Mensural Discrimination of Peromyscus 115 



PW). After a varimax rotation, the variables most highly correlated 

 with the first discriminant function were SL (0.74), TTL (0.74), BNL 

 (0.66), PFL (0.66), RL (0.65), PL (0.62), Foot (0.59), MTL (0.53), 

 ZB (0.53), and NL (0.50); the only variable highly correlated with the 

 second function was Tail (0.82, all others were less than ±0.17); at 

 -0.33, OC was most highly correlated with the third function. All the 

 misclassifications of the data were in separating P. leucopus and P. 

 maniculatus (Fig. 3b). This observation led us to implement a two- 

 step discrimination process as suggested by Thompson and Conley 

 (1983). First, we grouped P. leucopus and P. maniculatus and performed 

 discriminant analysis among P. gossypinus, P. polionotus, and P. leucopus- 

 P. maniculatus; then we separated P. leucopus and P. maniculatus. 

 However, this scheme did not improve the classification results. 



In analysis of species pairs, at least 98% of specimens could be 

 separated using only one or two external and/or skull measurements 

 (Table 2). In pairwise comparisons using only skull measurements, 

 we could separate at least 95% of the specimens (except for P. leucopus- 

 P. maniculatus). For this species pair two characters separate 82% of 

 the specimens. The scores generated by the discriminant functions 

 (Table 2) approximately fall on either side of zero, such that scores 

 for one species are positive, and scores for the other species are 

 negative. However, these models do generate a few misclassifications; 

 therefore, specimens with scores near zero (e.g., ±0.5) should be subjected 

 to the full discriminant models. 



DISCUSSION 



Discrimination of these Peromyscus species is difficult when 

 collection location information or skins are missing, and we did not 

 achieve the ultimate goal of this project which was to correctly classify 

 any skull without external information. However, the great majority 

 of specimens can be correctly assigned to species, and the discriminant 

 function was useful in identifying likely misclassified and questionable 

 specimens in our museum collections. Additionally, the function allows 

 evaluation of specimens collected at the periphery of species' ranges. 



The model using external and skull characters was reasonably 

 successful in classifying the test specimens, which suggests that we 

 captured enough of the variation within each species to make it useful 

 in classifying specimens from somewhat beyond the geographic distribution 

 of our samples. This is an improvement over the ad hoc approach 

 where each state or region requires a different discrimination model. 

 However, although the P. maniculatus test specimens classified correctly, 

 they tended to fall in the margins of the discriminant score distributions. 

 The model with only skull measurements was less successful in classifying 



