Should conservation strategies consider spatial generality? Farmland birds show regional not national patterns of habitat association

L E T T E RShould conservation strategies consider spatial

generality? Farmland birds show regional not

national patterns of habitat association

Mark J. Whittingham,1* John R.

Krebs,2 Ruth D. Swetnam,3 Juliet

A. Vickery,4 Jeremy D. Wilson5

and Robert P. Freckleton6

Abstract

A key assumption underlying any management practice implemented to aid wildlife

conservation is that it will have similar effects on target species across the range it is

applied. However, this basic assumption is rarely tested. We show that predictors [nearly

all associated with agri-environment scheme (AES) options known to affect European

birds] had similar effects for 11 bird species on sites with differing farming practice

(pastoral vs. mixed farming) or which differed in the density at which the species was

found. However, predictors from sites in one geographical region tended to have

different effects in other areas suggesting that AES options targeted at a regional scale

are more likely to yield beneficial results for farmland birds than options applied

uniformly in national schemes. Our study has broad implications for designing

conservation strategies at an appropriate scale, which we discuss.

Keywords

Conservation ecology, farming and wildlife, habitat selection, information-theoretic

modelling, spatial scale.

Ecology Letters (2007) 10: 25–35

I N T R O D U C T I O N

Many habitats are heavily managed for conservation because

the ecosystems in which they are found have been greatly

altered by human activities over a long period of time

(Sutherland 1998). A key assumption of many conservation

management practices is that the same action will have

similar effects wherever it is applied. This is equivalent to

assuming that organisms exhibit the same patterns of

preference or avoidance across the range of interest.

However, this assumption is rarely tested.

Several approaches to analysing species� responses to

management exist (e.g. experimental and observational), but

here we concentrate on habitat-association modelling.

Habitat-association modelling relates the occurrence of

organisms to the presence and amounts of key attributes of

the landscape (Rushton et al. 2004; Whittingham et al. 2005;

Meyer & Thuiller 2006). By using data from sites located in a

range of different geographical areas, the type of manage-

ment employed and the levels of key variables, robust

descriptions of species� distributions can be developed. On

the assumption that the observed relationships reflect cause

and effect, the models can then be used to predict whether

changing management will improve habitat quality.

One of the advantages of using this approach is that it

should be possible, given a large enough data set collected at

1Division of Biology, School of Biology and Psychology, Ridley

Building, University of Newcastle, Newcastle-Upon-Tyne NE1

7RU, UK2Zoology Department, University of Oxford, South Parks Road,

Oxford OX1 3PS, UK3Centre for Ecology and Hydrology, Monks Wood, Abbots

Ripton, Huntingdon, Cambridgeshire, UK4British Trust for Ornithology, The Nunnery, Thetford, Norfolk

IP24 2PU, UK

5Royal Society for the Protection of Birds, Scotland Headquar-

ters, Dunedin House, 25 Ravelston Terrace, Edinburgh EH4 3TP,

UK6Department of Animal and Plant Sciences, University of Shef-

field, Sheffield S10 2TN, UK

*Correspondence: E-mail: [email protected]

Ecology Letters, (2007) 10: 25–35 doi: 10.1111/j.1461-0248.2006.00992.x

� 2006 Blackwell Publishing Ltd/CNRS

the appropriate scale, to test the generality (robustness) of

model results across geographical locations. In the context

of predicting occurrence as a function of key habitat

variables, several elements of generality are important. First,

there is the question of whether a model developed at one

time or in one place, will predict well elsewhere (irrespective

of whether the model is correct or the parameters estimated

accurately). Second, it is necessary to know whether the

habitat variables identified as important in one place are also

important elsewhere. Finally, even if the same habitat

variables are identified as important, whether the precise

effects and ranking of importance are the same.

These three aspects of the model are reflected in different

model measures. The total model fit (e.g. R2) of a model can

be used to measure how well a model derived in one place

performs in total elsewhere. This measures the proportion

of variance the model explains, indicating broadly whether

the model performs well or poorly, but does not dissect

model performance any further. More detailed insight can

be obtained by looking at model parameters. For instance,

models could be fitted to data from two areas, and the

values of the estimated parameters compared. It might

be the case that parameters are numerically identical.

Alternatively, it is possible that precise effects are not

the same, but the ranking of variables is still the same

(or even the direction of their effects the same). Recognizing

these different elements of fit is important, and suggests

several ways that models might be compared between

localities.

Here, we consider a case study of species� responses to

habitat management. We explore habitat selection by

farmland birds at a national scale. This group of birds is

of great current conservation interest given the dramatic

decline of European farmland birds (Krebs et al. 1999) and

the costs of agri-environment schemes (AES) (€24 billion

was spent by the European Union between 1992 and 2003;

Kleijn & Sutherland 2003) designed, in part, to benefit

wildlife on farmland. In agricultural systems the key

variables that will affect habitat associations of bird species

are those relating to geography and farm management

(Kleijn et al. 2004; Peach et al. 2004; Bishop & Myers 2005).

An example of the different effects of regional context is

provided by a study in England by Peach et al. (2004) that

showed marked differences in habitat selection by song

thrushes Turdus philomelos L. in eastern arable sites and those

in southern mixed farming areas. Management of agricul-

tural land can also have dramatic effects on bird distribu-

tions and hence habitat selection. For instance, Peach et al.

(2001) described land management agreements that led to

dramatic increases in cirl bunting Emberiza cirlus L.

populations when compared with similar patches outside

agreement areas: such changes are likely to have impacts on

habitat selection not least due to buffer effects (see below).

A second factor that will be important in determining

occupancy is underlying habitat quality (Freckleton et al.

