Trade-offs between specificity and regional generality in habitat association models: a case study of two species of African vulture


*Correspondence author. E-mail:


1.  The habitat requirements of various species have been evaluated by statistical models. However, recent studies have shown that models are often not transferable between regions, limiting their applicability and ability to inform management decisions. One possible cause is that models tend to reflect dominant landscape features, which vary between regions. Transferability, and thus applicability, may be increased by developing models from multiple regions.

2.  We addressed this via a case study of two vulture species (white-backed and lappet-faced vultures, Gyps africanus and Aegypius tracheliotos) from six biogeographically different regions across southern Africa. Logistic models, developed using an information-theoretic approach, were used to predict nest occurrence based on explanatory variables derived from a Geographic Information System (GIS), the usual method for species with large ranges. Variables reflected key requirements at different spatial scales: food availability, human disturbance and nesting trees. We developed models using data from single and multiple regions, and tested the cross-regional transferability. We also collected field data to asses the adequacy of the GIS variables.

3.  There was a significant negative correlation between specificity and regional generality, multi-region models tending to be more consistently transferable than single-region models but having a weaker fit within the regions where they were developed. Multi-region models of nesting habitat were more structurally similar to each other than single-region models. GIS variables adequately represented the landscape but with differing adequacy between regions. There were no observed fitness benefits to the observed site selection.

4.Synthesis and applications. Models of species distribution are not transferable between regions, and use of models to inform management decisions in regions other than that used for model development should be undertaken with caution. Models are often built using GIS predictors only broadly related to the landscape properties of interest and the adequacy of such proxies can vary between regions, leading to models that emphasize dominant landscape features. Models developed from multiple regions partially overcome this problem by identifying predictors that apply across many regions and are more transferable. However, this increased generality trades off against reduced specificity. Models should be constructed with consideration to their intended use.


Habitat association models relate the occurrence of organisms to the presence and quantifications of key attributes of the habitat (Cowley et al. 2002). Such models are widely used in conservation programmes, to identify remaining areas of suitable habitat in regions with high levels of human disturbance (Early, Anderson & Thomas 2008) and to predict species’ response to changes in the habitat (Berry et al. 2008). An emerging issue with this type of modelling is regional variation in habitat relationships over large geographic ranges. Whittingham et al. (2007) found that predictors of lowland farmland birds in England differed in effect between regions. McAlpine et al. (2008) likewise found regional differences in habitat association of koalas in Australia. One reason for this may be that models tend to reflect the dominant features of the ‘training’ region used in their development (Whittingham et al. 2007; Zharikov, Lank & Cooke 2007) and these vary between regions. Zharikov et al. (2007) found that models with less discriminatory ability within their training regions were more transferable to other regions, and Whittingham et al. (2007) found that transferability increased when data were segregated based on habitat type rather than region, and thus developed models based on data from several regions.

The main application of habitat association models is in conservation programmes, and most have been for high-profile species, or keystone and umbrella species such as top predators. Raptors, our study system here, fall into both categories and the habitat associations of many large raptors have been evaluated. Raptors often occur at low densities over large ranges, and individuals range over large areas and many habitat types (Newton 1979), potentially making them particularly susceptible to a lack of regional generality. Sightings of organisms in non-ideal habitat can cause discriminatory problems in models (Titeux et al. 2007) and most habitat association models for raptors have been either limited to a large scale (Osborne & Tigar 1992; Bustamante & Seoane 2004; Balbontin 2005) or focused on a key requirement: nesting habitat (Martinez, Serrano & Zuberogoitia 2003; López-López et al. 2006, 2007; Morán-López et al. 2006a; Bamford, Monadjem & Hardy 2009).

