clustering approaches: binary data

Question

I am investigating clinical outcomes following an injury. I have 8 different outcomes, all binary, and there is correlation between them with different subjects have combinations of different outcomes. I have done a PCA of the variables, with loading plot showing grouping of different outcome variables.

I was wanting to examine with a clustering-type approach to assess if there are clusters of patients with certain combination of outcomes. I can then investigate what the biological signatures of these clusters are. I was wondering what clustering approach people on the forum would take. It is 8 binary variables with no continuous variables. All 8 are different outcomes. I mainly work on R.

Many thanks

Answer 1

Standard approaches would be any distance-based clustering (such as complete or average linkage or partitioning around medoids) using an appropriate distance measure. The standard distances would be simple matching and Jaccard (R-function daisy with metric="gower" and type "asymm" or "symm"); there are some more in the literature, but one of these two makes sense in many applications. Jaccard ("asymm") is the method of choice if joint presences (1) are meaningful but joint absences (0) are not; if 0 and 1 should be treated symmetrically, simple matching ("symm") is better. Alternatively, one could fit a latent class mixture model assuming local independence within clusters (R-package poLCA).

I'd absolutely plot a binary heatmap to get some feel for what is going on before deciding what clustering method(s) to take (note however that what is seen from this depends on the order of observations, and a standard approach for ordering them is hierarchical clustering, so this may already involve clustering).