Local extinction of British farmland birds and the prediction of further loss


  • Simon Gates,

    Search for more papers by this author
    • Present address: National Perinatal Epidemiology Unit, Institute of Health Sciences, Old Road, Oxford OX3 7LF, UK.

  • Paul F. Donald

    Corresponding author
      * Present address and correspondence: Royal Society for the Protection of Birds, The Lodge, Sandy, Bedfordshire SG19 2 DL, UK (e-mail paul.donald@rspb.org.uk).
    Search for more papers by this author

* Present address and correspondence: Royal Society for the Protection of Birds, The Lodge, Sandy, Bedfordshire SG19 2 DL, UK (e-mail paul.donald@rspb.org.uk).


1. The populations of many UK farmland birds declined between 1970 and 1990, frequently accompanied by contractions in breeding ranges. Ornithological atlas data, land use data and environmental data at the scale of 10-km squares were used to investigate the relationship between local extinctions and habitat suitability for six species, and to predict where future losses are most likely.

2. For each species we tested the hypothesis that local extinctions were concentrated in environments that were inherently less suitable. We also tested the hypothesis that spatial patterns of loss were not independent between species due to their concurrence in the same habitats.

3. Multivariate analyses (PCA) showed that areas where each species became extinct between 1970 and 1990 were more similar in land use type, climate and topography to areas where a species was never present than those where it was retained; local extinction was more likely in less suitable environments. Multiple logistic regression showed that for five of the six species the environmental gradient best predicting presence or absence in 1970 was also that best predicting loss between 1970 and 1990. For the six species studied, local extinctions were least likely in lowland arable areas.

4. For any pair of species, local extinctions were more frequent outside the area of overlap of the two species' ranges than inside. Within the area of overlap, species tended to be lost from the same squares. For each species, likelihood of local extinction declined with increasing number of the other five species present.

5. We used model parameters to map the probability of future local extinctions of the six species considered, allowing the identification of key areas for conservation management at a spatial scale appropriate to agri-economic incentives.


Most species' ranges are not static but change through the colonization of new sites and the extinction of existing populations. In the case of highly mobile groups such as birds, such changes can occur very rapidly. For example, the collared dove Streptopelia decaocto expanded from a foothold in the Balkans to most of north-western Europe in just 25 years (Fisher 1953; Hudson 1965). Brown-headed cowbirds Molothrus ater spread into Oregon and Washington (USA) during the 1940s and 1950s at an estimated 70–78 km year−1 (Rothstein 1994). In theory, range contractions could happen even more rapidly, because they do not require any movement of individual animals. Changes in range size are generally correlated with changes in population size and density (Wilcove & Terborgh 1984; Bart & Klosiewski 1989; Gaston 1994; Donald & Fuller 1998), although the mechanisms behind this relationship are not well understood. Furthermore, spatial patterns of range change are complex. Recent analyses of distributional decline have shown that, perhaps counter-intuitively, remnant populations tend to occur at the peripheries of their historical ranges (Channell & Lomolino 2000). Conversely, analysis of UK ornithological atlas data suggested that populations decline towards their cores. These results are not contradictory and are likely to arise from the underlying nature of the cause of the declines (Donald & Greenwood, in press).

Fuller et al. (1995) identified farmland birds as a group that suffered severe declines in population and range in Britain between 1970 and 1990, including many formerly common species now regarded as being of high conservation priority (Gibbons et al. 1996). For many species, the causes of decline are poorly understood, although co-occurrence with agricultural intensification (Chamberlain et al. 1999), and the fact that farmland birds have declined more than species associated with other habitats (Fuller et al. 1995), suggests that changes in farming are partly responsible. Detailed ecological studies (reviewed in Campbell et al. 1997) have established firm links between the declines of some species and agricultural change.

