Analysis of only one of the clusters found? I conducted a Cluster Analysis (CA) on my data and found 3 clusters that make theoretical sense. The average silhouette width is low (0.1), which according to Kaufman & Rousseeuw (1990) is interpreted as "no structure". However, the first cluster's silhouette width is 0.75 (Strong structure), but the other 2 have a low silhouette width - thus receiving a low ASW.
I want to analyze the clusters on my covariates (gender, age, etc.), for example, see if women use different cluster types than men. 
My question:Can I only use the first cluster type for my covariate analysis, and disregard the other two?
 A: You need to realize what a low silhouette means:
These points may as well belong to another cluster.
For example, they might also belong to the first cluster that appears to be well. But maybe is incomplete?
Silhouette is not perfect either. It assumes that distances to clusters are comparable. But if clusters have very different size (or non convex shapes) this is problematic. Just consider two Gaussians, N(0;1) and N(100;50). Draw 1000 points from each, and computer the Silhouette of this "correct" solution. You'll see it works fairly bad. The reason why Silhouette fails badly is that points at, say, x=40 are much more likely generated by the high-variance Gaussian. They are 40 standard deviations from the first, but only 1.2 standard deviations from the second Gaussian. Yet, the average distance to the left cluster is around 40, and to the right around 60, so the Silhouette is around -0.333.
Given that you have mixed data, your initial distance function probably is not reliable enough to draw any statistical conclusion, either.