As an additional complication, constructing habitat association models requires a priori knowledge of a species to allow suitable habitat characteristics to be chosen for inclusion (Rushton, Ormerod & Kerby 2004). Many rare species are not well studied and selection of habitat characteristics is problematic, with general characteristics often being picked. Habitat models for raptors often use the same variables (topography, land cover, land use and human disturbance) and, although they provide a good description of species occurrence in the study area, there has generally been little effort to confirm cause and effect in the relationships indicated. Many variables, especially those relating to topography, are chosen as proxy measures but the relationship between the variable and the property of underlying interest is usually assumed and not tested in the field.

Another method for validating models is comparing habitat association with fitness (usually reproductive success) but preference is often either unrelated or negatively related to fitness (Chalfoun & Martin 2007). This may be because fitness is similar across habitat types due to ideal free habitat selection but also because habitat selection is a complex process representing trade-offs between the factors considered (Chalfoun & Martin 2007). For nesting birds these considerations may be the risk of nest predation and food availability. Among raptors, nest site selection is negatively affected by human disturbance (Donázar et al. 2002; Morán-López et al. 2006b; Bamford et al. 2009), which also reduces nest success, as do climatic effects such as rainfall (Donázar et al. 2002; Morán-López et al. 2006b).

Our objectives are to: (i) construct models of breeding habitat preference; (ii) compare the transferability of single-region and multi-region models; (iii) test, using field data, relationships assumed in the choice of predictor variables; and (iv) test predicted preferences against fitness data. We address these objectives by focusing on African white-backed vultures (WbV), Gyps africanus (Salvadori), and lappet-faced vultures (LfV), Aegypius tracheliotos (Forster), family Accipitridae.

Study system

Vulture populations have generally declined as human population densities have increased. Although African vultures are less immediately threatened than European and Asian vultures, populations are declining and are becoming restricted to wildlife management areas (WMAs) throughout the continent (Monadjem & Garcelon 2005; Thiollay 2006), even in areas with low human population densities (Herremans & Herremans-Tonnoeyr 2000). Suitable habitat occurs outside of WMAs even when human population densities are high, suggesting that human persecution (Bamford et al. 2009) or a lack of available food limits vultures to WMAs. We focus on the two most widespread species: African WbV (G. africanus) and LfV (A. tracheliotos). Both are tree-nesting and found throughout sub-Saharan Africa, excluding central-western tropical forests, and co-exist across much of their ranges. WbV are relatively well studied and their habitat associations well documented at some sites. They are gregarious at feeding and roosting sites, nesting in loose aggregations, sometimes referred to as colonies (Bamford et al. 2009), which can reach high densities (Monadjem & Garcelon 2005). LfV are less studied; they are territorial and occur at low densities with habitat associations known only in very general terms, making the choice of explanatory variables more difficult.

Materials and Methods

Conceptual model

We broadly considered three types of factors influencing nesting distributions: presence of suitable trees, food availability and human disturbance.

Nesting trees

WbV nest in savannas and woodland in a variety of tree species; LfV nest in savanna to desert regions, mainly in Acacia species. Both build nests in the tops of large trees. Bamford et al. (2009) found that topography and soil types predicted WbV nest occurrence and, within their study site, these variables predict the availability of large trees. We follow the approach of using correlates of vegetation structure measured at a fine resolution using a Geographic Information System (GIS).

Food availability

Nesting Gyps vultures are food limited (Houston 1976), so food availability in the area around the nest may influence site selection (Murn & Anderson 2008). Both LfV and WbV are regularly sighted on large mammal carcases. As surveying large areas for carcass availability is impractical, we used as an estimate the Normalised Difference Vegetation Index (NDVI, the ratio of blue to near infra-red light), which is positively associated with herbivore abundance (van Bommel et al. 2006). Carcass availability is affected by land use and management (Murn & Anderson 2008). Sixty per cent of movements from roost site to feeding site by cape vultures Gyps coprotheres were within 20 km (Bamford et al. 2007), and we consider this distance as the foraging range of a nesting vulture. Vultures locate food through an information network (Jackson, Ruxton & Houston 2008), which requires other vultures to be in the area, and the presence of other nesting WbV is a major determinant of WbV nesting preferences (Bamford et al. 2009).