2005, 2006). An important consequence of variation in

habitat quality is that species may initially preferentially

select high-quality sites, and subsequently move into low-

quality sites as the high-quality habitat is filled. This is

termed a buffer effect (e.g Gill et al. 2001). In terms of

habitat-association modelling this poses a problem, as

habitat associations will very likely change with density.

To address this requires that the fit of models is

compared between areas varying in the density at which

species occur.

In this article, we use an extensive data set based on

individual fields and field boundaries to examine habitat

associations of farmland birds, one key group of wildlife

targeted by AESs. We ask for the first time, on the scale at

which both the birds and AESs are operating, whether

farmland bird–habitat associations are general across larger

geographical areas, in areas with differing species density

and across different landscape types. We use these findings

to inform the question of whether AES implementation

needs to be more sensitive to the limits on generality. We

test the generality of models based on habitat variables, 11

of 12 of which are associated with AESs for European

birds (see Appendix S1). We show that predictors had

similar effects on bird distribution on sites with differing

farming practice (pastoral vs. mixed farming) or which

differed in the density at which the species was found.

However, predictors from sites in one geographical region

tended to have different effects in other areas suggesting

that the scale on which current AESs are targeted may not

be appropriate.

M E T H O D S

Birds and boundaries

Our data set is based on 42 study sites across lowland

agricultural land in England and Wales (Fig. 1) for 11 bird

species (namely blackbird Turdus merula L., blue tit Parus

caeruleus L., chaffinch Fringilla coelebs L., dunnock Prunella

modularis L., great tit Parus major L., greenfinch Carduelis

chloris L., linnet Carduelis cannabina L., reed bunting Emberiza

schoeniclus L., robin Erithacus rubecula L., song thrush,

yellowhammer Emberiza citrinella L.).

Data were collected in 2002 (mean area per site ¼70.8 ± 4.2 ha, 1 SE) in England and Wales (Fig. 1). These

sites were all located on farmland and were chosen by

observers; it is possible that the sites may have higher than

average densities of bird species (a common bias of sites

chosen by observers) but nevertheless they were widely

distributed across England and Wales and across a range of

different landscapes. Furthermore, the relationships we

26 M. J. Whittingham et al. Letter


sought were based on habitat selection not on species

density per se and so the influences of site selection should

not affect our modelling results.

All bird species were surveyed on boundary sections twice

per month from April to June (a minimum of six visits were

made to each site, range 6–12), using Common Birds

Census methods (Marchant et al. 1990). This methodology

gives an accurate standardized estimate of breeding birds on

each site and, because all sites are of the same broad habitat

type, there is unlikely to be significant bias in detection

probability between different sites. Thus the assumption in

our study is that there were no detection differences

between sites.

Counts of birds were made between 07:00 and

13:00 hours, but not in wet or windy (> force 4 on the

Beaufort scale) weather. The locations of all individuals were

mapped, and records from all censuses over the course of

the visits were collated. Territories were identified from the

spatio-temporal clusters of records using established meth-

ods (Marchant et al. 1990).

We compared the characteristics of occupied territories

with unoccupied territories. Unoccupied territories were

constructed using an algorithm based on the size

(i.e. number of boundary units observed from each species)

and a simple rule that boundaries > 100 m apart could not

be linked as part of the same unoccupied territory. An

unoccupied territory was a boundary or series of clustered

boundaries of similar size to observed territory sizes

(see Whittingham et al. 2005 for a full explanation of the

methods). We found that territory models (i.e. models in

which single or multiple boundaries formed the unit of

replication) gave parameter estimates that were very similar

to those derived from models based on treating each

boundary unit as a separate unit of replication; in this study,

we present results of models based on the territory scale.

The advantage of using models based on territories is that

territories better account for the spatial configuration of

resources (for discussion of boundary vs. territory models

see Whittingham et al. 2005).

Sufficient data existed to investigate patterns for 11

species (see above and Appendix S2a–c) (mean number of

territories per species across all sites ¼ 442.7 ± 98.6, mean

number of sites occupied ¼ 34.5 ± 2.1).

Information was also collected about every boundary

(n ¼ 3266) and the surrounding fields on each of these sites

in summer 2002 (see Appendix S1 for details of habitat

information collected). Boundary sections (sampling units)

were defined as any contiguous length of field boundary

between points of intersection with other boundaries

(all boundary sections were included in the analysis

irrespective of whether they were hedgerows or some other

feature, e.g. fence or wall). If the nature of the boundary

changed abruptly between intersections, it was further

subdivided into separate sampling units.

Statistical methods

Model fitting

We examined correlates of the probability of occurrence of

each species using a generalized linear model (presence or

absence of a territory along one or more sampling units,

assuming a binomial error distribution and a logit link, i.e.

logistic regression). The response variable was specified as

either an occupied territory (1) or an unoccupied territory (0).

We used the methods described by Burnham & Anderson

(2002). The approach compares the fits of a suite of

candidate models using Akaike’s information criterion

(AIC). The AIC is a measure of relative model fit,

proportional to the likelihood of the model and the number

of parameters used to generate it. The absolute size of the

AIC is unimportant, instead the difference in AIC values

between models indicates the relative support for the

different models. To compare models we calculated an

�Akaike weight�, wi, for each model. For a set of R models,

the wi sum to 1 and have a probabilistic interpretation: of

these models, wi is the probability that model i would be

selected as the best fitting model if the data were collected

again under identical circumstances.

11

1

2

22

1

3

1

3

2

1

1

2

1

3

2

3

11

1

1

1

11

3

1

2

1

1

3

1

1

3

1

2

1

1

11

3

1

Figure 1 Map of England and Wales showing the location of our

42 study sites. The sites were grouped in three different ways

[1, south-east (SE); 2, south-west; 3, north] (note that two SE sites

are very close together and cannot be distinguished).

