How do you setup your deciles? Do you use PROC RANK and TIES=LOW option to
prevent the same score being split over 2 deciles?
Also, why just Test vs Control? Why not set up an experimental design to
test multiple factors?
JP MORGAN CHASE BANK
On Tue, 6 Sep 2005 15:42:03 -0400, Talbot Michael Katz <topkatz@MSN.COM>
>As Peter Luo has advised you, you should choose your control group from
>across your entire target population, if you want to measure the
>difference between using a model and targeting at random. You might wish
>to check out a previous SAS-L thread concerning "Control Group Assignment"
>from the second week of July. I gave the standard formula for computing
>sample size needed to discriminate between the test and control response
>But you will also want to pay close attention to David Cassell's critique
>of that (http://listserv.uga.edu/cgi-bin/wa?A2=ind0507B&L=sas-
>-- TMK --
>"The Macro Klutz"
>On Tue, 6 Sep 2005 13:35:17 -0500, Nick . <ni14@MAIL.COM> wrote:
>>I have a question regarding the selection of a CONTROL GROUP in a direct
>marketing campaign (banking) setting.
>>I have built a (logistic) model and I have deciled my PROSPECT population
>from DECILE1 (low probability of rersponse) to DECILE10 (high probability
>of response). For example, if I have a prospect population of 1 million,
>then each decile gets 100K. I want to campaign to DECILE7 through
>DECILE10. At the same time, I wish to put aside a CONTROL GROUP to see how
>the model will perform against it. Here is my thought:
>>Since I have 1 million prospects, I will select 400K prospects--100K from
>DECILE7, 100K from DECILE8, 100K from DECILE9, and 100K from DECILE10.
>That's 400K, the population I will campaign to. I want to put aside 10% as
>a control group.
>>Do I randomly select the 10% from the 400K records (so I will end up
>campaigning to 360K) or do I randomly select the 10% from the 1 million
>>If I select the 10% from the 400K prospects (my high probability of
>response prospects according to the model), execute the campaign, the
>results come back and I see a response rate from the campaign (i.e. the
>model) and the control about the same, say about 5% response for the
>mailed (modeled) group and the control group, then am I saying that the
>model is not a good model because it has not provided a higher response
>than the control?
>>If I select the 10% from the 1 million prospects (my high probability of
>response prospects), execute the campaign, the results come back and I see
>a response rate from the campaign (i.e. the model) and the control much
>different, say 5% response for the mailed (modeled) group and 2% for the
>control group, then am I saying that the model is a good model because it
>has provided a (much) higher response than the control?
>>I guess, I don't quite know how to select the control group so as to
>compare model response rate to control group response rate.
>>Your thoughts are much needed and appreciated.
>>Sign-up for Ads Free at Mail.com