In this paper, we combine ornithological atlas data from two periods around 20 years apart (1968–72 and 1988–91) with environmental data at the same 10-km square level to test hypotheses relating to spatial patterns of decline of six farmland species (grey partridge Perdix perdix L., lapwing Vanellus vanellus L., turtle dove Streptopelia turtur L., tree sparrow Passer montanus L., reed bunting Emberiza schoeniclus L. and corn bunting Miliaria calandra L.). These species were selected by virtue of their range contractions (range 9–32%), by being conspicuous species whose presence or absence could be confidently assumed from the atlas data, and by occupying a large number (> c.1000) of 10-km squares in both atlas periods. All these species are on the Red List (high conservation concern) of the Birds Species of Conservation Concern in the United Kingdom, Channel Islands and Isle of Man (Gibbons et al. 1996), with the exception of the lapwing, which is on the Amber List (medium conservation concern). The grey partridge population has declined largely as a result of decreased chick survival due to increased pesticide usage (Potts & Aebischer 1995), whereas the lapwing is likely to have declined due to a loss of mixed arable–grass rotations (Hudson, Tucker & Fuller 1994). The loss of spring tillage and the consequent decline in the area of winter stubbles have been implicated in the decline of the corn bunting (Donald 1997). Analyses of ringing data show that declines in annual survival of reed buntings, possibly brought about by a loss of winter food, are sufficient to explain the population decline (Peach, Siriwardena & Gregory 1999). The reasons for the decline of the other two species are poorly understood, although both are likely to have been affected by agricultural change (Fuller et al. 1995).

The link between population decline and distributional change is more complex than simply the disappearance of populations from areas where habitats have been most altered; most farmland bird species have disappeared from the margins of their range in the west and north of Britain, but certain of the agricultural changes implicated in their population declines have been most severe in the south and east, where many remain most abundant (Chamberlain & Fuller 2000). For example, Chamberlain & Fuller (2000) found that local extinctions of farmland birds and declines in overall avian diversity were most pronounced in areas showing the fewest changes in crop type.

Hill (1991) used a multivariate ordination technique to model the distribution of bird and plant species along environmental axes, and we examined distributional changes using similar methods. We started by testing the hypothesis that local extinctions have occurred primarily in inherently less suitable environments. This could have implications for determining the causes of decline and could have relevance to the practical conservation of these species because it could identify areas where conservation efforts should be concentrated. We developed this method to predict which areas of the UK are most or least likely to suffer local extinction if range declines continue. The second hypothesis we tested was that losses of each of the six species were not independent, i.e. that they have tended to be lost from the same squares. If these species have similar ecological requirements, poor environments for different species might coincide or local environmental changes may have similar adverse effects on several species. Alternatively, if losses of each species occur independently in response to different local environmental changes, or if ecological differences mean that peripheral populations of each species are in different places, independent patterns of loss would occur. The detection of independence or non-independence in patterns of decline between species could elucidate the reasons behind the declines and could help to identify key areas for conservation.


Sources of data

The British and Irish breeding distributions of all bird species are known from two atlases, the first carried out between 1968 and 1972 (Sharrock 1976) and the second between 1988 and 1991 (Gibbons, Reid & Chapman 1993). The presence or absence of each breeding species was recorded in each 10 × 10-km grid square (hereafter referred to as 10-km squares). For convenience, the distributions of breeding birds recorded in each atlas are referred to as the 1970 distribution and the 1990 distribution throughout. By comparing the two distributions, it was possible to identify, for each species, four classes of 10-km square:

  • 1. 10-km squares that were occupied in both atlas periods (retained);
  • 2. 10-km squares that held birds in 1970 but not in 1990 (abandoned);
  • 3. 10-km squares that were occupied in neither atlas period (never present; although the species may have occupied these squares prior to 1970 but become extinct before the first atlas period);
  • 4. 10-km squares where the species was absent in 1970 but present in 1990 (gained).

Because we selected species on the basis of their severe distributional declines, there were few gained squares and they were discounted in some analyses. Ireland was excluded from the analyses because environmental data were not available at the same scale as the bird data. This left a total of 2641 10-km squares in the analyses falling into one of the four categories described above.

Environmental data, referenced to the same 10-km grid scale as the bird distribution data, were obtained from several sources. The data related to the period of the first breeding atlas (1968–72), although most of them were topographical, pedological and climatic variables that would apply to both atlas periods. Descriptions of the data, their units and sources are given in Table 1.

Table 1.  Environmental variables used in the principal component analysis
  1. Key to data sources: 1, Ball, Radford & Williams (1983); 2, MAFF/DAFS June agricultural census for 1969, converted to 10 × 10-km square grid by Edinburgh University data library; 3, J.R.G. Turner & J. Lennon, Leeds University Genetics Department. Units of variables: prop., proportion of the land area of the 10-km square; m, metres; mm, millimetres; °C, degrees Celsius; n, frequency.

