Hi, gang.

This is the posting I foretold of :-)

I haven't done a lot of repeated measures work, so I'm looking for some advice (while I'm waiting for Will Potts' book on Survival Data Mining to appear). Here's the set-up. We've got direct marketing campaigns to try to enroll people into a bonus program; obviously, the hope is that this will increase spending and loyalty, because enrollment in the program without increased spending and loyalty is a money loser. Each campaign sends out offers to quite a lot of people. We have campaign data going back for a few years. Looking back through all the campaigns of this particular type since 2003, we have about 4.5 million pieces going out to 2.5 million individuals. So, as you can see, quite a number of people who didn't enroll the first time they got an offer, were included in one or more subsequent campaigns.

Now, I am quite accustomed to one-shot offers in direct marketing. Typically the results of one campaign are used to build predictive models for the next campaign. First the non-responders and responders are classified as a 0,1 dependent variable for a first-stage response / propensity model; then the amount of the spending response is modeled for a second-stage spending model (hey, weren't we just talking about two- stage models in a separate thread?).

We certainly have enough responders to build a response / propensity model from only the most recent campaign, where each offer represents a single individual. But the reason we want to use some of the older campaigns is because profitability is a longer-term issue; we want to examine post- enrollment behavior over several months, or even a year. Consequently, I was given the following proposal. Start with a universe consisting of every offer in each campaign (the 4.5 million), and treat them all as separate independent individuals. In such a case, the data for each appearance of a single person may be quite different; the number of previous offers will have changed, and the recent spending patterns may change. And since, with a 5% response rate, you can build very good response models on samples of 20,000 or 30,000, you may not get a lot of multiple observations for the same people in a particular sample (that's a tough probability to compute).

So, the first question is, will this give a valid response model? I tend to feel that it will underpredict response. If it's not good to treat each offer as an independent individual, what is the best way to deal with it? I could model the response for a single campaign, and score every enrollee on each campaign by that model for use in the profitability model. Does that make more sense? Or is there a better way?

I hope I've presented the situation clearly, and I look forward to receiving your ideas. Thanks!

-- TMK -- "The Macro Klutz"