ABSTRACT
Up to now we have been transforming data matrices to maps where the rows and columns are displayed as points in a continuous space, usually a twodimensional plane. An alternative way of displaying structure consists in performing separate cluster analyses on the row and column profiles. This approach has close connections to CA and decomposes the inertia according to the discrete groupings of the profiles rather than along continuous axes. In the case of a contingency table there is an interesting spin-off of this analysis in the form of a statistical test for significant clustering of the rows or columns.
Partitioning the rows or the columns . . . . . . . . . . . . . . . . . . . . 113 Between-and within-groups inertia . . . . . . . . . . . . . . . . . . . . . 114 Calculating the inertia within a single group . . . . . . . . . . . . . . . 115 Data set 8: Age distribution in food stores . . . . . . . . . . . . . . . . . 115 Clustering algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 Tree representations of the clusterings . . . . . . . . . . . . . . . . . . . 117 Decomposition of inertia (or χ2) . . . . . . . . . . . . . . . . . . . . . . 118 Deciding on the partition . . . . . . . . . . . . . . . . . . . . . . . . . . 118 Testing hypotheses on clusters of rows or columns . . . . . . . . . . . . 118 Multiple comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 Multiple comparisons for contingency tables . . . . . . . . . . . . . . . . 119 Cut-off χ2 value for significant clustering . . . . . . . . . . . . . . . . . 119 Ward clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 SUMMARY: Clustering the Rows and Columns . . . . . . . . . . . . . . 120