Letter Conservation and generality 27


We calculated confidence sets of models fitted to each

data set. A confidence set is the smallest subset of candidate

models for which the wi sum to a given value, in this case

0.95. This set represents a set of models for which there is

95% probability that the set would contain the best

approximating model to the true model were the data

collected again under the same circumstances. It is

important to note that it is not the set with 95% probability

of containing the true model since we do not know that the

set of models considered actually contains the true model.

Because the wi are probabilities, it is also possible to sum

these for models containing given variables (Burnham &

Anderson 2002). For instance, if we consider some variable

k, we can calculate the sum of the Akaike weights of all the

models including k, and this is the probability that of the

variables considered, variable k would be in the best

approximating model were the data collected again under

identical circumstances.

Model averaging uses the average of parameter estimates

or model predictions from each candidate model, weighted

by its Akaike weight. For parameter bj the model-averaged

estimate was calculated as:

�bj ¼XR

i¼1

wi bþj ;i ; ð1Þ

where wi is the Akaike weight of model i, and bþj;i is the

estimate of bj if predictor j is included in model i or is zero

otherwise. Model-averaged estimates were compared with

estimates from a general linear model (GLM) including all

variables to assess the potential impact of model selection

bias on parameter estimates. The estimated selection bias for

parameter j was calculated as:

biasj ¼�bj � bGLM

�bj

��

��: ð2Þ

Prediction by model averaging using a set of GLMs is

complicated by the link function: unless an identity link

function is used, the predicted value for a given set of

predictors is not a linear function of the parameters, b. The

predicted value for given data is:

�l ¼XR

i¼1

wi liðxiÞ: ð3Þ

The model-averaged prediction (�l) is the weighted average

of the predicted values (l) of the R candidate models.

Although initially our data set contained many possible

predictors, we reduced this number to 12 based on a

literature review of the likely determinants of distribution

for these species (see Appendix S1). We then explored all

possible subsets of these 12 predictors as candidate models.

A variable coding for site was included in all models as a

fixed effect (although including it as a random effect made

no quantitative difference to the results), allowing large-scale

variation across the sites to be controlled for in every fitted

model. In our exploratory analysis we considered interaction

terms, however, did not find that these significantly

improved model fits. Moreover, we had no a priori reason

to expect interactions between the variables we included to

be important; therefore, we did not include any interaction

terms in the results reported below.

Stratification

We stratified our data in three different ways in order to test

the robustness of habitat-association models to variation in

three important broad-scale ecological factors: (i) farming

practice: we divided sites into two groups based on the

proportion of grassland on the site (21 sites with most grass,

mean proportion of area with grass ¼ 0.80 ± 0.02 and

0.15 ± 0.01 with arable crops; 21 sites with least grass, mean

proportion of area with grass ¼ 0.50 ± 0.04 and

0.38 ± 0.04 with arable crops); (ii) bird density: data were

split into two groups based on density for each of the 11

species (half of sites where species recorded at higher

density ¼ group 1; sites where species was not recorded

were excluded). All species were recorded at significantly

higher densities in group 1 than group 2 (t-tests, P < 0.01 in

all cases). On average across all species abundance was 3.45

times higher in group 1 than group 2 (see Appendix S2b).

(iii) Geographical location: sites divided into three groups,

south-east England, south-west England and Wales (SW)

and northern England (see Fig. 1).

We chose to stratify the data in these three ways because,

as outlined in the introduction, management, geographical

location and species density are all known to affect species

habitat-association patterns (see Introduction) and if AES are

to have general beneficial effects then they must produce

positive responses in differing areas. The alternative

approach to stratifying the data would have been to analyse

the data with factors coding for each of the habitat types,

along with interactions of these factors with the 12

predictors. The problem with doing this would have been

that the resultant model would have been complex, with a

large number of estimated parameters. For clarity, therefore,

we analysed the data in the stratified form.

For each stratification the analysis generated a net model

(using the information-theoretic method, this is formed

from a model-averaged set of parameters), with an estimate

of the regression coefficient for the effect of each predictor

on each bird species (i.e. 12 parameters · 11 species ¼ 132

parameters per stratum).

Model analysis

The standard approach for measuring model performance in

one area from a model derived elsewhere has been the



subject of much debate in presence/absence models with

many different criterion put forward to measure error with

kappa and receiver operating characteristic1 curves being the

most favoured at present (Manel et al. 2001). The problem

with such approaches is that they do not measure the

structural adequacy of models, and they do not allow the

three levels of model adequacy outlined in the fourth

paragraph of the introduction to be distinguished. Here we

employ model comparisons based on comparisons of model

parameters and on model fits at the scale of individual

farms: these are more appropriate measures of model

performance for the question in hand of determining how

well models developed in one area perform elsewhere at the

scale of whole farms.

We analysed the models and parameter estimates in

three ways. First, we measured the goodness of fit for each

model (at the scale of the 42 sites), measured as the

correlation between the fitted proportion of territories

occupied in each site and that observed. For each stratum

we then measured the correlation between the observed

occupancy and that �cross-predicted� by the model fitted to

the data from the others in that stratification. For example,

we compared the occupancy observed in the high-density

sites, predicted using the parameters of the model fitted to

the density of the low-density sites. This tests the ability of

the models from one stratum to predict occurrence in

another.

Second, for each stratum, we compared the values of the

parameter values fitted to each variable and species. We did

this through a GLM approach – the parameter values (i.e.

the 132 per stratum) were entered as the dependent variable

and the species and stratum category entered as factors.

Significant differences between strata, species and the 12

predictors then indicate quantitative structural differences

between the models fitted in each type or across species.

One criticism of this approach to the analysis is that we are

mixing model-fitting paradigms in our analysis, which is

cautioned against (e.g. Burnham & Anderson 2002).

