicitly assumes that global warming is both geographically and temporally uniform, allowing comparisons between time series of 34 years (Sweden) and only seven years (Poland), or between southern and northern European populations. Although these assumptions might not generally hold, we consider them to be acceptable in our case for the following reasons. Regarding geographic variation, almost all 12 countries included in this study have experienced a temperature increase during the twentieth century (range from -0.05°C to 1.27°C, mean (SE) = 0.67 (0.12) °C; temperature data obtained from [57]), while the relationship between temperature change and latitude for the 12 countries is far from significant (Pearson correlation, r = -0.250, n = 12, P = 0.43). In relation to temporal variation, and using data from the EEA (http://www.eea.europa.eu), the relationship between annual temperature change in Europe during 1975-2008 and year is also not statistically significant (Pearson correlation, r = 0.074, n = 34, P = 0.68). These results suggest that warming has been similar in southern and northern Europe and constant over the 34 years considered in the present study. We here use temperature variation as a proxy for climate variation that is necessarily more complex than just differences in temperature.

Statistical Analyses

The hypothetical relationships between marginality or latitude and population trends or fluctuations were tested among populations of every species. We obtained information on population indices and marginality for a total of 231 bird species, although the number of countries with information for every species differed greatly. For 55 species we could only find information in one country, for 45 species in two countries, 16 species in three countries, 14 species in four countries, five species in five countries, six species in six countries, 16 species in seven countries, ten species in eight countries, 12 species in nine countries, 15 species in ten countries, 15 species in 11 countries, and 22 species in all 12 countries. Since we were interested in testing for relationships among populations within species (i.e., among countries within species), it would make no sense to include in the analyses species with information for only one or two countries. On the other hand, if only species with information for all 12 countries were included, sample size (number of species) would be dramatically reduced, consequently decreasing statistical power. Therefore, we attempted to find a compromise between these extremes, including as many species as possible but only species with information for a reasonable number of countries. We chose species with data for at least eight countries (74 species) because this allows sufficient variation among populations of the same species, while simultaneously maintaining a relatively large sample size. However, we repeated the analyses including species with data for at least seven countries (90 species) and including only species with data for at least nine countries (64 species). In general, qualitatively similar results were obtained with the three data sets, with virtually no difference in the effect of our variables of interest, i.e., population marginality and latitude (see Appendix S1 in Supporting Information). The only qualitative difference referred to the relationship between SEE and habitat fragmentation (see Appendix S1).

To test for the hypothetical effect of marginality or latitude of a population on among-year variability in population size of European birds we performed a General Linear Mixed Model (GLMM), with CV of population indices among years (corrected for sample size, see above) as the dependent variable and marginality (as defined above) and latitude of the population (in degrees) as predictor variables. Species was included as a random factor and, as a result, all putative relationships between dependent variables (e.g. CV) and predictors (e.g. latitude) were tested within species, i.e., among populations of the same species. We also included in the model the density-dependence estimate, sampling effort, census method, habitat fragmentation (see estimate or definition of these four variables above), and number of years with population indices. As explained in the Introduction, among-year variability in population indices can be split into two components, namely the slope of the regression of population indices on years, and the dispersion of points around the slope, i.e., SEE of the regression. Therefore, we regressed population indices on years for every population, and repeated the same GLMM analysis described above, but with slope or SEE of these regressions as dependent variables. We noticed that two populations were clear outliers, namely the slope for stock pigeon Columba oenas in Hungary, which was 1.57 (all other slopes ranged from -0.18 to 0.67), and SEE for coot Fulica atra in Hungary, which was 4.01 (all other SEE ranged from 0.02 to 1.76) (Figure 1). Thus, we excluded these two populations from the analyses, implying exclusion of all coot data, because when excluding one population, the coot was no longer a species with data for eight countries. Analyses including the outliers in general yielded similar results, especially regarding population marginality and latitude (see Results).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Frequency distribution of population parameters in 767 populations of 74 European breeding bird species.

