Response modeling generates a formula that predicts tendency to respond to a certain ‘target’ customers set. For this example, we will use a life insurance policy mailing, where the top 100,000 candidates will be mailed. The model ultimately leverages variables found to explain how current policyholders differ from the non-policyholders.  In the end the model will provide a score for each record respective of the degree to which those non-policyholders ‘look like’ the actual policyholders. 

In layman’s terms, the modeling process is an extensive and comprehensive series of mathematical computations wherein you define a target audience, in this case existing life insurance policyholders, and compare and contrast them to non-policyholders.  This is done by inputting hundreds of variables that might potentially predict the differences between the groups and slowly reducing the number of variables based on interactions between those variables and the target audience and those variables WITH EACH OTHER.  Adding or removing a single variable from the mix can potentially alter the entire relationship and cause other variable sets to be the most predictive.  You then test the performance of alternate top models against one another using a hold out sample to determine which single model that you predict will perform the best.

The final output of the process is a simple algebraic equation of the top model with each final variable having a specific weight so that when the entire target audience is scored, each record comes back with a score (or weight) of 1-100, with 100 being the strongest score.  (I often speak in terms of deciles, which is breaking the scores, 1-100, into ten equal groups).  Depending on campaign/marketing goals, you then decide how deep to mail into the target group.  One campaign may only mail the top decile, another the top 5 deciles, and another approach might be to mail the top 2 or 3 deciles and test smaller quantities lower in the model sometimes all the way down to the lowest decile to test response rates, validate assumptions, and/or give a fresh audience to build a new model for future campaigns.  This is an ongoing and iterative process with repeated attempts to improve a model’s performance by making adjustments based on actual results. 

For the purposes of approving variables to be involved in modeling, it is important to know that the variables in the model are not used to ‘pre-select’ records - according to any one variable.  Even though one variable may point to a record being unlikely to respond, others may say the opposite.  Because of the interaction of variables in the model formula, each record has a chance to be in the top of the model.  In other words, the modeling process does not consider any variable by itself, so no selects are performed.  In the modeling process, all of the variables work together to create the score very similar to an orchestra – it works beautifully in unison, and one single element may make a small but necessary contribution to the success of the whole.