How do I assess quality of hierarchical clustering?

Question

[reubenmcgregor88]

Hi all,

Very generic question and perhaps a bit of a strange one.

I have some data that includes cases and controls. I have created a heatmap and tried some hierarchical clustering using the morpheus software (https://software.broadinstitute.org/morpheus/).

What is clear is that no clustering occurs between cases and controls, and this is in fact what we expected and want to show. We want to show that the measured variable does not distinguish between cases and controls.

Apart from saying visually no clear clusters emerge are there ay objective measures that can be used to say no significant clustering occurs between two groups?

Any pointers or reading much appreciated, I am even struggling to think about what search terms to use to look this up.

Thank you

[Joe]

Are the heat maps on equivalent scales?

[reubenmcgregor88]

Yes all the data is one measure on the same scale.

Answer 1

[raunakms]

You could perform bootstrapping experiment (i.e. permuting the data for a large number of times) and then compute the bootstrap p-values for the dendogram. This will provide you p-values of each nodes in the dendogram indicating the confidence on the dendogram structure. If you are using R, the R-package pvclust has in-built functions to perform the bootstrap and visualizing the dendogram.

[reubenmcgregor88]

Thank you for the reply, and I will be using R indeed so pvclust seems very useful!

So this would tell me how robust the nodes in the dendrogram are? Is there a way of carrying out hierarchical clustering but forcing the "cases" and "controls" into groups? If I have interpreted correctly, in this way I could use bootstrapping as suggested to generate a p-value and in my case show there is no good clustering between the two groups?