Human disturbance

Direct persecution of nesting vultures, and indirect disturbance such as land management (see above), may affect nesting distributions of vultures (Monadjem & Garcelon 2005; Bamford et al. 2009).

Figure 1 illustrates our multiscale conceptual model of how the above factors might operate. While differences between the two species are likely, lack of knowledge of LfV habitat associations lead to the use of the same variables for both.

Figure 1.

 Multi-scale conceptual model of nest site selection by white-backed and lappet-faced vultures.

Study sites

Nest site data in this study are taken from six regions across southern Africa (Fig. 2), although both species are found together in only three of these. These regions differ in terms of biogeography, fragmentation and human disturbance: further details are given in Appendix S1.

Figure 2.

 Locations of the study regions and surveyed areas within the regions: (a) Linyanti; (b) Makgadikgadi; (c) Kimberley; (d) Swaziland; (e) Namib; and (f) Zululand (Northern KwaZulu-Natal).

Presence/absence data

Swaziland, Kimberley and the Namib have been surveyed from 2002–2007, Zululand from 2004–2007, Linyanti 2006–2007 and Makgadikgadi 2006–2007 for LfV and for WbV in 2007. In all regions except Linyanti and Makgadikgadi, the surveyed region was not continuous, but divided into several subregions. All regions were surveyed from the air at least once and additionally from the ground. Swaziland and Zululand were considered as one region for LfV due to the small numbers of nests found.

Human persecution and land use and management affect vulture nesting distributions (Herremans & Herremans-Tonnoeyr 2000; Monadjem & Garcelon 2005; Thiollay 2006) and habitat suitable for nesting may not be utilized if it is unprotected (Bamford et al. 2009), or if insufficient food is available. The inclusion of unprotected areas in our analysis may lead to many ‘false negatives’ in the data and, as we wished to examine habitat preferences without the complicating factor of human persecution (which can be difficult to evaluate due to very small numbers of nests recorded outside WMAs), analysis was restricted to WMAs (nature reserves, controlled hunting areas and private game farms). Nest presence data were edited so that no two nest-containing points were closer than 200 m. Absences were randomly generated so that the number of absences for each species in each region was equal to the number of presences. For each multi-region model, 200 points (presences and absences) were randomly selected from each of three regions for the WbV models (600 points overall), and 100 points from each of two regions for the LfV models (200 points overall). This ensured that no one region dominated the multi-region data sets. We generated three multi-region data sets for LfV and five for WbV, giving equal numbers of single- and multi-region data sets.

Explanatory variables

Ten variables, selected following the models of Bamford et al. (2009) and our conceptual model (Fig. 1), were mapped and measured using the GIS software arcview 3.2 (Environmental Systems Research Institute, Redlands, CA, USA). Further details on the variables are given in Appendix S1.

Vegetation predictors

As correlates of vegetation structure, we measured distance to the nearest river, slope, standard deviation of elevations surrounding nest sites (testing this measure over various areas) and soil type (broadly classified and assigned ordinal fertility rankings: no or infertile soil, shallow/poorly developed, predominantly sandy, predominantly clay). Vegetation type, derived from Landsat images, was considered as five categories: clear ground, savanna (sparse vegetation), woodland (dense vegetation, predominantly trees), thicket (dense vegetation, predominantly shrubby) and lush (thick vegetation close to water). The Landsat model correctly predicted vegetation in 83% of test cases, significantly better than random expectation (χ2 = 57·4, d.f. = 6, < 0·001).

Human disturbance

We included the one variable relating to indirect human disturbance that Bamford et al. (2009) found significant within WMAs: the area of undisturbed (very little or no tree removal) habitat surrounding the nest.

Food availability

Explanatory variables estimating food availability, measured across a 20-km foraging range from nest sites, were: the long-term mean NDVI (from 8 km resolution Advanced Very High Resolution Radiometer images) and the proportion of land with food available to vultures (equating to WMAs), hereafter %WMA.


