Utilizing the scores from several domain specific metrics, I have run hierarchical clustering to determine where specific domain scores from different metrics tend to cluster. My questions are
1.) What can this tell me about the relationship between each measurement domain?
2.) What are the strengths and drawbacks from this analysis?
In my eyes, the main difficulty is weighting measurements between domains, especially if the number of measurements differs significantly between them. Simply putting them together in one dataset may end up giving a result completely dominated by measurements from one domain. As for their relationship, you can obviously see how the variables cluster in "sample space", but perhaps other correlation metrics will inform better on that.
Strengths are that, if done "correctly", the analysis could give very powerful results, with the potential drawback of what I mentioned above.
Feel free to supply some more detail about the specific setting or your analysis.
Excellent, yes it is 5 surveys, varying from 2 to 5 domains each. From a correlation standpoint, several domains correlate well, which was partially reflected using Wards method, procedurally right out of the SPSS manual. I'm curious if there are other tests I can run on cluster sets which may reveal more.
I may not be totally clear on your method, but in general, here are some challenges with hierarchical clustering:
1) You need to think about linkage method. How are clusters being formed?
2) It is up to you how many clusters are produced, which can introduce bias.
3) It is sometimes difficult to link clusters and descriptors. For this reason it might make more sense to perform some sort of ordination, maybe NMDS, apply your clusters, and apply (potentially) a PERMANOVA or similar (see adonis package in R) to infer how descriptors are affecting similarities.
Thank you Daniel, that may be what I need to do next. I did set the cluster limit in one analysis, and no cluster limit on the second, though I have yet to compare the two, or to really derive a good reason for limiting the clusters. However limiting the clusters did not create significantly different grouping for the first two clusters. I also ran the clusters by cases, rather than by measure, but I am not exactly sure what the significance there might be. Thank you for your input.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS