SEARCH

SEARCH BY CITATION

Keywords:

  • Area under the ROC curve;
  • calibration;
  • classification;
  • contingency matrix;
  • discrimination;
  • probability;
  • reliability;
  • species distribution modelling;
  • uncertainty

Abstract

Aim

When faced with dichotomous events, such as the presence or absence of a species, discrimination capacity (the ability to separate the instances of presence from the instances of absence) is usually the only characteristic that is assessed in the evaluation of the performance of predictive models. Although neglected, calibration or reliability (how well the estimated probability of presence represents the observed proportion of presences) is another aspect of the performance of predictive models that provides important information. In this study, we explore how changes in the distribution of the probability of presence make discrimination capacity a context-dependent characteristic of models. For the first time, we explain the implications that ignoring the context dependence of discrimination can have in the interpretation of species distribution models.

Innovation

In this paper we corroborate that, under a uniform distribution of the estimated probability of presence, a well-calibrated model will not attain high discrimination power and the value of the area under the curve will be 0.83. Under non-uniform distributions of the probability of presence, simulations show that a well-calibrated model can attain a broad range of discrimination values. These results illustrate that discrimination is a context-dependent property, i.e. it gives information about the performance of a certain algorithm in a certain data population.

Main conclusions

In species distribution modelling, the discrimination capacity of a model is only meaningful for a certain species in a given geographic area and temporal snapshot. This is because the representativeness of the environmental domain changes with the geographical and temporal context, which unavoidably entails changes in the distribution of the probability of presence. Comparative studies that intend to generalize their results only based on the discrimination capacity of models may not be broadly extrapolated. Assessment of calibration is especially recommended when the models are intended to be transferred in time or space.