Do habitat association models have any generality?: Predicting skylark Alauda arvensis abundance in different regions of southern England

Much of conservation science is based upon determining the use by organisms of different resources. However, the field data used to construct habitat association models generally come from a small number of sites covering a fraction of the area of interest. It is important therefore to assess the generality of those models for species occurring over large geographical areas. In this paper we test the generality of models describing skylark Alauda arvensis abundance across farmland in southern England in relation to crop type, crop structure and field structure (i.e. height of surrounding boundaries). Skylarks responded to most predictors we measured in similar ways across three regions of differing farming practices (arable-dominated, pasture-dominated and a mixture of the two). Most of the regional differences in habitat associations could be related to differences in the speed of crop development. For example, the sowing of cereals in spring, a much lauded strategy to increase skylark populations, is likely to have less of an effect in regions where cereal development is slow than in regions where it is fast. Most studies that explicitly test the performance of a model developed in one place elsewhere use presence/absence models. We adopt a more sensitive and novel approach by using counts. We found regression equations developed in one region performed poorly when tested as a direct predictor (i.e. a 1:1 relationship) on data from other regions. However, the skylark territories observed in any one region were positively correlated with the territory numbers predicted by models built using data from other regions, so models were good predictors of relative abundance. The results suggest that, for this species at least, conservationists should have confidence when advocating management strategies based upon habitat-association models and extrapolating their generality to other areas. However, our results warn against using regression equations developed in one place to make absolute quantitative predictions elsewhere. Decision-makers should beware of using models in this way.