(a) Coefficient of variation (CV) of population size estimates (population indices) corrected for sample size (see text); (b) slope and (c) standard error of the estimate (SEE) after regressing population indices on year. Outliers are included (see text). Mean (SE) for the three parameters is CV: 21.95 (0.51); slope: 0.008 (0.003); SEE: 0.202 (0.008).

https://doi.org/10.1371/journal.pone.0077654.g001

In addition to testing for a possible relationship between latitude and population trend, we also estimated whether trends were positive or negative for southern and northern populations. First, all populations within each species were divided in two halves, the northernmost and southernmost populations, and the mean slope was calculated for each half. When the number of populations was an odd number, the most central population was excluded from the analyses. Second, slope means for northern and southern populations were averaged across species.

We acknowledge that slopes after regressing population indices on years are rough estimates of population trends (for more refined methods, see e.g. [58,59]). However, we are mainly interested in relative trends among populations (within species). As all slopes have been calculated in the same way, comparisons among them should be appropriate. Intrinsic variation associated with the estimates of annual population indices was not taken into account because it was unavailable for most countries. Nevertheless, as factors potentially affecting variation in these estimates (e.g., habitat fragmentation, sampling method or sampling effort) have been controlled in the analyses, we assume that the remaining variation is random (unbiased) and will add only noise to our analyses, making any significant relationship between variables conservative.

Statistical analyses in this study implicitly assume that population parameters in different countries are independent. However, this assumption might not be met if population parameters tend to be more closely related in neighbouring than in more distant countries. To address this issue, we checked whether CV, slope and SEE were more strongly correlated in contiguous than in non-contiguous countries. We considered contiguous countries to be those that shared a land border. In our case, there was a maximum of 13 pairs of contiguous countries (e.g., Germany-Austria, Germany-Poland, Spain-France, Norway-Finland, etc.). For each bird species, we correlated the three population parameters between all pairs of contiguous countries and also between the same number of pairs of non-contiguous countries. Pairs of non-contiguous countries were chosen randomly except for the fact that we only included countries already present in the pairs of contiguous countries for that species. Mean Pearson correlation coefficients across the 73 bird species were 0.0002 (CV), 0.083 (slope) and 0.012 (SEE) for contiguous countries and -0.079 (CV), -0.136 (slope) and -0.085 (SEE) for non-contiguous countries. In none of the three cases was the absolute value of the mean correlation coefficient larger in contiguous than in non-contiguous countries. This implies that population parameters were not more closely related in nearby than in more distant countries.

The number of years with population indices varied greatly among countries (Table 1), and we partially controlled for this variation by including the number of years as a covariate in the models. However, these analyses implicitly assume that population parameters are relatively constant through time. We checked this assumption in the four countries with data for more than 20 years (Czech Republic, Denmark, Finland and Sweden) by calculating CV, slope and SEE for every species in the last 10 years and in the previous 10 years, checking whether the two sets of data were related. In GLMM analyses including country as a random factor, we found positive and significant relationships between decades for the three population parameters (F_1,254 ≥ 11.49, P < 0.001 in the three cases). This means that population changes for a given species were similar in different decades, at least in these four countries. In some countries (Austria, Hungary and Poland), the number of years with population indices was rather small (< 10; Table 1), making population parameters more prone to error and thus less reliable. However, these countries did not have a strong influence on the results, because exclusion of them from the analyses generally provided qualitatively similar results (Appendix S1). The most remarkable difference was that the relationship between slope and latitude was statistically significant only after removal of non-significant factors from the model (see Appendix S1).

All statistical analyses were two-tailed with a significance level of 0.05, and performed with Stastistica v. 9.0 (http://www.statsoft.com), except RMA regressions that were performed with RMA Software for Reduced Major Axis Regression v. 1.17 (http://www.bio.sdsu.edu/pub/andy/RMA.html).

An example of how population parameters were calculated in a common bird species (the great tit Parus major) can be found in Appendix S2 (Supporting Information).