The Influence of Probability of Detection When Modeling Species Occurrence Using GIS and Survey Data By

Databases of species observations from surveys and ad hoc observations are frequently maintained by managers for a particular area. Obtaining information from this type of survey data about the habitat associations of species can be an efficient method of predicting habitat suitability across a landscape. Many multivariate statistical methods have been used to develop models of habitat associations. The objective of this study was to compare the effectiveness of discriminant function analysis, logistic regression, ecological niche factor analysis (ENFA), and Bayesian logistic regression methods to predict species occurrences using a large database of species occurrences. The four methods were used with the same data set to create models of habitat suitability for 29 amphibian and reptile species. Sensitivity, specificity, and Kappa statistics were used to assess model discrimination (the ability to correctly classify locations as present or absent) and compare the four modeling methods. Overall, model discrimination was poor and did not differ among modeling methods. Rana virgatipes, the carpenter frog, was the only species that was modeled with a greater than 0.4 Kappa statistic using all four methods. Discriminant function analysis and ENFA predicted presence locations with greater success than logistic regression and Bayesian logistic regression methods. Logistic regression and Bayesian logistic regression correctly predicted absence locations more frequently than discriminant function analysis. All three methods predicted absences more successfully than ENFA. The large number of absence locations and method of defining absence locations influenced the model results. Without knowledge of the detection probability of amphibian and reptile species, it is unknown if the absence locations truly represent absences instead of non-detections. ENFA methods, which only use presence data, are better suited to the survey data set. If information about the detection probability of species exists, the Bayesian modeling methods can be modified to account for detectability. Overall, database collections of species occurrences appear to have limited use as data sources for habitat suitability modeling.