We incorporated Augustin, Mugglestone & Buckland’s (1996) autologistic technique, which is useful in cases where the cause of autocorrelation is interspecific attraction. This introduces an extra variable: the density of nests (weighted by distance) surrounding each target nest site to an arbitrary distance. We generated two values, over 500 m and 5 km distances.

Statistical analysis

We developed habitat association models using an information-theoretic approach (Burnham & Anderson 2002). Logistic models, in which a binomial distribution of residuals is assumed (Crawley 2002), were used to test explanatory variables. We used GenStat v11 (VSN International, Hemel Hempstead, UK) with a modified version of the RSEARCH procedure. Models were compared using Akaike’s information criterion (AIC), calculated as AIC = G + 2N, where G is the deviance explained by the model (equivalent to −2 × maximum log-likelihood of the model) and N is the number of parameters in the model.

Spearman’s rank test was used to check for collinearity between explanatory variables. If a pair of variables had a correlation coefficient >0·5, the variable with the lower AIC was removed in order to avoid interpretational difficulties. For each single or multi-region data set, we constructed sets of models from all linear combinations of the explanatory variables. Within each set, we calculated the Akaike weight (wm) for each model, which is a measure of the probability that it is the Kullback–Leibler best model given a set of models (Burnham & Anderson 2002). The direction and magnitude of the effect of each explanatory variable was determined by averaging the parameter estimates, weighted by wm, from each model in the set. Parameter uncertainty was quantified by calculating the unconditional standard error of the estimates (Burnham & Anderson 2002). We summed the Akaike weights of each model in which the variable occurred (∑ wm) and used this to rank the variables, larger values indicating the variable was more important relative to the others. To quantify model uncertainty, we constructed 95% confidence sets of models by summing the weights of the models, starting with the highest and working down, until the sum exceeded 0·95. Making predictions from logistic models is complicated by the link function: the predicted value is not a linear function of the parameters, so the predicted value is calculated as the weighted average of the predicted values from the 95% confidence set of models (Whittingham et al. 2007).

Model analysis

A frequently used method for measuring model discrimination ability is to measure the area under receiver operating characteristics curves (AUCs). AUC values range from 1·0 (perfect fit of the model to the test data set) to 0·0 (perfect inverse fit of the model), with 0·5 indicating model performance expected by chance. We calculated AUC values for each model within the region(s) it was developed in and for cross-predictions to the other regions. Models were fitted using the mean value for autocorrelation for that region for all data points.

We tested models for quantitative similarity by comparing the parameter estimates generated by single and multi-region models: we calculated the differences between the model-averaged estimates for each parameter within each model type (five single-region data sets gives five model-averaged estimates for each parameter, and 10 differences between all combinations). Two-way anova was used to test the groups of differences. The qualitative similarity of the models was tested by anova comparing differences in the ranks of the explanatory variables.

Field validation

Vegetation predictors

To test the adequacy of GIS vegetation predictors, we collected data on nesting trees and trees without nests (105 nesting trees for WbV, 35 for LfV, and an equal number of non-nesting trees), randomly selected from across Swaziland, Zululand, Kimberley and the Namib. Data collected were: tree species, height, height relative to surrounding trees, surrounding tree and shrub densities, and structural complexity of tree (ranked 1–5). Tree species were classed as broad-leafed non-thorny trees, thorn trees or fig trees. Logistic models tested the effect of these variables on nest site preference, with each variable assigned a P-value based on the change in deviance, G, it caused on removal from the model (Crawley 2002). Using the resultant model, we visited 140 random sites from the study regions to determine if they contained suitable nesting trees (0 = no tree, 1 = suitable tree present). These binary results were correlated with the probabilities predicted by a logistic model developed using only GIS vegetation predictors.

