Help System (web edition)

CLASSI <data> 
classifies observations in Survo <data> to g groups
according to Mahalanobis distances and derived measures.
The groups are defined by CORR and NSN specifications
of the form
   CORR=CORR1,CORR2,...,CORRg
   MSN=MSN1,NSN2,...,MSNg
giving the correlation matrices and matrices of means and
standard deviations. These matrices are usually computed
by the CORR operation for g different groups with same
variables and transformed into corresponding matrices of
canonical discriminant functions (discriminators) with
lower dimensions by the /DISCRI (sucro) operation.
When discriminators are used as the basis for classifi-
cation (this is strongly recommended), the specification
   COEFF=DISCRL.M
must be included since these coefficients transform the
original variables into discriminant scores.

The classification is based either on Mahalanobis distan-
ces or Bayes probabilities (assuming that the samples are
multivariate normal). The classification rules are
selected by activating variables in <data> as follows:
  D = Mahalanobis distances, equal covariances
  d = Mahalanobis distances, unequal covariances
  B = Bayes probabilities, equal covariances
  b = Bayes probabilities, unequal covariances
In B and b alternatives, numbers proportional to prior
probabilities are give by a specification
  PRIORS=P1,P2,...,Pg .
Default is PRIORS=N1,N2,...,Ng where Nk is # of observa-
tions in the k'th group (taken from the MSN file).

In alternative B, posterior measures used in
classification are computed as
  Pk*exp(-0.5*Dk^2)
where Dk^2 is the (squared) Mahalanobis distance.
In alternative b, the corresponding measure is
  Pk*exp(-0.5*Dk^2)/sqrt(det(Sk))
where Sk is the covariance matrix in group k.

All above rules can be used simultaneously by indicating
unique D,d,B,b variables.
Also posterior probabilities (or distances in cases D,d)
can be saved in g variables activated by P's.
This, however, is possible only for one of the alternatives
at a time. The precedence order is b,B,d,D.

 1 = More information on additional multivariate operations 
 M = More information on multivariate analysis