At 11:34 23.6.2000, you wrote: >Hi everyone, >can someone tell me the limits on the number of variables in relation to >sample size? Are there any good references on this topic? Thanks in advance, >Winnie The good empirical rule for discriminant analysis (classification) is: N>=p, where N is the numer of observations (sample size); p is the number of variable. However this rule is appropriate for two classes. In general the minimum sample size depends on the procedure you will use. The parametric statistical procedures require less N, while the nonparametric ones require more N. But you have to use as more as possible training observations, except if you have tremendous data set. I any case you need some unbiased estimation of the classification accuracy (cross-validation; leave-one-out; test sample) in order to determine the particular classifier's efficiency. And the most important questions are: a) selection of variables (the best subset) Even you have many variables and moderate N, you could use different variables slection procedures and you will decrease p b) choice of an appropriate discriminant procedure Ognian Asparoukhov -- Ognian Asparoukhov Phone: ++(359) 2 700-528 Centre of Biomedical Engineering ++(359) 2 700-326 Bulgarian Academy of Sciences Fax: ++(359) 2 723-787 Acad. Georgi Bonchev Street, Bl. 105 E-mail: [log in to unmask] 1113 Sofia, BULGARIA [log in to unmask]