Hierarchical Cluster Analysis – Various Approaches to Data Preparation


No 3/2013, September

Abstract

The article deals with two various approaches to data preparation to avoid multicollinearity. The aim of the article is to find similarities among the e-communication level of EU states using hierarchical cluster analysis. The original set of fourteen indicators was first reduced on the basis of correlation analysis while in case of high correlation indicator of higher variability was included in further analysis. Secondly the data were transformed using principal component analysis while the principal components are poorly correlated. For further analysis five principal components explaining about 92% of variance were selected. Hierarchical cluster analysis was performed both based on the reduced data set and the principal component scores. Both times three clusters were assumed following Pseudo t-Squared and Pseudo F Statistic, but the final clusters were not identical. An important characteristic to compare the two results found was to look at the proportion of variance accounted for by the clusters which was about ten percent higher for the principal component scores (57.8% compared to 47%). Therefore it can be stated, that in case of using principal component scores as an input variables for cluster analysis with explained proportion high enough (about 92% for in our analysis), the loss of information is lower compared to data reduction on the basis of correlation analysis.

Keywords

Hierarchical clustering, PCA, correlation, Pseudo t2, Pseudo F Statistic, e-communication, Internet satisfaction index, Mobile phone satisfaction index.

Full paper

  Full paper (.pdf, 389.62 KB).