Agricultural land class 1prop.1
Agricultural land class 2prop.1
Agricultural land class 3prop.1
Agricultural land class 4prop.1
Agricultural land class 5prop.1
Difference in altitude between highest and lowest pointsm1
River frequency scoren1
Lake areaprop.1
Average altitude (above sea level)m1
Area of woodlandprop.1
Area of urban landprop.1
Area of crops and fallow landprop.2
Area of grasslandprop.2
Area of rough grazingprop.2
Area of cereal cropsprop.2
Area of non-cereal cropsprop.2
Area of horticultural cropsprop.2
Area of vegetable cropsprop.2
Average spring temperature°C3
Average summer temperature°C3
Average autumn temperature°C3
Average winter temperature°C3
Average annual rainfall 1941–70mm1

Statistical analysis

We assumed in the analyses that environmental axes correlated with presence/absence could be treated as measures of environmental suitability. This assumption was justified for the species studied here, as there was no indication that they were excluded from otherwise suitable habitat by persecution or other factors. All of them are widespread in Britain and have almost certainly colonized all suitable areas at some point in the past, although some populations have subsequently become extinct (Holloway 1996).

We used principal component analysis (PCA) to summarize the overall environmental attributes of the 10-km squares. The use of PCA was considered appropriate because the suitability of an area for a particular species is likely to be determined by a combination of many environmental factors (Harrison, Murphy & Ehrlich 1988; Hill 1991; Lawton & Woodroffe 1991; Osborne & Tigar 1992; Buckland & Elston 1993; Fuller, Trevelyan & Hudson 1997). The environmental variables relating to the first atlas period (given in Table 1) were summarized into a small number of PCA axes, and the axis scores for each 10-km square were used in subsequent analyses, which compared the environmental attributes of the four classes of square (retained, abandoned, never present and gained). The axis scores for each 10-km square were then related to the presence or absence of each species in 1970 and the loss or retention of each species between 1970 and 1990 using binary logistic regression (Manel et al. 1999). This allowed us to assess whether the PCA environmental gradients best determining whether a species was present or absent in 1970 also best predicted its loss between 1970 and 1990. If this were found to be the case, it would support the argument that losses occurred primarily from less suitable areas. Non-linear relationships were allowed for by including quadratic terms in the set of explanatory variables. Multiple regression models were selected using backward elimination of variables, with α = 0·001 being the criterion for retention of a term in the model. This stringent critical value of α was thought to be appropriate to the very large number of replicates available in the analyses. Checks on the models thus produced, by iterative addition and subtraction of variables, showed that, in every case, backward elimination produced a model in which all the terms were significant at P < 0·001 and no other variables had an additional effect. A backward elimination process ensured that all quadratic terms were entered with their respective unit terms. Analyses were carried out using SAS (SAS 1990) and GLIM (Crawley 1993) statistical software.

Model parameters from the best retention/loss model for each species were plotted spatially using the mapping package DMAP. The maps showed, for each species, the predicted extinction likelihood in all 10-km squares still occupied by that species in the 1988–91 atlas. These maps can therefore be used to identify regions of Britain where future local extinctions of each species are likely.

To test whether losses occurred principally from less suitable environments (i.e. those most similar to areas where the species was never present), two methods were used. First, we calculated the correlation between the percentage occupancy of a land use type (assumed to be a measure of environmental suitability) and the probability of distributional losses occurring from it. This was done by determining the bivariate model best explaining loss–retention between 1970 and 1990 (all possible models using two PCA axes with eigenvalues > 1 were examined). Each of the four classes of 10-km square described above were plotted in the two-dimensional space defined by these PCA axes. This two-dimensional space was then divided into equal-sized PCA regions measuring an arbitrary 0·5 by 0·5 PCA units. Weighted Pearson correlation coefficients were calculated between the proportion of 10-km squares in each PCA region in which each species was present in 1970 and the proportion of these from which the species was lost by 1990. PCA regions in which the species was not present in any 10-km squares were excluded from the analysis. All proportions were subjected to angular transformation prior to analysis. This effectively gave a correlation between the percentage occupancy of a land use type (assumed to be a measure of environmental suitability) and the probability of distributional losses occurring from it (‘gained’ squares were excluded from this analysis). This test could be extended to use more than two principal axes, but this was not done here as the results of the two-dimensional tests were sufficiently clear. The second test of the hypothesis used anova to test for significant differences between the three categories of 10-km square (‘gained’ squares were again excluded) in the PCA axis scores of the two axes (i.e. those explaining the greatest amount of variance in the environmental data set). An overall anova was calculated in each case, followed by post hoc pairwise tests between the three classes of 10-km square.

