Null-model simulations.

Number of IEM models with accuracy not better than expected by chance among the 100 built on the 100 learning data sets. We counted the number of sites turned from absences to presences by the IEM procedure. The same number of sites predicted as absences by EM were randomly selected and replaced by presences. The accuracy of the resulting model was evaluated using the percentage of well-predicted sites, the TSS and Kappa indices. The accuracy was considered as lower if at least one of the three indices of the IEM was lower than that evaluated on the random predictions. The random sampling was repeated 10000 times.