A distribution-free, multivariate discriminating method. Initial experience with the discrimination of patient groups...

We communicate a distribution-free quasi graphic procedure for obtaining a linear discriminating function. The method is based on the following considerations: Suppose a group A is to be separated from a group B using two parameters X and Y. The centre of each group is defined as the median (or mean or mode) of its points. Of all straight lines passing through any point of A and any point of B those are retained which intersect the segment joining the centres of A and B. For each of these the number of wrongly allocated points is calculated (i.e. the points which do not lie on the same side of the straight line as their group centres). In this way one obtains straight lines with maximal separating power (for the two groups given). Finally, each optimal line is rotated in such a way that its defining points are also correctly allocated. If more than two parameters are available a stepwise procedure can be used: the distance from the separating straight line obtained from the first two parameters is introduced as a new parameter, which is then combined with the third parameter to yield a new discriminating function which depends on all three parameters. Iterating this step one can combine any number of parameters. The method was implemented on a personal computer. It was first applied to a textbook model (chances of survival for M. haemolyticus neonatorum estimated by concentrations of haemoglobin and bilirubin in cord blood).(ABSTRACT TRUNCATED AT 250 WORDS)