Question & Answer
How can I run a cluster analysis solution using centroid clustering and Ward's method, using Mahalanobis distance as the distance measure?
It is actually physically possible to construct a matrix of distances of whatever type "by hand" or programmed computations, and dress that up as a matrix file with a ROWTYPE_ of PROX to feed into CLUSTER. However, the Ward, median and centroid methods in SPSS CLUSTER should only be used with squared Euclidean distances. When you start with a distance matrix and join from that, recalculating distances using the joining formulas in our algorithms (which are taken from Anderberg's 1973 Cluster Analysis for Applications) the results will only match those from recalculating distances using raw data for these methods if you use squared Euclidean distances. This is why we say in the manuals to only use this measure with these methods. So it is technically possible to get CLUSTER to work on a matrix of Mah. distances, but it's not advisable to do so.
16 June 2018