Validation tests of predictive models of butterfly occurrence based on environmental variables

Erica Fleishman, R. Mac Nally, J.P. Fay

    Research output: Contribution to journalArticle

    41 Citations (Scopus)

    Abstract

    Ecologists often seek to predict species distributions as functions of abiotic environmental variables. Statistical models are useful for making predictions about the occurrence of species based on variables derived from remote sensing or geographic information systems. We previously used 14 topographically based environmental variables from 49 locations in the Toquima Range ( Nevada, U.S.A. ) and species inventories conducted over 4 years ( 1996–1999 ) to model logistically the occurrence of resident butterfly species. To test the models, we collected new validation data in 39 locations in the nearby Shoshone Mountains in 2000–2001. We used a series of “classification rules” based on conventional logistic and Bayesian criteria to assess the success rates of predictions. The classification rules represented a gradient of stringency in the “certainty” with which predictions were made. More stringent rules reduced the number of predictions made but greatly increased the success rate of predictions. For comparisons of classification rules making similar numbers of predictions, conventional logistic and Bayesian criteria produced similar outcomes. Success rates for predicted absences were uniformly higher than for predicted presences. Increasing the temporal extent of data from 1 to 2 years elevated success rates for predicted presences but decreased success rates for predicted absences, leaving the overall success rates essentially the same. Although species occurrence rates ( the proportion of locations in which each species was found ) were correlated between the modeling and validation data sets, occurrence rates for many species increased or decreased substantially; erroneous predictions were more likely for those taxa. Model fit ( measured by the explained deviance ) was an indicator of the probable success rate of predicted presences but not of predicted absences or overall success rates. We suggest that classification rules for predicting likely presences and absences may be decoupled to improve overall predictive success. Our general framework for modeling species occurrence is applicable to virtually any taxonomic group or ecosystem.
    Original languageEnglish
    Pages (from-to)806-817
    Number of pages12
    JournalConservation Biology
    Volume17
    Issue number3
    DOIs
    Publication statusPublished - 2003

    Fingerprint Dive into the research topics of 'Validation tests of predictive models of butterfly occurrence based on environmental variables'. Together they form a unique fingerprint.

  • Cite this