To test whether loss of the six species from individual 10-km squares was independent or not, the pattern of loss and retention of each species was compared with that of each other species (making a total of 15 pairwise comparisons). The area of distributional overlap between each pair of species was determined; this was the set of 10-km squares in which both species were present in 1970. The number of these squares that contained both species, one species alone and neither species in 1990 was then determined. This was compared with the expected number of each type of outcome expected under the null hypothesis, that squares from which each species was lost within their area of overlap were distributed independently. Binary logistic regression was used to determine whether there was a relationship between the likelihood of a species being lost from a particular 10-km square and the number of the other five species present in that square in 1970.


Environmental gradients

The first six principal components together explained nearly 80% of the variance in the environmental data, with the first two alone accounting for more than 56% (Table 2). The first principal component represented an axis from upland to lowland 10-km squares, and the second a transition from agriculture dominated by grazing to arable land. The others were hard to interpret ecologically, although ecological interpretation was not necessary to the testing of the hypotheses under question.

Table 2.  Eigenvectors of the first two principal components (PRIN1, PRIN2)
  1. Key to variables: ALC1–5, agricultural land class 1–5; HTDIFF, altitudinal range; RIVER, river frequency; LAKE, lake area; AVALT, mean altitude; AVRAIN, average annual rainfall; FOREST, area of woodland; URBAN, urban area; CROPFAL, area of crops and fallow land; TOTGRASS, area of grassland (excluding rough grazing); ROUGH, area of rough grazing; TCEREAL, area of cereals; TNONCER, area of non-cereal crops; TOTHORT, area of horticultural crops; TOTVEG, area of vegetable crops; WINTEMP, winter temperature; SPRTEMP, spring temperature; SUMTEMP, summer temperature; AUTTEMP, autumn temperature. % Var. gives the cumulative percentage of variance explained. For ease of interpretation, axis scores with an absolute value > 0·2 are in bold type.

% Var.41·856·4

Effects of environment on distribution and distributional change

Logistic regression showed that the presence or absence in 1970 of each species was significantly related to at least three of the six principal environmental axes in univariate regressions, although no two species were related to exactly the same set (Table 3). Similarly, the loss or retention of each species between 1970 and 1990 was related to at least one of the principal axes. For five of the six species the axis that best predicted loss was the same as that best predicting presence or absence in 1970. The exception was grey partridge, for which the first principal axis best predicted presence or absence in 1970, but the third axis best predicted loss. Multiple logistic regression models of presence–absence in 1970, and loss–retention between 1970 and 1990, differed between species both in the axes selected and the signs of the coefficients (Table 3). In most cases the first axis was the best predictor of both presence–absence in 1970 and loss–retention when assessed by the increase in residual deviance when deleted from the model. The two exceptions were for the loss–retention of grey partridge, for which the third axis again had a slightly larger predictive power, and the presence–absence in 1970 of reed buntings, where the fourth axis dominated. These models provided good discrimination between the different classes of 10-km square, with a minimum of 77·2% being classified correctly (Table 3), and therefore provided an adequate description of the data.