However, the two types of analysis we employ are being

used for rather different tasks: the IT-AIC methods are used

to generate parameter estimates. The GLM approach is used

here to test differences between estimates in a fully factorial

design. We believe that there is nothing contradictory or

invalid in using these very different methodologies in this

way (e.g. see Stephens et al. 2005 for a detailed discussion of

these and related points).

Third, we conducted an identical analysis to that above,

but using the rank of the parameter values in each strata,

rather than the actual parameter values. This asks whether

the rank order of the strength of the predictors differs

between species and strata and hence whether there are

qualitative structural differences between the models fitted in

each stratum or across species.

R E S U L T S

In general, the overall fitted models provided a good

description for most species for all three different data

stratifications (Fig. 2), in some cases with the correlation

between observed and fitted being as high as 0.8–0.9. The

cross-predictions, as would be expected, provided lower

correlations. However, for the bird density and farming

practice stratifications the predictions were generally fair to

good, the correlations being as high as 0.8 in some

cases (Fig. 2). In contrast the cross-predictions from the

geographical stratification were much lower, in particular

with occurrence in the SW being predicted only poorly

based on models fitted in the other locations (Fig. 2b; see

Appendix S2c).

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Density

Low

Hig

h

Gra

ss

Ara

ble

SE

from

SW

SW

from

N

SE

from

N

Nfr

omS

W

Nfr

omS

E

FarmingType Geography

SE

from

SE

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Low High Grass Arable SE SW NDensity FarmingType Geography

Cor

rela

tion

betw

een

obse

rved

and

fitte

d

Cor

rela

tion

betw

een

obse

rved

and

pred

icte

d

(a)

(b)

Figure 2 Correlations of fitted parameter estimates from habitat-

association models derived in each strata (e.g. bird species density –

high or low) and observed values. (a) Results of correlations

between observed data from strata �A� and predictions derived

from a model derived using data from strata �A�, e.g. predicted

values from a model derived from grassland sites explained 69% of

the variation in observed data from those same grassland sites. (b)

Predictions from models derived in strata �A� and correlated with

observed data from strata �B�, e.g. a model derived from low-

density sites explained 35% of the observed variation on high-

density sites.



The generally good predictive ability of the models for

the farming practice and bird density stratifications is

reflected in generally close quantitative structural similarity

of models between strata. Although there is variability in the

precise strength of the effects of different predictors

between species and predictors, these effects do not differ

systematically between farming practice or bird density

strata (see interaction terms shown in bold in Table 1a,b).

This quantitative structural similarity was also reflected in

qualitative similarity between the models in the different

strata (Fig. 3b,d).

Conversely, we found that the models fitted to the

geographical strata yielded no evidence of either qualitative

or quantitative structural similarity (Fig. 3e,f; Table 1c).

Although there is a broad-scale positive correlation between

parameter values and their ranks, these tend to be rather

weak, with large outliers. Consequently, our analysis

indicates that models fitted in one geographical area tend

to be unrepresentative of those fitted in other areas.

Several predictors that are currently included within AES

were generally found to be positively related to species

occurrence namely: taller, wider hedges; boundary strips;

trees within boundaries (see Appendix S2a,b).

D I S C U S S I O N

The question of how to determine whether a model predicts

species occurrence well or not has been the subject of a

huge number of studies (e.g. see Rushton et al. 2004 for a

recent review). To some extent the focus of this literature

has tended to be on statistical and methodological aspects of

the problem. In this study, we have taken a more practical

perspective on the issue of model fit. In our application, we

have focused on several specific aspects of model fit and

performance. We have taken problems of understanding

how management affects distributions of individuals and

used these to define different criteria for judging model

performance based on both absolute and structural aspects.

By concentrating not only on model fit in the absolute

sense, but also examining the parameter estimates for

models fitted to data from different places, we are able to

examine in detail how robust models are, and how variable

Table 1 General linear models testing the

generality and influence of species identity

and predictor identity on parameter esti-

mates from habitat association models of

territories (which could encompass one or

more boundary sections, see Whittingham

et al. 2005 for further details, of 11 different

farmland bird species) fitted to data split into

three different strata: (a) different farming

areas; (b) sites with differing species density;

and (c) sites from different geographical

regions

Variable d.f. Deviance

Residual

d.f.

Residual

deviance F P (> F)

(a)

Null 248 293.757

Species 10 6.446 238 287.311 0.9766 0.469

Predictor 11 68.104 227 219.208 9.3801 <0.001***

Farming type 1 0.970 226 218.237 1.4700 0.228

Species · predictor 107 137.947 119 80.290 1.9533 <0.001***

Predictor · farming type 11 11.898 108 68.392 1.6387 0.100

Species · farming type 10 3.709 98 64.684 0.5619 0.841

Error 98

(b)

Null 252 342.26

Species 10 17.79 242 324.47 2.1621 0.026*

Predictor 11 65.71 231 258.77 7.2612 <0.001***

Density 1 1.84 230 256.93 2.2328 0.138

Species · predictor 107 148.78 123 108.15 1.6903 0.004**

Predictor · density 11 13.79 112 94.36 1.5237 0.134

Species · density 10 10.45 102 83.91 1.2705 0.257

Error 102

(c)

Null 299 439.43

Species 9 18.91 290 420.52 2.5529 0.009**

Predictor 12 71.06 278 349.46 7.1949 <0.001***

Geographical area 2 7.06 276 342.40 4.2912 0.015*

Species · predictor 95 136.87 181 205.53 1.7507 0.001**

Predictor · geographical area 22 74.06 159 131.46 4.0906 <0.001***

Species · geographical area 16 13.78 143 117.69 1.0462 0.412637

Error 143

Terms were fitted sequentially starting from the top. *P < 0.05, **P < 0.01, ***P < 0.001.



