Leo,

I'd like to take you up on your offer to handout your SAS code.

Since you posted this quite awhile ago, I suppose you have already
received many responses, but I'll add mine here.  Coincidentally, I
used to work with Victor and can attest that he is an astute
individual and usually correct in what he says.  In this paper in
particular, I would also concur with his conclusions, although there
are some alternative ways of approaching the problem.

As Victor, points out when you are faced with a significant response
rate among customers who will NOT receive the promotional campaign,
the model should focus on the 'incremental' lift of the campaign.  My
standard approach would have been to include variable interaction
parameters between the prior mail campaign and the other predictors.
This approach allows the interaction by interaction testing of
significance and therefore only selects those interactions that are
highly likely to be incrementally effective.  The 'true lift' would
therefore be the incremental lift of just the interaction parameters.

Victor's approach is to create two models independently, one for
customers who received the prior campaign and another for those who
did not, and then compare the two models.  Victor's approach has some
advantages and disadvantages over a single model with variable
interactions.  On the negative side, it measures all differences
between the models as 'true lift' which is equivalent to saying that
all the variable interactions were statistically significant.

On the other hand, his approach can be applied to any methodology,
including techniques like CART, Neural Networks, etc. that do not
allow the analyst to specify specific variable interactions.  Another
advantage of his approach is that it is possible that the incremental
effect of the promotional campaign might exhibit a complex
relationship that would not be captured by testing one variable
interaction at a time.  His approach might or might not be easier to
evaluate also, but it depends more on the situation and how many
variable interactions there turn out to be.

Those are my thoughts.  Hope some of it was useful.

-- 
Best regards,

David Young
Marketing and Statistical Consultant
Madrid, Spain
+34 913 540 381
http://www.telefonica.net/web2/davidyoung



Saturday, May 3, 2008, 3:26:04 AM, you wrote:

LG> Hi All,

LG> I wanted to share this and, if possible, know your thoughts regarding the
LG> following:

LG> The paper in the link below proposes a methodology of target selection in
LG> situations where there is a high probability of response regardless of
LG> whether the customer receives the campaign contact (e.g. mail, call). I
LG> replicated the simulation outlined in the paper to measure what the author
LG> calls "True Lift" for the current and proposed methodology. I coded this in
LG> SAS which I'll be happy to send to you in case you are interested.

LG> http://www.sigkdd.org/explorations/issues/4-2-2002-12/lo.pdf

LG> Have anyone used this methodology before, or an alternative one for this
LG> situations?

LG> Many thanks in advance for your thoughts.

LG> Regards,

LG> Leo.