I am working on a segmentation model that has been fed into the company by an external agency which created the segmentation based on surveys. I have a number of 6 attitudinal segments and after ascribing them to our database via proxy variables, I wish to test if it makes sense from a statistical point of view to further split (or combine for that matter) the segments.

What kind of methods should I consider applying to tackle this? I have limited experience with clustering (k-means), and not so much from a statistical point of view, but more from a perspective where I'd only care about measures of predictive accuracy and not so much about exploratory modelling, so any directions would be helpful.

Many thanks!

P.S. The original segmentation will have to stay unchanged (business requirements...)

  • $\begingroup$ KNN clustering? Are you sure you don't mean classification? $\endgroup$ Mar 4 '19 at 19:39
  • $\begingroup$ Thanks, edited :) Shouldn't have listed kNN $\endgroup$
    – Nolatar
    Mar 6 '19 at 12:55

There have been many 'tests' proposed in literature. But if you apply them to the business-given clusters it will likely not consider them to be good either...

Popular tests include AIC and BIC that trade point approximation vs. model size. So of you need just a few more parameters (clusters) to get a much better model of your data, then these clusters are good to keep.

You could try an ANOVA test, but as k-means optimizes this, the significance values will be way too extreme. Beware of multiple testing and "data snooping" when trying to judge the significance of cllusterings.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.