Abstract
A new method of error-rate estimation in linear discriminant analysis appropriate to situations with several small groups is introduced. It is compared with three other methods that can be generalized from the two-group situation. Simulation studies show that in most situations it outperforms the other techniques.