Vulture fitness

Habitat association is assumed to be adaptive but the consequences of observed selection are rarely quantified (Chalfoun & Martin 2007). One fitness measure is straightforward for nests: whether or not the nest was successful. Parental quality is an important determinant of nest success but with sufficient data habitat effects may also be detectible. Nest successes were recorded in all regions in the same years given for the nest surveys. Fieldwork in the Namib was carried out primarily to ring chicks rather than record nest success; these data are not comparable with the other regions and were analysed separately. Logistic models tested effects of the explanatory variables on nest success. Two methods were used: first, the binary response unit was the individual nest (1 = successful or 0 = failed). Second, population responses were scored as the number of successful nests in the subregion in a given year, with the numerical denominator set as the total number of nests in that region and year (Crawley 2002).


Nest surveys

WbV nests in Swaziland and Linyanti were mainly in riparian woodland in several tree species. Riparian woodland was also favoured in Zululand but with most nests in Acacia and Ficus. Nests in Kimberley and Makgadikgadi were in Acacia savanna. The two Botswana regions had far lower nest densities than the other three regions. LfV nests in Swaziland and Zululand were in Acacia nigrescens in open savanna, nests in the Namib were in Acacia erioloba along watercourses, and nests in Makgadikgadi were mainly in Acacia close to the edge of the salt pan. Nest densities were much higher in the Namib than the other regions. Further survey results are in Table S1.


There was collinearity between slope and S.D.elevation: for both species in all regions, S.D.elevation had greater support and slope was excluded. The strongest support was for measuring S.D.elevation at the largest scale considered, an area of 1 km2 around the nest site. The autocorrelation variables (large and small scale) also show collinearity. Among WbV, there was weak support for the small-scale measure in Swaziland and Zululand, while in the other regions the large-scale measure had strong support. Among LfV, there was weak support for the large-scale measure in all regions. For consistency across regions and species, we excluded the small-scale measure. There remained eight variables for consideration in the models: Autocorrelation, River, S.D.elevation, Soil, VegetationType, Habitat, %WMA and NDVI. These final explanatory variables showed acceptably low levels of collinearity for all regions (rs<0·5).

Model fit and discrimination across regions

For both species, single-region models provided good-to-excellent descriptions of habitat association in the region of their development (Fig. 3). AUC values ranged from 0·93 ± 0·008 for WbV in Swaziland to 0·7 ± 0·033 for WbV in Kimberley. Cross-regional predictions provided weaker descriptions, ranging from good to incorrect with several intermediates where model performance was no better than expected by chance (Fig. 3). In a few cases, model fit was better in regions other than the region of development (e.g. the Kimberley WbV model had higher AUC in Swaziland than in Kimberley), most probably due to the relative variability of the habitat parameters in the regions. Multi-region models provided weaker descriptions of the regions they were developed in than did single-region models (Fig. 3c). There was a trade-off between specificity (the fit of a model to the region(s) it was developed in) and generality (the fit of a model to other regions) (Spearman’s rank test: rs = −0·58, P =0·005; Fig. 4): thus models with high specificity tended to have lower generality.

Figure 3.

 Discriminatory ability, measured as AUC (area under an Roc curve), of: (a) single region white-backed vulture models fitted to each of the five study regions containing this species; (b) single region lappet-faced vulture models; (c) multi-region models for both species, showing the fit to the regions the model was developed in, and to the other regions. *indicates AUC not significantly different from 0·5, where model performance is no better than expected from chance. SWZ, Swaziland; KZN, Zululand; KBY, Kimberley; MKG, Makgadikgadi; LNT, Linyanti; ZSW, Zululand/Swaziland.

Figure 4.

 Negative correlation between specificity (AUC of models within the region(s) used to develop them) and generality (AUC of models applied to other regions). Spearman’s rank test: rs = −0·58, = 0·005.