Table 3.  Multiple binary logistic regression models of presence–absence and loss–retention of each species on the first six principal components (P1–P6). The models relate the logit of the response variable [ln(p/1 − p)] to the explanatory variables. The column headed % gives the percentage of squares correctly allocated to each category by each model
Grey partridge89·33·40 + 2·40(P1) − 0·46(P1)2 + 1·05(P2) − 1·52(P3) + 0·39(P4) − 0·20(P4)2−0·91(P6) − 0·21(P6)2
Lapwing93·15·12–1·19(P1)2 + 1·44(P2) − 1·31(P3) − 0·41(P4) − 0·84(P6)
Turtle dove85·7−0·72 + 2·81(P1) + 0·96(P1)2−0·46(P2) − 0·23(P3) − 0·36(P3)2 − 0·91(P6) − 0·23(P6)2
Tree sparrow82·11·87 + 1·64(P1) − 0·69(P1)2 + 1·03(P2) − 0·71(P3) − 0·43(P3)2 − 0·60(P5) − 0·43(P6)
Reed bunting91·03·97 + 1·11(P1) − 0·46(P1)2 + 1·62(P2) + 0·75(P4) − 0·24(P4)2 − 0·37(P6) − 0·15(P6)2
Corn bunting79·60·66 + 1·39(P1) − 0·77(P1)2 + 0·75(P2) − 0·70(P3) − 0·25(P4) − 0·35(P5)
Grey partridge82·41·84 + (P1) − 1·57(P1)2 + 0·82(P2) − 0·91(P3) − 0·27(P3)2−0·47(P5) − 0·34(P6)
Lapwing89·53·89–1·13(P1)2 + 1·14(P2) − 0·63(P3) − 0·51(P5) − 0·18(P5)2−0·65(P6)
Turtle dove82·7−1·91 + 4·40(P1) − 1·03(P3) + 0·47(P4) − 1·25(P6)
Tree sparrow81·61·51 + 1·38(P1) − 1·13(P1)2 + 0·87(P2) − 0·69(P3) − 0·50(P5) − 0·16(P5)2−0·29(P6)
Reed bunting85·72·51 + 0·55(P1) − 0·55(P1)2 + 0·51(P2) − 0·22(P6) − 0·10(P6)2
Corn bunting77·2−0·18 + 1·68(P1) + 0·31(P2) − 0·49(P4) − 0·13(P5)2

For each species there was a highly significant negative correlation between the proportion of occupied squares in each region of PCA space and the proportion of squares from which that species was lost (Table 4). This indicated that decreasing suitability of environment was associated with increasing probability of local extinction. The anova results demonstrated significant differences between the three classes of square (abandoned, retained and never present) in terms of land use and environment (Fig. 1 and Table 5). For each species there were significant differences in PCA scores of the first two environmental axes between the three classes of square (Table 4), and in most cases abandoned squares were intermediate between never present and retained squares (Fig. 1). Axis scores of the first axis of environment differed significantly in all pairwise tests, with the exception of the test of abandoned and never present squares for lapwing. At least one significant difference for each species was found in pairwise tests of the second axis of ordination (Table 5), although the overall model was not significant in the case of lapwing. In each case, retained squares were at the upper limit of the first axis of ordination, suggesting that populations in lowland squares were more resistant to local extinction than those in upland squares. In most cases, retained squares were also at the upper limit of the second axis of ordination, suggesting that populations in arable areas were more resistant to local extinctions than those in grass-dominated areas.

Table 4.  Correlation coefficients between the arcsine square-root transformed proportions of squares in each region of a two-dimensional PCA space (defined by the two axes best explaining loss–retention) occupied by a species and the proportion of occupied squares from which a species was lost, weighted by the number of squares in each PCA region. All correlations were significant at P < 0·0001
SpeciesPCA axes usedCorrelation coefficientNumber of PCA regions
Grey partridge1,3−0·8990
Turtle dove1,6−0·9573
Tree sparrow1,3−0·8183
Reed bunting1,2−0·7965
Corn bunting1,4−0·8688
Figure 1.

Plots of means of axis scores on first two axes of ordination (means with SE) of four classes of 10-km square (RT, retained; GA, gained; NP, never present; AB, abandoned). Significance of pairwise differences are given in Table 5.

Table 5. anovas comparing PCA scores of the first two axes of ordination (PRIN1 and PRIN2) between abandoned (AB), never present (NP) and retained (RT) 10-km squares. Sample sizes differ slightly due to differing numbers of squares in the category gained, excluded from this analysis. The overall F-value is given followed by levels of significance for pairwise tests between the three classes of 10-km square. *  P < 0·05, **  P < 0·005, ***  P < 0·0001. Values of all categories are plotted in Fig. 1
SpeciesOverallRT vs. ABRT vs. NPAB vs. NP
Grey partridgeF2,2587 = 686·1************
LapwingF2,2591 = 49·1*********NS
Turtle doveF2,2602 = 1342·0************
Tree sparrowF2,2542 = 761·4************
Reed buntingF2,2557 = 184·1************
Corn buntingF2,2554 = 829·2************
Grey partridgeF2,2587 = 23·2***********
LapwingF2,2591 = 2·9 NS*NS 
Turtle doveF2,2602 = 100·3************
Tree sparrowF2,2542 = 23·1******NS***
Reed buntingF2,2557 = 3·4*NS*NS
Corn buntingF2,2554 = 62·6************