–1.0 –0.5 0.0 0.5 1.0 1.5 2.0

–1.0

–0.5

0.0

0.5

1.0

1.5

2.0

Parameter in grass

Par

amet

er in

mix

ed

2 4 6 8 10

2

4

6

8

10

Parameter rank in grass

Par

amet

er r

ank

in m

ixed

–0.5 0.0 0.5 1.0

–0.5

0.0

0.5

1.0

1.5

2.0

2.5

Parameter in high

Par

amet

er in

low

2 4 6 8 10

2

4

6

8

10

Parameter rank in high

Par

amet

er r

ank

in lo

w

(a) (b)

(c) (d)

(e) (f)12

8

Parameter rank in SEParameter in SE

Par

amet

er in

N /

SW

Par

amet

er r

ank

in N

/ S

W

4 5 6 7 8 9

2

4

6

8

0.0 0.5 1.0 1.5

0

1

2

3

10

2

11

8

53

9 6

12

7

1

4

17

10

294

126

3

5

11

8

11

8

53

9

2

7

1

4

6

12

10

7

1 12

4

118

5

103

2

9

6

1

4 1

7

7

12

11

11

8

8

6

6

55

9

9

2

2

33

1

1

7

7

12

12

9

9

4

4

3

3

8

22

6

6

5

5

11

114

Figure 3 Summary graphs of parameter estimates derived for an �average� species (i.e. mean estimate across all species) from three different

groupings of habitat-association models. Models derived from sites with: (a) and (b) predominantly grassland vs. sites with a mix of grass and

arable (c) and (d) low species density vs. high species density; (e) and (f) models derived in three different geographical regions (south-east,

south-west and north). The left-hand side graphs (i.e. a, c and e) illustrate models derived using the actual parameter values and the right-hand

side graphs (i.e. b, d and f) illustrate models derived using the rank parameter estimates. Note that the numbers correspond to the following

codes: 1, boundary height; 2, boundary strip; 3, boundary width; 4, brassica fields; 5, ditch; 6, height of vegetation in summer; 7, presence of

hedge; 8, tilled fields; 9, tree presence; 10, winter set-aside fields; 11, winter stubble fields; 12, woodland edge (for further details of predictors

see Appendix S1). The dashed line represents a perfect 1 : 1 relationship.



species� responses are to changing habitat variables between

areas.

In the Introduction we highlighted three components of

model fit that may be useful in characterizing how useful a

model may be when applied elsewhere. The total model fit

(e.g. R2) is an indicator of net model performance. This

would almost certainly be expected to indicate that a model

performs less well when applied elsewhere, as this quantity

is optimized in model fitting for a given data set. Although

measures of total model fit are a useful guide, they may be

deficient in two respects. First, if one predictor is

particularly influential, then this may yield models with

good fit, irrespective of the effects of other predictors.

Second, a model may perform relatively poorly in terms of

total model fit when applied elsewhere, but may be

structurally correct in terms of the direction and relative

effects of predictors. In the former case it may be that

model predictions are applied with undue faith, in the latter

they could be rejected unnecessarily.

In our analyses we have attempted to dissociate these

aspects of model fit, and have performed a detailed analysis

of model structure by analysing two further components of

model fit based on parameter estimates. When data were

stratified according to density and farming type the results

indicated that both the absolute and rank effects of

parameters were well correlated in different habitat types.

Consequently, we can conclude with reasonable confidence

that these models are essentially transferable between

habitats. When the data were stratified according to region,

however, not only were measures of total model fit poor, we

also found that both the numerical values and rank effects

of predictors were poorly correlated. In this case we can be

confident that models are not generally transferable.

How can these results be interpreted in practical terms or,

in other words, how poor do models have to be in order for

the practical implementation to be compromised? The

answer to this question is probably not general, but depends

on how models are used. The type of application we

envisage for the models we report would include deciding

the extent to which changes to given habitat features impact

on bird populations, and ranking the effectiveness of

different habitat variables ultimately with the aim of

informing AES. It is clear that for both purposes total

model fit is not an appropriate measure as these applications

focus on individual habitat variables and their effects. The

two further measures that we employ indicate quite clearly

how robust predictions might be: for instance, significant

correlations between parameter estimates show that the

quantitative effects (i.e. increase in bird density per unit

change in predictor) are similar between habitats; correlated

rank effects show that the relative importance of different

predictors does not change markedly. In summary, the

measures of model performance that we have used in the

analysis allow us to translate our model analysis into

practical recommendations and we have been able to

indicate how robust these predictions are to landscape-scale

habitat features.

In some senses the issue of generality of habitat selection

may seem trivial: Wiens (1989) noted that the general habitat

associations of many bird species are already known to any

good birdwatcher; so why do ecologists continue to study

habitat associations for such species? The answer lies in the

details. Using Johnson’s (1980) hierarchical levels of habitat

selection most birdwatchers are likely to know details of the

geographical range (e.g. yellowhammers occur in farm-

land) and home ranges within the geographical range

(e.g. yellowhammers nest along field boundaries) of a given

species but at the finer scales of patches within home ranges

and microhabitats within patches (e.g. Bradbury et al. 2000;

Whittingham et al. 2005) even a very knowledgeable

birdwatcher may struggle to predict optimal habitat patterns.

It is at precisely these finer scales that our study is based. We

are not arguing that coarser scale selection is not important

but within our case study of farmland birds, the manipu-

lation of farmland habitats to increase abundance of a range

of fairly abundant and widespread species is the target of

AESs. Thus the issue of generality of habitat selection across

a range of agricultural landscapes within which the

organisms are known to occur is key to their success.

