Error and uncertainty in habitat models


Jane Elith, School of Botany, The University of Melbourne, Parkville 3010, Victoria, Australia (fax + 61 39347 5460; e-mail


  • 1Species distribution models (habitat models) relate the occurrence or abundance of a species to environmental and/or geographical predictors that then allow predictions to be mapped across an entire region. These models are used in a range of policy settings such as managing greenhouse gases, biosecurity threats and conservation planning. Prediction errors are almost ubiquitous in habitat models. An understanding of the source, magnitude and pattern of these errors is essential if the models are to be used transparently in decision making.
  • 2This study considered the sources of errors in habitat models. It divided them into two main classes, error resulting from data deficiencies and error introduced by the specification of the model. Common and important data errors included missing covariates, and samples of species’ occurrences that were small, biased or lack absences. These affected the types of models that could be developed and the probable errors that would occur. Almost all models had missing covariates, and this introduced significant spatial correlation in the errors of the analysis.
  • 3A challenging aspect of modelling is that species’ distributions are affected by processes operating in both environmental and geographical space. We differentiated between global (aspatial) and local (spatial) errors, and discussed how they arise and what can be done to alleviate their effects.
  • 4Synthesis and applications. This study brings together statistical and ecological thinking to consider the appropriate techniques for habitat modelling. Ecological theory suggests models capable of defining optima, while allowing for interactions between variables. Statistical considerations, including impacts of data errors, suggest models that deal with multimodality and discontinuity in response surfaces. Models are typically simple approximations of the true probability surface. We suggest the use of flexible regression techniques, and explain what makes such methods superior for ecological modelling. The most robust modelling approaches are likely to be those in which care is taken to match the model with knowledge of ecology, and in which each is allowed to inform the other.