Step 1: Examine the similarity and distance levels
At each step in the amalgamation process, view the clusters formed and examine their similarity and distance levels. The higher the similarity level, the more similar (correlated) the variables are in each cluster. The lower the distance level, the closer the variables are in each cluster.
Ideally, the clusters should have a relatively high similarity level and a relatively low distance level. However, you must balance that goal with having a reasonable and practical number of clusters.
Step 2: Determine the final groupings for your data
Use the similarity level for the clusters that are joined at each step to help determine the final groupings for the data.Look for an abrupt change in the similarity level between steps. The step that precedes the abrupt change in similarity may provide a good cut-off point for the final partition. For the final partition, the clusters should have a reasonably high similarity level. You should also use your practical knowledge of the data to determine the final groupings that make the most sense for your application.
For example, the following amalgamation table shows that the similarity level decreases slightly from step 1 (93.9666) to step 2 (93.1548). The similarity then decreases abruptly in step 3 (87.3150), when the number of clusters changes from 3 to 2. These results indicate that 3 clusters may be appropriate for the final partition. If this grouping makes intuitive sense, then it is probably a good choice.
Step 3: Examine the final partition
After you determine the final groupings in step 2, repeat the analysis and specify the number of clusters (or the similarity level) for the final partition. Minitab displays the final partition table, which shows the variables that form each cluster in the final partition.
Examine the clusters in the final partition to determine whether the grouping seems logical for your application. If you are still unsure, you can repeat the analysis, and compare dendrograms for different final groupings, to decide which one is the most logical for your data.