Parameter estimates and rankings

For both species, each model was dominated by a small subset of the eight variables considered but with little consistency between models as to which variables these were (Table S2). For WbV, river and slope were dominant in Swaziland and Zululand, slope in Linyanti and VegetationType in Kimberley, Makgadikgadi and all the multi-region models. For LfV, soil and river were dominant in the Namib, NDVI in Namib and Swaziland/Zululand (although with opposing effects), and habitat and %WMA were important in all models, single- and multi-region. For both species, the estimate for autocorrelation was apparently affected by nesting density: among WbV there was a positive effect, strongest in the regions with lowest nesting densities; whilst among LfV there was a negative effect, strongest at the highest densities. The variables relating to human disturbance, habitat and %WMA, similarly had the strongest effects in the more disturbed regions for both species. Among WbV S.D.elevation had a strong negative effect in Swaziland but a strong positive effect in Linyanti.

Model uncertainty varied: for WbV, the 95% confidence set contained 10 models in Swaziland (out of 256), 22 in Zululand, 25 in Kimberley, 74 in Makgadikgadi and 28 in Linyanti; and for LfV, 64 models in Swaziland/Zululand, 49 in Makgadikgadi and 7 in the Namib.

The differences between parameter estimates in multi-region models were smaller than in single-region models for WbV (anova: parameter, F11,216 = 10·57, < 0·001; model type, F1,216 = 149·6, < 0·001; interaction, F11,216 = 11·45, < 0·001) and for LfV (anova: parameter, F8,36 = 6·52, < 0·001; model type, F1,36 = 8·22, P = 0·007; interaction, F8,36 = 3·06, P = 0·01), significant interactions indicating that not all parameters estimates showed increased similarity. Parameter rankings were more similar in multi-region than single-region models for WbV (parameter, F7,144 = 5·59, < 0·01; model type, F1,144 = 4·59, P = 0·03; interaction, F7,165 = 3·49, P = 0·002), but not for LfV (parameter, F7,32 = 1·13, P = 0·37; model type, F1,32 = 0·20, P = 0·66; interaction, F7,32 = 1·29, P = 0·29).

Field validation


Two variables were significant in explaining WbV nesting preferences at the tree level: relative tree height (G = 8·8, d.f. = 2, P =0·003) and tree type (G = 4·5, d.f. = 2, P =0·035). Trees equal in height or taller than the surrounding vegetation were preferred, with no difference in response to these two categories (G = 1·9, d.f. = 1, > 0·1). Fig trees were the preferred type, with no difference in the response to thorn trees and broadleaved non-thorny trees (G = 0·0, d.f. = 1, > 0·9). Occurrence of preferred trees in ground surveys correlated with a GIS model containing only vegetation correlates (Swaziland: rs = 0·54, < 0·001; Kimberley: rs = 0·55, < 0·001).

For LfV, the only significant variable was the surrounding tree cover (G = 6·5, d.f. = 1, P =0·03), showing preference for open areas. Tree type was invariant in the sample analysed: all nests were in Acacia and the random trees, taken from savanna areas, were also all Acacia.

Vulture fitness

Individual nest responses for WbV showed that the only correlate of nest success was NDVI (G = 6·0, d.f. = 1, P =0·016, positive effect; Fig. 5a). Considering population success, overall nesting density showed a negative effect (F1,21 = 16·5, P =0·012; Fig. 5b). For LfV in the Namib, NDVI was also significant when individual nests were considered (G = 5·3, d.f. = 1, P =0·02), but had a negative effect (Fig. 5c). Effects of NDVI were not significant in other regions. No other variable was significant. In the population response model, nesting density had no effect for LfV (F1,19 = 0·35, P =0·7). None of the measurements taken from nesting trees was correlated with nest success.

Figure 5.

 Effect of parameters on nest success (chicks per nest) of vultures: (a) NDVI and white-backed vulture nest success; (b) nesting density and white-backed vulture nest success; and (c) NDVI and lappet-faced vulture nest success.


