This could be done by conducting univariate F tests for each independent variable across the intact groups. As it is similar to already discussed techniques, this method is redundant. Instead, it is better to construct linear combinations of the original variables which eliminate intercorrelations between variables. This composite function will best discriminate between the groups on a multidimensional basis.

Note that LDFA is, in many ways, similar to Multiple Regression. More importantly, note that the purposes and techniques of LDFA are very similar mathematically to MANOVA (except that the roles of x and y are interchanged). The composite is determined in the same way, such that the groups maximally differ. The maximum number of dimensions which can be calculated is the smaller value of the following two: (a) the number of groups minus one, or (b) the number of continuous variables. And like before, each composite is formed from the residual of the previous, thereby making each orthogonal.

Like MANOVA, LDFA is tested for significance by the use of Wilk's Lambda. Follow-up analyses are the same as those for MANOVA (see the previous page). An analysis of the significant findings of LDFA can give a Confusion Matrix, a table showing how accurate group membership is predicted by the composite function(s). Each "axis" is labeled by the groups: one axis being the actual group membership, and the other being the group membership predicted by the composite function; the percentages indicated on the diagonal are "hits"--correct predictions- and the percentages that are not on the diagonal are "misses"--incorrect predictions or errors.