I have a data set that clusters block groups in the US into either 15 broad neighborhood categories or 72 fine-grained segments with goofy names. The segments were constructed using factor analysis using mainly age, income, ethnicity, education, marital status, dwelling type, and the presence of children. The categories take into account urbanization, affluence, age, family status, and ethnicity. The segments are nested in the categories. For example, high income, college educated, married with children category contains 6 segments that differ largely in age and sources of income.
I have 2 types of variables:
- I know the dominant and the second largest (if one exists) segment/category for each block group
- I know the population/households in each category/segment in each block group
I've never worked with this sort of data, but I am curious to see how useful it is compared to using traditional demographics. I am mainly interested in how people employ segmentation data to understand their customers. Other than saying that particular groups are enthusiastic consumers, what are the best practices? Any solid references would be welcome. I've not seen this question discussed a lot at CV.
Some less vague questions:
- Is there a way to avoid throwing away the detailed data without having 72 explanatory variables, most of which will have zeros? I've usually seen people use dominant/secondary segments as dummies, but that seems like a poor choice since it throws away a lot of quantitative info. I am mainly looking into functional form advice here.
- Is there any way to exploit the fact that segments are nested in categories? If I find that some category is a bad customer, while another is great, does it work to separate out the good category into its segments, while treating the bad group as one category? Can you mix the classification schemes?
- Are there any variables that are useful in supplementing the geodemographic groups? Would anything that's not in the original factor analysis, is relevant, and not too collinear with the clusters be OK?