Model transferability

Habitat association models are increasingly used tools in conservation, as powerful desktop GIS and statistical analysis packages are readily available. A key question is whether such models adequately predict distributions but there have been few practical approaches to addressing this. In recent papers Whittingham et al. (2007), constructing models using field data representing key determinants of habitat suitability, and McAlpine et al. (2008), whose models used field data and GIS-derived proxies, found that models were not transferable between regions. Our models, built entirely using proxies (the usual method for species with large ranges), show the same pattern. For both species tested, cross-regional predictions of occurrence were sometimes good but generally weak, and some had no predictive value at all. We highlight this pattern as many studies of habitat association validate models on an independent data set collected from the same region as that used to develop the model (Gavashelishvili & McCrady 2006; López-López et al. 2007) and conclude that the model could be widely applied.

We suggest that this lack of generality could partly be due to inter-regional variation in dominant landscape features and that models developed across several regions might have greater transferability. Multi-region models were more consistently transferable for both species examined but at the cost of reduced specificity. Multi-region models were more structurally similar to each other than single-region models. There was less improvement for LfV, perhaps because of the smaller sample sizes or because the explanatory variables were less suitable for this species. Interpretation of logistic models can be difficult when variables explain small amounts of deviance (Crawley 2002); this can cause problems with inadequate proxy measures, which are not related to habitat preferences (Rushton et al. 2004). Alternatively, the response to some variables may have been nonlinear (e.g. response to NDVI); because logistic regressions assume a linear response, two regions where opposing responses were found would give no apparent response.

Evaluation of habitat relationships

Evaluation of assumed relationships between proxy measures and habitat association is omitted from many studies constructing models using GIS data sources, with the result that the adequacy of proxy measures cannot be assessed. Our results suggested that WbV are adaptable and will simply nest in the tallest trees available, an interpretation supported by records of nests in electricity pylons (Anderson & Hohne 2007), while LfV simply favour open areas of ground. GIS-derived variables accurately identified the locations of suitable trees, suggesting that they were adequate proxies. All models were dominated by a few GIS predictors and the dominant predictors differed between regions, suggesting that the adequacy of proxies varies between regions. In some cases, GIS predictors had opposing effects between regions. For example, variation in elevation was negative in effect on WbV nest site selection in Swaziland but positive in Linyanti. In Swaziland, slopes are associated with mountainous regions and poor soils but, in Linyanti, slopes are river valleys containing rich soils, so the model was reliably incorrect when making predictions in Swaziland. Better soil or vegetation maps would solve this problem, but variables relating to topography are often chosen for inclusion in models precisely because they relate to soils and are in higher resolution than available vegetation maps (Bustamante & Seoane 2004). A land cover map derived from Landsat images was consistent in effect across regions and was the most important habitat variable in all five multi-region models, but was often ranked low in importance in the single region models as one of the other vegetation predictors better described the vegetation within that region.

Habitat selection and fitness

Nest success was explained only at the landscape scale but variables affecting nest success were either unimportant in site selection (e.g. NDVI) or had contrary effects on selection and fitness (e.g. nest density). The unimportance of habitat and disturbance variables in determining nest success may reflect the unimportance of predation relative to food availability in determining nest success. This may also be due to restriction to WMAs constraining the ranges of the variables, particularly those relating to human disturbance that did not affect fitness in this study despite being shown to do so in other vulture species (Donázar et al. 2002; Morán-López et al. 2006a). Humans are the only major cause of mortality of vultures, and this has a major effect on species distributions (Bamford et al. 2009).

WbV nest success showed negative density dependency (as in other Gyps species; Fernendez, Azkona & Donazar 1997), yet there was a strong selection for areas already containing nests. Nest site selection is more complicated than simply selecting advantageous habitat (Chalfoun & Martin 2007). Food availability during the breeding season is unknown when sites are selected, perhaps explaining the unimportance of NDVI. Gyps vultures do re-use nests but often abandon a site after breeding failure (Sarrazin et al. 1995). WbV may select sites close to conspecifics as an indicator of the quality of the area, as areas with nesting pairs are likely to have hosted successful nesting attempts in previous years. In long-lived, slow-breeding birds such as vultures, adult survival is more important than breeding success in determining population levels; thus, given a food shortage, adults would be expected to value their own survival more greatly than that of their chicks. This would make nest success, although relatively unimportant in species demography, a more sensitive indicator of habitat suitability than adult survival.