Recent studies have raised concern that AESs may be

ineffective or have benefits that are small or affect only a

few species (Swash et al. 2000; Kleijn et al. 2001; Kleijn &

Sutherland 2003; Bradbury et al. 2004). The recommenda-

tions used to design such schemes are applied over national

and international areas (Kleijn et al. 2004; Vickery et al.

2004) but are usually derived from small-scale studies, often

conducted at a limited number of sites [e.g. birds (Bradbury

et al. 2000; Perkins et al. 2000); mammals (Smith et al. 2004);

invertebrates (Dauber et al. 2005)]. This disparity between

the scales at which management recommendations are

derived and implemented is one factor that may constrain

the effectiveness of schemes to benefit wildlife. Our results

suggest that this is likely to be the case because we

demonstrate that parameter estimates of the effects of

management options, such as grass margins and hedges,

from models constructed in different geographical regions

were significantly different from one another (Table 1c and

Appendix S2c). Therefore, options that deliver benefit for

farmland birds in one region may not necessarily do so

elsewhere. This suggests that schemes targeted at smaller

spatial scales, e.g. regional, may be more effective than the

larger spatial scales on which schemes are currently applied.

Why are species models different in different geograph-

ical areas? We can think of at least four possible

explanations of these regional differences. First, if a species

is limited by different factors in different parts of its range



(e.g. nest sites in one area and chick food supplies in

another), then habitat associations are likely to differ

accordingly [e.g. it is likely that suitable breeding habitat

limits curlew Numenius arquata L. populations on Orkney,

which supports the highest recorded UK breeding popula-

tion densities, where moorland is far preferred to nearby

improved grass fields (Gibbons et al. 1993); in contrast, in

Northern Ireland curlews make more use of improved grass

fields where populations appear to be limited by nest and

chick predation (Grant et al. 1999)]. Second, although we

found that farming type did not significantly affect habitat

association models our regional strata are likely to represent

some differences in farming (e.g. there is more pastoral

farming in western England) as well as differences in

topography, underlying soils and climate which may all

combine to affect habitat choice. Third, there may be

context-dependent habitat selection. A species may show a

stronger selection for habitat X at site A (where habitat X is

rare and/or there are no habitats similar to X) than at site B

(where habitat X is more widely available and has other

habitats similar to X present). For example, one study found

arable crops were more strongly favoured by granivorous

birds in grassland-dominated landscapes (in other words

arable crops were more strongly selected when they were

less available) (Robinson et al. 2001). Finally, species may

have evolved different habitat selection across different

regions. However, this seems unlikely given the genetic

similarity of even sedentary farmland species across the UK

such as the yellowhammer (Lee et al. 2001) and the

migration behaviour of some of the species we studied

(e.g. linnets, blackbird and song thrush populations often

move considerable distance both within and outside the

UK; Wernham et al. 2002).

One previous study has attempted to determine whether

AESs differ in their effects on birds in different landscapes

across 40 different fields (and their surrounding area) in the

Netherlands (Kleijn et al. 2004). They found no differences

in the effects of the schemes in the three different landscape

areas (determined by soil type). Our work differs from this

earlier study because we concentrate on the broader

question of spatial variation in habitat association, which

underlies not just AESs but conservation schemes in

general.

Our results indicate that management options targeted at

a regional, rather than a national scale, may be more

effective. For example, local targeted conservation effort,

via the Countryside Stewardship Scheme, for cirl buntings in

SW England has increased their populations by 83% on

Countryside Stewardship land but by only 2% on land

adjacent to, but outside, the scheme (Peach et al. 2001).

However, AES options (and conservation programmes in

general) targeted at regional scales are likely to be much

more difficult to implement when a species� geographical

range encompasses a variety of landscapes and agricultural

systems. Effective management is likely to depend on a

detailed knowledge of the variation in response shown by a

species across its range and a translation of this knowledge

into locally or regionally targeted management options. A

good example is provided by the northern lapwing Vanellus

vanellus L. which breeds widely across arable and grassland

agricultural systems in both lowland and upland contexts in

Europe. In all these habitats, lapwings have faced declining

habitat quantity and quality for several decades, and severe

population declines have resulted (e.g. Wilson et al. 2001,

2005; Taylor & Grant 2004). Despite local successes

through intensive management mainly on nature reserves

(e.g. Ausden & Hirons 2002), declines have not yet been

reversed by AES implementation. However, there is rich

information on the behavioural and demographic responses

of lapwings to agricultural management across a wide range

of systems (e.g. Galbraith 1988; Baines 1990; Johannson &

Blomqvist 1996; Sheldon 2002), so knowledge does exist to

develop management for this species that is better targeted

towards regional variation in its associations with agricul-

tural practice but to date management options have been

based at the national scale.

Clearly within our analysis we have shown that predictors

affect species in different ways (species · predictor interac-

tion was significant in Table 1a–c; see also Supplementary

material for details). For example, whilst most species prefer

taller hedges the reverse is true of reed buntings. In this

study, we have looked at average responses across species.

We did this because AES are applied not to the benefit of

any single species, but across all species. We intend to

analyse species-specific responses in detail elsewhere;

however, here we discuss one important point. Within the

group of species we have analysed, there are species that

specialize to varying degrees on farmland. For instance,

linnets and yellowhammers are closely associated with

farmland and could be considered specialists, whereas

arguably other species, such as song thrushes and greenfin-

ches, are less specialized. Our analysis has averaged across

all species, and there is an important question of how such

averaging may influence our results. We repeated the

analyses reported above for only four species that might

be regarded as better adapted to farmland (linnet, yellow-

hammer, chaffinch and reed bunting); in addition, three of

these species have experienced significant population

declines (http://www.bto.org/psob/redlist.htm2 ) and are

thus a focus for conservation effort on farmland. We

obtained essentially the same results for these species

(M.J. Whittingham and R.P. Freckleton, unpublished results3 );

e.g. a significant effect for the �predictor · region� interac-

tion term, similar to that shown in Table 1c. However, as

mentioned above, species-specific responses undoubtedly

exist. There is an important issue of the degree to which



individual species are benefited by broad-brush management

such as AES. Analysis of the species-specific variation about

the average response should reveal such effects.

