It is slighly different approach, but is useful to do bootstrap to see stablitity of the clusters. You can use R package Pvclust.

library (pvclust)

result < - pvclust(data, nboot=10000)

plot(results)

This is a very hard decision to make. First, you can plot the within group sum of squares along the number of clusters to find the optimum number then cut the tree (dendrogram) accordingly. Then, there are very sophisticated dynamic tree cutting approaches being developed, for example: http://www.genetics.ucla.edu/labs/horvath/CoexpressionNetwork/BranchCutting/ These methods are more involved so you may need to read more about them.