Model Selection for Major Giving

This past couple weeks have been active on Prospect-DMM, the discussion group for data mining in fundraising. One discussion called into question the use of certain modeling techniques for major giving. I thought my brief response might be useful to this forum.

From Prospect-DMM:

For identifying major giving prospects, it makes sense to try various avenues.

A reason binary logistic regression or a decision tree on a binary variable (C5) maybe effective is because of the nature of major giving donors.

Certainly, a pathway to major giving via increasing annual support is a pattern found among major donors. These donors might be making gifts informed by cash flow (what can I afford to give this year?). And, using different ordinal or linear techniques makes sense.

However, there tends to be a large group of major donors--usually the majority of many files I review--that have very inconsistent "pre-major giving" gift behaviors. They tend to give gifts out of assets and are motivated by an investment frame of mind. Since they are very different from the overall donor population and tend to be a small pool, categorizing these large outright donors as a 1/0 makes a lot of sense.

Only using factors related to levels may miss many new opportunities (it also may not - each data file is different). Likewise, undocumented planned gift donors and future volunteers are difficult to predict using other modeling techniques.

I would rather be equipped with a large arsenal of techniques and cater my approach/es to the specific modeling need. In major gift modeling, this may require diversifying the models to find as many new leads as possible.