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 i
t outperforms the other techniques.