SPSS TWOSTEP CLUSTER–A FIRST EVALUATION∗JohannBacher†, KnutWenzig‡, Melanie Vogler§Universit¨at Erlangen-N¨urnbergSPSS 11.5 and later releases offer a two step clustering method. According to the authors’knowledge the procedure has not been used in the social sciences until now. This situationis surprising:The widely used clustering algorithms, k-means clustering and agglomerativehierarchical techniques, suffer from well known problems, whereas SPSS TwoStep clusteringpromises to solve at least some of these problems. In particular, mixed type attributes can behandled and the number of clusters is automatically determined. These properties are promising.Therefore, SPSS TwoStep clustering is evaluated in this paper by a simulation study.Summarizing the results of the simulations, SPSS TwoStep performs well if all variables arecontinuous. The results are less satisfactory, if the variables are of mixed type. One reasonfor this unsatisfactory finding is the fact that differences in categorical variables are given ahigher weight than differences in continuous variables. Different combinations of the categor-ical variables can dominate the results. In addition, SPSS TwoStep clustering is not able todetect correctly models with no cluster solutions. Latent class models show a better perfor-mance. They are able to detect models with no underlying cluster structure, they result morefrequently in correct decisions and in less unbiased estimators.Key words:SPSS TwoStep clustering, mixed type attributes, model based clustering, latent class models1 INTRODUCTIONSPSS 11.5 and later releases offer a two step clustering method (SPSS 2001, 2004). Accordingto the authors’ knowledge the procedure has not been used in the social sciences unti...
