I have a set $A$ that consists of 5 million objects. Each object has a color, size and cost. I'd like to do a cluster analysis, e.g. K-means, on cost. However, it is computationally not feasible to use all 5 million objects to do the analysis.
I have 2 options:
Option 1: Someone else performed a stratified random sample of the set $A$ on color and size. I can take that sample and do my cluster analysis.
Option 2: Perform a simple random sample of size $k$ on the entire set $A$ and then do my cluster analysis on the simple random sample.
My question is from a statistical standpoint, i.e. bias, consistency, etc. is their a preference between options 1 and 2? In other words, is expected that I should get different results from the two options. If I cluster into 2 groups, perhaps with option 1 I'll get $10$ as the dividing point between the clusters and with option 2, I'll get $40$ as the dividing point.