C O N C L U S I O N S

Our study highlights that ecologists should beware of

placing too much confidence in the use of habitat-

association modelling to make predictions. Even within a

relatively homogenous environment, such as that created by

lowland farming practices in our study areas, we have found

that patterns of habitat association vary on a regional basis.

Our study suggests that it should not be assumed that

management recommendations that are based on habitat

associations derived from studies in a small subset of a species�range will necessarily solve that species� conservation prob-

lems over its entire range. Although AESs targeted at species

occupying small ranges have met with some success, AESs

aimed at species� with a geographical range encompassing a

variety of landscapes have been less successful. For some

species (e.g. Lapwing, see above) sufficient knowledge may

already exist for regional scale programmes, but for the

majority of less well-studied farmland species, the most cost-

effective approach may be to use AES themselves as the basis

for trialling management options across multiple locations,

and adapting recommended management as necessary in the

light of population response. Although our work focuses on

farmland systems, the results have wider implications for the

scale on which conservation programmes for widespread

species should be based.

A C K N O W L E D G E M E N T S

Many thanks to the CBC volunteers who collected the bird

data used in this study. Thanks also to Richard Thewlis, Rick

Goater and Dan Chamberlain. The work in this project was

funded by BBSRC (ref no: 43/D13408) and by a David

Phillips Fellowship to MJW. RPF is a Royal Society

University Research Fellow.

R E F E R E N C E S

Ausden, M. & Hirons, G.J.M. (2002). Grassland nature reserves for

breeding wading birds in England and the implications for the

ESA agri-environment scheme. Biol. Conserv., 106, 279–291.

Baines, D. (1990). The roles, of predation, food and agricultural

practice in determining the breeding success of the lapwing

(Vanellus vanellus) on upland grasslands. J. Anim. Ecol., 59,

915–929.

Bishop, C.B. & Myers, W.L. (2005). Associations between avian

functional guild response and regional landscape properties for

conservation planning. Ecol. Indic., 5, 33–48.

Bradbury, R.B., Kyrkos, A., Morris, A.J., Clark, S.C., Perkins, A.J.

& Wilson, J.D. (2000). Habitat associations and breeding success

of yellowhammers on lowland farmland. J. Appl. Ecol., 37,

789–805.

Bradbury, R.B., Browne, S.J., Stevens, D.K. & Aebischer, N.J.

(2004). Five-year evaluation of the impact of the arable stew-

ardship pilot scheme on birds. Ibis, 146, 171–180.

Burnham, K.P. & Anderson, D.R. (2002). Model Selection and

Multimodel Inference. Springer, New York.

Dauber, J., Purtauf, T., Allspach, A., Frisch, J., Voigtlander, K. &

Wolters, V. (2005). Local vs. landscape controls on diversity: a

test using surface-dwelling soil macroinvertebrates of differing

mobility. Glob. Ecol. Biogeogr., 14, 213–221.

Freckleton, R.P., Gill, J.A., Noble, D. & Watkinson, A.R. (2005).

Large-scale population dynamics, abundance-occupancy

relationships and the scaling from local to regional population

size. J. Anim. Ecol., 74, 353–364.

Freckleton, R.P., Noble, D. & Webb, T.J. (2006). Distributions of

habitat suitability and the abundance–occupancy relationship.

Am. Nat., 167, 260–275.

Galbraith, H. (1988). Effects of agriculture on the breeding ecology

of Lapwings Vanellus vanellus. J. Appl. Ecol., 25, 487–503.

Gibbons, D.W., Reid, J.B. & Chapman, R.A. (1993). The New Atlas

of Breeding Birds in Britain and Ireland: 1988–1991. T & A.D.

Poyser, London, UK.

Gill, J.A., Norris, K., Potts, P.M., Gunnarson, T.G., Atkinson,

P.W. & Sutherland, W.J. (2001). The buffer effect and large-scale

population regulation in migratory birds. Nature, 412, 436–438.

Grant, M.C., Orsman, C., Easton, J., Lodge, C., Smith, M.,

Thompson, G. et al. (1999). Breeding success and causes of

breeding failure of curlew Numenius arquata in Northern Ireland.

J. Appl. Ecol., 36, 59–74.

Johannson, O.C. & Blomqvist, D. (1996). Habitat selection and

diet of Lapwing Vanellus vanellus chicks on coastal farmland in

S.W. Sweden. J. Appl. Ecol., 33, 1030–1040.

Johnson, D.H. (1980). The comparison of usage and availability

measurements for evaluating resource preference. Ecology, 6,

65–71.

Kleijn, D. & Sutherland, W.J. (2003). How effective are European

agri-environment schemes on conserving and promoting bio-

diversity? J. Appl. Ecol., 40, 947–970.

Kleijn, D., Berendse, F., Smit, R. & Gilissen, N. (2001). Agri-

environment schemes do not effectively protect biodiversity in

Dutch agricultural landscapes. Nature, 413, 723–725.

Kleijn, D., Berendse, F., Smit, R., Gilissen, N., Smit, J., Brak, B.

et al. (2004). Ecological effectiveness of agri-environment

schemes in different agricultural landscapes in the Netherlands.

Conserv. Biol., 18, 775–786.

Krebs, J.R., Wilson, J.D., Bradbury, R.B. & Siriwardena, G.M.