Applicability of models

Our models were restricted to sites within WMAs, which cover a small proportion of most African countries, and this has implications for their applicability. Both species we considered are close to extinction outside of protected areas across much of Africa due to human persecution (Thiollay 2006; Bamford et al. 2009) and lack of food; their current distribution is strongly influenced by land management and policing, raising the possibility that nests now occur in non-ideal, but protected, habitat in preference to unprotected but ecologically ideal habitat. Our models assume that adequate protection for nesting vultures exists in WMAs and also that the major food source derives from wild populations of ungulates. Food was estimated directly using NDVI, although a model of ungulate responses to NDVI may be better (e.g. Rasmussen, Wittemyer & Douglas-Hamilton 2006), but this only applies when the presence of wild ungulates can be assumed and carcasses are not removed. Data on carcass availability would constitute a superior estimator but is time-consuming to collect, even in small areas (Murn & Anderson 2008). The small number of nests outside of WMAs makes the effect of human persecution difficult to model (Bamford et al. 2009). As such, our models may only be applicable within WMAs, although, as the habitat preferences of both species appear to be constant across their ranges, the models could be used to predict the occurrence of habitat that could be utilized for nesting should there be sufficient food and little persecution (Bamford et al. 2009).


Habitat-association models of species distribution are not always applicable in regions other than those used in model development. This limitation can arise without the habitat preferences of focal species varying across their ranges as it may be due to the methods of model development. Habitat-association models require easily available and updateable habitat data to be useful management tools, but for species with large ranges this often means using GIS data, which may only be broadly related to the landscape that affects the focal species. Although the use of proxy measures is unavoidable, particularly for endangered or poorly studied species with large ranges or in remote areas, effort should be made to asses their adequacy as our results show this can vary substantially between regions. Models thus tend to emphasize dominant landscape features of the region used to develop them and the use of models to inform management decisions in new regions should be undertaken with caution, especially with poorly studied species where there is little information to inform the choice of explanatory variables.

Models developed from multiple regions overcome this problem to some extent. Multi-region models are in general more transferable as they place less emphasis on dominant landscape features and identify predictors that do work in many regions, but at the cost of reducing the fit within the regions used to develop the model. Such models are therefore of limited use in informing local management decisions, but are likely to be more useful for identifying areas on which to focus in regions that have not been surveyed, and for identifying suitable habitat for protection in species, like vultures, in which protection is increasingly vital to survival.


This paper is dedicated to our colleague and co-author James Wakelin who died before the manuscript was submitted. We thank Roger Payne (VSN International) for the modified GenStat code. We thank S. Thirgood, M. Whittingham and two anonymous referees for helpful comments. AJB was supported by a NERC (UK) studentship. Assistance with the fieldwork has been provided by many people. Swaziland: Mickey Reilly, Ngwane Dlamini, Dave Ducasse and Alan Howland allowed us access to reserves/ranches; Mduduzi Ngwenya and All Out Africa volunteers assisted with the fieldwork. Zululand: Doug van Zyl undertook much of the survey work. Kimberley: Assistance was provided by De Beers, Department of Tourism, Environment & Conservation, McGregor Museum, Hawk Conservancy Trust, and Birds of Prey Working Group of the Endangered Wildlife Trust, Johannesburg and especially Tania Anderson, Eddie McFarlane, Julius Koen, Cornè Anderson, Beryl Wilson, Campbell Murn, Andy Hinton, Graham Main and Johan Kruger. Namib: Nedbank Go Green Fund, SGA Chartered Accountants, and the Birds of Prey Working Group funded the fieldwork. Linyanti and Makgadikgadi: The Department of Wildlife and National Parks is thanked for their support of BirdLife Botswana’s vulture monitoring work; M. Muller and B. Bridges financed and undertook most of the vulture survey work.