The model parameters of the loss–retention models were plotted for each square where the species still occurred during the 1988–91 atlas (Fig. 2). Generally 10-km squares in the north and west of the six species' British ranges had higher extinction probabilities than squares in the south and east, although the models identified other regions with a high probability of local extinctions. For example, a number of species were predicted to have high extinction likelihoods along the south coast of England. Some striking interspecific differences in the distribution of extinction-prone and extinction-resistant squares also emerged. For example, the grey partridge was predicted to be at a high risk of local extinction in the fens of Lincolnshire, whereas tree sparrow and corn bunting were predicted to be at lowest risk of local extinction there. Generally, the distribution of extinction-prone squares matched the distribution of low territory densities given in Gibbons, Reid & Chapman (1993), and the distribution of squares with low extinction likelihoods with areas of high territory density.

Figure 2.

Maps showing the predicted extinction likelihood of each species within the range occupied in 1990. The extinction probability was derived from the models best predicting the loss or retention of each species between 1970 and 1990. For ease of interpretation, data were broken down such that each of the four dot sizes occurs in equal numbers on each map. Therefore the four dot sizes denote different extinction likelihoods for each species. In descending order of dot size, the ranges of extinction probabilities for each species are:
inline image

Between-species dependence of local extinctions

For all combinations of species pairs, more 10-km squares lost and retained both species and fewer squares lost just one species than would have been expected if local extinctions were independent. Considering just the area of distributional overlap between each species pair, there was a tendency for losses to occur from the same 10-km squares. In all cases, the observed number of 10-km squares from which neither or both species were lost was greater than would be expected by a random pattern of loss, and the observed number of squares where one of the species was lost and the other retained was fewer than expected (χ2 values with 3 d.f. ranged from 14·2 to 172·2, P < 0·005 in all cases).

For each pairwise comparison, the number of squares from which each species was lost from within and outside their area of distributional overlap was also calculated. This showed that there was a consistent pattern of a higher probability of loss from 10-km squares where each species occurred alone than from where they occurred together (χ2 tests with 3 d.f., P < 0·005 in all cases). Thus, although where two species occurred together they tended to be lost from the same squares, there was a relatively higher rate of loss of each species from squares where it occurred alone. For each species, there was a strong negative relationship between the likelihood of a species disappearing and the number of the other five species originally present in each 10-km square (binary logistic regression, −2 log likelihood statistics: grey partridge 262·1, lapwing 216·3, turtle dove 174·8, tree sparrow 231·9, reed bunting 221·1, corn bunting 182·1; all significant at P < 0·0001).


Relationships between distribution and environmental gradients

The environmental suitability of any particular area is probably determined by several environmental variables (Brown 1984; Maurer & Brown 1989), so combining many variables into a few summary gradients by multivariate techniques is a sensible way of classifying sites environmentally. This method also has the advantage of reducing problems of collinearity between large numbers of predictor variables (Donald & Fuller 1998). The regression models based on PCA scores discriminated well between occupied and unoccupied 10-km squares, and between those in which each species was lost and retained between the two atlas periods, showing that the principal component axes reflect some measure of environmental suitability. That the axes best explaining presence–absence were the same as those best explaining loss–retention suggests that the approach is robust temporally as well as spatially. However, these models provide little information on the relative importance of the different individual environmental factors, and some variables irrelevant to avian distribution probably contribute to the principal components. Attempting to discover which are the crucial variables by using the raw environmental data rather than the principal components derived from them is difficult. There are likely to be other important environmental factors that are not included in the data set, but variables correlated with these factors may be good predictors of distribution (Gates et al. 1994). Data reduction methods, such as ordination, appear to be a more pragmatic approach to assessing large scale patterns of bird distribution (Chamberlain & Fuller 2000).

For the six species examined here, Fig. 1 suggests that local populations were most resistant to local extinction in arable areas, and in four of the six species 10-km squares where the species was gained between 1970 and 1990 lay further along the arable axis than squares where the species was lost. This supports the findings of Chamberlain & Fuller (2000) that losses of these species have been greater in grass-dominated areas. For all six species, extinction probabilities were greater in the western regions of their ranges, where grass-dominated agriculture prevails, although for several species the maps of extinction likelihood (Fig. 2) identified an area of high risk in coastal regions of south-east England, an area of pastoral farming in an otherwise arable part of the country.

