Segmentation analysis
Message posted by Stuart (via 195.157.144.29) on November 19, 2001 at 9:20 AM (ET)

Hi all,
Hope someone can help with this one, as I'm driving myself mad.

I vaguely remember someone once describing a type of segmentation approach, where you segment primarily on attitudes (as normal with k-means) but where you also include demographics variables to help make the segmentation more actionable.

The example given was of segmenting according to height. With k-means (k=2) you might result in a segment of tall people and one of short people. If all members of the short group except one were female, and all members of the tall group except one were male, you may wish to alter the clusters according to the "gender" variable so you have a "taller/male" group and a "shorter/female" group. This would make your cluster solution less good in theory (since the clusters would be less well-separated), but it would be better in practice.

Has anyone heard of this sort of approach, and what on earth is it called? I have a feeling it was referred to as a sort of latent segmentation (since gender is not an active segmentation variable), but it's clearly not the same as latent-class segmentation.

Thanks for any ideas!
Stuart