Can I use HCA-Ward's cluster-centers to run a K-means including a new item, to see to which cluster is more similar to?

Question

Thank you for reading my question. I have an archaeological case-study, that we can call "Site1", that I want to compare with 9 others "Sites" studied by other scholars. For all of them I have 8 economic (indipendent) variables based on the frecuency of certain artifacts. How can I do to explore the 'economnic' variabiliy of this sites, and more specifically to see to which sites my case-study is more similar to?

This are the steps I followed:

• I changed from frecuency to percentages in order to avoid the among sites variability. There are site with a lot of observations and sites with only twenty observations. But I am not interested in absolute frecuency, but in relative ones.

• I calculated Z-values for the entire sample from the percentage values.

• I run a HCA with Ward method, excluding my 'Site1' from the analysis. I obtain a three-cluster solution, with three Sites each one.

• I save the cluster centers.

• I run a K-means analysis using the Ward's cluster-centers, but this time including also my 'Site1'. As result, I see that the cluster-composition is not changed and that my 'Site1' is added to cluster2. So, I can run a One-way Anova and a Tukey to see which variables contribuite at most to each cluster.

Is it a right procedure? Would you suggest some different strategy?

Thanks,

Mark

Answer 1

Why don't you just assign the new observation to the nearest cluster center by least SSQ? Running k-means on a Ward clustering result is like doing a nice steak on the BBQ and then mincing it to make a frozen hamburger.

Technically, you use the cluster centers as new data set, then do a 1NN classification to classify your new observation.