Although the ultimate goal of restoring an endangered species is establishment of a viable population, population viability analyses (PVA) are rarely conducted. If PVAs are constructed, they are rarely validated, particularly for plants. We used 15 years of demographic data from a restored population of the federally threatened Pitcher’s Thistle (Cirsium pitcheri) to develop eight stochastic element selection (ES) and matrix selection (MS) PVA models that differed in their methods used to incorporate environmental stochasticity and the degree to which they incorporated within- and between-year correlations among vital rates (i.e., survival, growth, fecundity). ES models drew vital rates from a specified distribution, while the MS model randomly drew yearly matrices. ES models that incorporate within- and between-year correlations among vital rates are reportedly more realistic. Because population viability projections (i.e., stochastic growth rate (λ_s), time to extinction, population size and distribution) varied among models, we compared predicted to observed population persistence and size to evaluate each model.
Results/Conclusions

Among the ES models, increasing correlations among vital rates resulted in increased variation and decreased mean of projected population size as well as decreased λ_s. Adding between-year correlations had a larger negative impact on projected population viability than using within-year correlations. The only two models that projected λ_s less than one included between-year correlations. Projected population size and λ_s for the MS model were intermediate between ES models including and not including correlations, presumably because the MS model preserves correlation structure among the vital rates. Overall, validation indicated that the MS model provided the best fit between observed population persistences and predicted persistence probabilities when compared to the ES models. Use of ROC curves and logistic regression diagnostic statistics as validation methods were not sensitive to differences between MS and ES models nor among ES models differing in within- and between-year correlations. The validation methods that best distinguished among the models included comparison of predicted and observed short-term population sizes and stage distributions, the deviance test for bias and spread, and persistence probability curves. Because the choice of PVA model and correlations among vital rates can produce opposing projections, we recommend these validation methods to determine if ES models are in fact better predictors of population viability.