Model selection criteria affect measures of temporal variation in animal–landscape regression models




Temporal variation in animal–landscape relationships hinders management decisions and development of conservation policy. Animal–landscape relationships may be modelled using several model selection criteria (MSC), but how such criteria affect measures of temporal variation of modelled relationships is unclear.


Appalachian Mountains, USA.


We used results from linear and logistic regression models for 98 bird species over a 5-year period to compute two measures of interannual variation in bird–landscape relationships: model consistency reflected the degree to which variables and directions of relationships in a model were the same. R2 coefficient of variation (R2 CV) reflected the degree to which the strength of relationships was the same. We calculated both metrics for each species' models after using four common MSC: AICc, BIC, stepwise-0.15 and adjusted R2.


Although positively correlated (= 0.43−> 0.99; = 0.030−< 0.001), interannual variation metrics for species differed up to 80% among different MSC. One exception was that using stepwise-0.15 and AICc resulted in identical bird–landscape models > 90% of the time and thus produced the same interannual variation metrics for those models. MSC affected R2 CV less than model consistency. Model consistency did not differ significantly among MSC. Use of BIC led to the highest R2 CV (greatest interannual variation), and use of AICc and stepwise-0.15 led to the lowest R2 CV.

Main conclusions

Comparisons of interannual variation metrics across studies or regions should be based on models derived from the same MSC. When cross-study comparisons must be made, effects of MSC on interannual variation in the strength of relationships (R2 CV) may be smaller than effects of MSC on interannual variation in the identity of relationships (model consistency). Model consistency metrics could potentially be used in concert with other model evaluation criteria to refine landscape-based species conservation planning.