Alternative Numerical Procedures

Abstract

To deal with the Mahalanobis algorithm, some limitations must be observed. Two further algorithms exist with identical or very close results. Both are less sensitive to the initial boundary conditions but are more calculation oriented. One algorithm makes use of an orthogonalization process for the variables. This transforms the set of data in linear, independent variables, taking advantage of the fact that the number of variables can be nearly as numerous as the number of objects defining a pattern. The other algorithm is useful if some variables are perfectly correlated. In this case, the two previous algorithms are useless, as some steps of the calculation cannot be performed. This is a crucial property in case of the third algorithm as multicollinearities can contain essential information . For comparison, screening of the variables and root-cause-analysis are performed with the same clinical case study.