Mechanisms of distributional change

The results of this work suggest that local extinctions occurred primarily in the least favourable parts of their range. This is consistent with several possible explanations. In general, two types of factors will determine what distributional changes will occur in response to environmental change: the nature of the environmental change (whether it affects few localities or habitats, or applies to the whole range), and the structure of the species' distribution (the spatial arrangement of good and poor habitats and the levels of migration between them).

It has been suggested (Pulliam 1988; Pulliam & Danielson 1991) that emigration from productive habitats into surrounding poor quality habitats leads to the existence of productive source populations and associated sinks. On a larger geographical scale, productive source areas in the core of a species' range may support sink populations in poor quality habitats at the periphery (Lawton 1993). The most favourable habitats tend to be found in the centre of the range, away from which conditions deteriorate (Brown 1984; Maurer & Brown 1989), although the existence of sources and sinks is controversial and difficult to prove (Watkinson & Sutherland 1995; Dias 1996). However, it is possible that environmental changes in source areas may affect distant populations in sinks, and conversely that populations in sink areas may be relatively unaffected by local environmental changes, as long as the productivity of sources is unaffected. This may be one of the explanations for the extinction of farmland bird populations in areas where certain agricultural changes have been least severe.

Relationships between population density and distribution

In the analyses reported here, we have ignored changes in population density. There were two reasons for this; first, data on the change in abundance of each species within each 10-km square do not exist; and, secondly, extinction of a species in a 10-km square is particularly interesting from a conservation point of view. Nevertheless, changes in population density are a response to environmental change, and at present our understanding of their links to distributional changes is incomplete. It has been suggested, based on metapopulation models (Lawton et al. 1994), that the number of sites a species occupies and its population density at those sites may be directly connected, so that a change in one of these parameters will lead to a change in the other. Whether this is so, and whether it applies over large geographical areas, is unknown, although the true situation is likely to be considerably more complex than can be modelled. There are examples of increase and decrease of population density with range contraction (Wilcove & Terborgh 1984; Gaston 1994; Donald & Fuller 1998); it is even possible for an increase in overall population size to occur at the same time as a decline in the distribution, as in the European nightjar Caprimulgus europaeus L. in Britain (Morris et al. 1994).

Gaston (1994) considered the possible ways in which population density could change with range size; as a species distribution shrinks, average population density may decrease if there is a disproportionately high loss of core (high density) populations, or it may increase if losses occur mostly from the periphery. For the six farmland bird species studied here, distributional losses have been concentrated in peripheral areas, so an increase in overall population density of occupied squares might be expected. However, the Common Birds Census (Marchant et al. 1990), which covers predominantly the lowland areas of Britain in which these species are most abundant, has shown severe declines of all six species. Thus, it appears that population density in core habitats has declined as the ranges have shrunk. The reason for the discrepancy between these results and Gaston's (1994) models is probably that the loss of peripheral populations represents a relatively small loss of individuals; the bulk of the population losses have occurred in core areas, but have not been sufficient to cause local extinctions. Moreover, other farmland species such as linnet Carduelis cannabina L. and skylark Alauda arvensis L. are also suffering declines in population density, but their ranges have not yet contracted appreciably (Gibbons, Reid & Chapman 1993). It appears that for these farmland birds, population losses have occurred faster than distributional losses, and population declines in the core of the range have preceded distributional contraction. Previous studies of atlas data have revealed similar patterns (Donald & Fuller 1998). Comparison of the predicted extinction likelihood estimates shown in Fig. 2 with the breeding density maps given in Gibbons, Reid & Chapman (1993) suggests that, generally, there is a negative relationship between breeding density and extinction likelihood.

Spatial associations of distributional changes

Spatial patterns of loss were not independent. The likelihood of local extinction was greater outside the area of overlap of any two species' ranges than within. Furthermore, within areas of overlap of two species, local extinctions tended to occur in the same squares. A suggested interpretation of these results is that the most suitable areas are common to several species, but peripheral areas are occupied by fewer species. Environmental preferences for these species may be broadly similar, but each one may be able to tolerate environmental conditions in peripheral areas to varying degrees. Hence they tend to coincide in good habitats, but to occur alone, or only with one or two other species, in the least suitable environments, and hence in the areas where pairs of species' distributions did not overlap. The high incidence of cases where both species were lost within the area of distributional overlap could be due either to the occurrence of both species in a poor quality habitat, or to degradation of a high quality habitat, rendering it uninhabitable by either species.

