tag:blogger.com,1999:blog-38857904.post8744842293492745081..comments2008-07-09T13:29:01.163Znjrhttp://www.blogger.com/profile/08980758986023344486noreply@blogger.comBlogger2125tag:blogger.com,1999:blog-38857904.post-86590278997039027162008-07-09T13:29:00.000Z2008-07-09T13:29:00.000ZHi Anonymous<BR/> <BR/>You didn't leave an email address, so who knows whether you'll see this, but...<BR/> <BR/>You raise an general point. For an indivual, we cannot know the probability. Exactly as you say, an individual either responds or doesn't respond. It's almost a philosophical question as to whether that individual even really *has* a probability other than 1 (if (s)he responds) or 0 (if (s)he doesn't).<BR/><BR/>However, we can estimate a probability by using either full-blown modelling techniques (regression, for example) or simply by grouping customers.<BR/><BR/>For example, suppose we know each customer's gender and it transpires that 20% of females purchase and only 10% of males do. Then one simple estimate we can make is that, based on gender alone, the probability that a female will purchase is 20% and the corresponding probability for a male is 10%.<BR/><BR/>This is almost as simple as it gets. (We could simplify even further by not even looking at gender and just observing the overall purchase rate. Maybe this is 14%, in which case that would be a simple estimate we could apply to anyone.<BR/><BR/>Hopefully this makes sense.<BR/><BR/>Broadly speaking, this is what "response modelling" does (though usually using something like logistic regression or a decision tree method like CART, rather than simple averages like the above.<BR/><BR/>Penetration models do something similar except that they don't look at people who've been mailed (or otherwise treated); they just look at which customers have a given product. So, using the same numbers but in a different context, we might measure how many of our customers have a given product (not particularly after or as a result of any particular marketing campaign). And then again, we might observe that 20% of females have the product and 10% of males have the product.<BR/><BR/>Personally, I think it's even more dubious to use a penetration rate/model to predict campaign response than to use a response model, but the point of this piece is that neither is really what you want, because the thing that should really be of interest is the EXTRA sales that result from a campaign --- i.e. those that wouldn't have happened without the campaign.<BR/><BR/>Hope this helps (and that you see this response.)<BR/><BR/>Nicknjrhttp://www.blogger.com/profile/08980758986023344486noreply@blogger.comtag:blogger.com,1999:blog-38857904.post-71266913394719933522008-06-29T16:56:00.000Z2008-06-29T16:56:00.000ZI'm curious about how you model the probability. From a campaign, if we can only get whether a person responded or not - how do we actually know the probability of him purchasing?Anonymousnoreply@blogger.com