(1999). The second silent spring? Nature, 401, 611–612.

Lee, P.L.M., Bradbury, R.B., Wilson, J.D., Flanagan, N.S., Rich-

ardson, L., Perkins, A.P. et al. (2001). Microsatellite variation in

the yellowhammer Emberiza citrinella: population structure of a

declining farmland bird. Mol. Ecol., 10, 1633–1644.

Manel, S., Williams, H.C. & Ormerod, S.J. (2001). Evaluating

presence-absence models in ecology: the need to account for

prevalence. J. Appl. Ecol., 38, 921–931.

Marchant, J.H., Hudson, R., Carter, S.P. & Whittington, P. (1990).

Population Trends in British Breeding Birds. British Trust for Orni-

thology, Thetford, UK.

Meyer, C.B. & Thuiller, W. (2006). Accuracy of resource selection

functions across spatial scales. Divers. Distrib., 12, 288–297.



Peach, W.J., Lovett, L.J., Wotton, S.R. & Jeffs, C. (2001). Coun-

tryside stewardship delivers cirl buntings (Emberiza cirlus) in

Devon, UK. Biol. Conserv., 101, 361–373.

Peach, W.J., Denny, M., Cotton, P.A., Hill, I.F., Gruar, D., Barritt,

D. et al. (2004). Habitat selection by song thrushes in stable and

declining farmland populations. J. Appl. Ecol., 41, 275–294.

Perkins, A.J., Whittingham, M.J., Bradbury, R.B., Wilson, J.D.,

Morris, A.J. & Barnett, P.R. (2000). Habitat characteristics

affecting use of lowland agricultural grassland by birds in winter.

Biol. Conserv., 95, 279–294.

Robinson, R.A., Wilson, J.D. & Crick, H.Q.P. (2001). The

importance of arable habitat for farmland birds in grassland

landscapes. J. Appl. Ecol., 38, 1059–1069.

Rushton, S.P., Ormerod, S.J. & Kerby, G. (2004). New paradigms

for modelling species distributions. J. Appl. Ecol., 41, 193–201.

Sheldon, R. (2002). Lapwings in Britain – a new approach to their

conservation. Br. Wildlife, 13, 109–116.

Smith, R.K., Jennings, N.V., Robinson, A. & Harris, S. (2004).

Conservation of European hares Lepus europaeus in Britain: is

increasing habitat heterogeneity in farmland the answer? J. Appl.

Ecol., 41, 1092–1102.

Stephens, P.A., Buskirk, S.W., Hayward, G.D. & Del Rio, C.M.

(2005). Information theory and hypothesis testing: a call for

pluralism. J. Appl. Ecol., 42, 4–12.

Sutherland, W.J. (1998). Conservation Science and Action. Blackwell

Science, Oxford, UK.

Swash, A.R.H., Grice, P.V. & Smallshire, D. (2000). The con-

tribution of the UK Biodiversity Action Plan and agri-environ-

ment schemes to the conservation of farmland birds in England.

In: Ecology and Conservation of Lowland Farmland Birds (eds Aeb-

sicher, N.J., Evans, A.D., Grice, P.V. & Vickery, J.A.). British

Ornithologists� Union, Thetford, UK, pp. 36–42.

Taylor, I.R. & Grant, M.C. (2004). Long-term trends in the

abundance of breeding Lapwing Vanellus vanellus in relation to

land-use change on upland farmland in southern Scotland. Bird

Study, 51, 133–142.

Vickery, J.A., Bradbury, R.B., Henderson, I.G., Eaton, M.A. &

Grice, P.V. (2004). The role of agri-environment schemes and

farm management practices in reversing the decline of farmland

birds in England. Biol. Conserv., 119, 19–39.

Wernham, C., Toms, M., Marchant, J., Clark, J., Siriwardena, G. &

Baillie, S. (2002). The Migration Atlas: Movements of the Birds of Britain

and Ireland. T & A.D. Poyser, London, UK.

Whittingham, M.J., Swetnam, R.D., Wilson, J.D., Chamberlain,

D.E. & Freckleton, R.P. (2005). Habitat selection by yellow-

hammers Emberiza citrinella on lowland farmland at two spatial

scales: implications for conservation management. J. Appl. Ecol.,

42, 270–280.

Wiens, J.A. (1989). The Ecology of Bird Communities. Cambridge

University Press, Cambridge, UK.

Wilson, A.M., Vickery, J.A. & Browne, S.J. (2001). Numbers and

distribution of Northern Lapwings Vanellus vanellus breeding in

England and Wales in 1998. Bird Study, 48, 2–17.

Wilson, A.M., Vickery, J.A., Brown, A., Langston, R.H.W., Small-

shire, D., Wotton, S. et al. (2005). Changes in the numbers of

breeding waders on lowland wet grasslands in England and

Wales between 1982 and 2002. Bird Study, 52, 55–69.

S U P P L E M E N T A R Y M A T E R I A L

The following supplementary material is available for this

article:

Appendix S1 List of habitat parameters used as potential

explanatory predictors in bird–habitat models.

Appendix S2 Summary of the parameter estimates derived

from habitat-association models for 11 species.

This material is available as part of the online article from:

http://www.blackwell-synergy.com/doi/full/10.1111/

j.1461-0248.2006.00992.x

Please note: Blackwell Publishing are not responsible for the

content or functionality of any supplementary materials

supplied by the authors. Any queries (other than missing

material) should be directed to the corresponding author for

the article.

Editor, Mark Schwartz

Manuscript received 7 August 2006

First decision made 14 September 2006

Manuscript accepted 12 October 2006



Should conservation strategies consider spatial generality? Farmland birds show regional not national patterns of habitat association

Documents