Spatial autocorrelation and prediction of distributional changes

Environmental variables probably exhibit positive spatial autocorrelation; 10-km squares tend to resemble their neighbours more than they resemble squares further away. Thus squares that are close together on a map are likely to be close together in PCA space as well. The associations in PCA space between occupied squares and those where a species was retained are therefore likely to be reflected in similar associations in real space. Such a correlation was found by Donald & Greenwood (in press). Thus the distributional losses of these birds have appeared as contractions from the range edges, which is where the poorest environments are concentrated. Other species, however, show no such association of distributional losses with a range boundary, two good examples being the grasshopper warbler Locustella naevia L. and barn owl Tyto alba L., both of which have disappeared from large numbers of squares that are apparently scattered randomly throughout their range in Britain (Donald & Greenwood, in press). It is possible that good habitats for these species may be much more fragmented than for farmland birds, with lower spatial autocorrelation, and that their losses have been from the intervening poor quality habitats. This hypothesis could be tested by the methods presented here.

Implications for conservation

A high proportion of declining farmland bird species in western Europe (including those examined here) remains relatively common and widespread, despite qualifying for urgent conservation priority because of severe population declines (Tucker & Heath 1994). Site-orientated conservation methods are therefore inappropriate, and resources available for the conservation of common farmland species are currently most effectively distributed through agri-environment incentives (Pain & Pienkowski 1997). Such schemes could be used to conserve declining farmland bird populations by improving conditions within the existing range of these species in order to maintain the current population and range, or by improving conditions in areas from which the species have disappeared in order to try to re-establish former ranges. Our results suggest that, although each species has its own particular habitat requirements, areas of farmland suitable for one species tended to be suitable for other species and that areas abandoned were inherently less suitable than areas where the species were retained. Therefore, in the case of farmland birds, concentrating conservation efforts in regions where a high proportion of the species still occur is likely to be more effective than directing resources to less suitable areas on the peripheries of the species' former ranges. This is likely to be true for countries with a long history of human modification of habitat, but in countries where human influence on populations is more recent, range peripheries may support important populations that require conservation (Channell & Lomolino 2000). A conservation strategy will be most successful if it maintains or improves the best habitats, rather than attempting to maintain populations in poor habitats while allowing the key areas to deteriorate. Not only are the best areas conserved, but increases in numbers within such areas are likely to help restore peripheral populations.

The methods presented above can be used to target areas of conservation priority. By plotting estimates of extinction likelihood, as we have done in Fig. 2, it is possible to identify 10-km squares that, if declines in distribution continue to occur, would be most or least likely to suffer local extinction. This could be used not only to identify areas of high conservation priority but also to assess the importance of future distributional change. The extinction of one or more species in a 10-km square with a low extinction probability is likely to be of greater conservation concern than extinction in a 10-km square with a high extinction probability. Such an approach has yet to be implemented as a practical conservation tool. The predictive power of this method can, however, only be tested when the results of future atlases are available.

Despite the potential applications of the methods described above to conservation management, it is important to remember that the underlying causes of decline are unlikely ever to be understood in any detail from studies carried out at the large geographical scale of atlas studies (Donald & Fuller 1998); atlas data merely allow an assessment of where declines are taking place. Nevertheless, we believe that methods such as those described above can elucidate the process of distributional decline and be used to identify priority areas for conservation at a spatial scale appropriate to national agri-economic incentives.


This work would not have been possible without the efforts of the thousands of volunteers who took part in the two British breeding bird atlas surveys. This research was partly funded by the Natural Environment Research Council, the Royal Society for the Protection of Birds and the Joint Nature Conservation Committee (on behalf of English Nature, Scottish Natural Heritage, Countryside Council for Wales and the Department of the Environment Northern Ireland). We are grateful to John Turner and Jack Lennon (Leeds University), Alison Bayley and Peter Burnhill (University of Edinburgh Data Library) and the Institute of Terrestrial Ecology for providing the environmental data. DMAP was written by Dr Alan Morton and the maps produced by Simon Gillings. The grid square agricultural data were adapted from the Parish Summaries of the Agricultural Census for England and Wales and for Scotland. We thank Dr Rob Fuller, Dr David Gibbons, Professor John Lawton, the editor and two anonymous referees for valuable discussion and comments on the manuscript.

Received 14 August 1998; revision received 7 April 2000