When the discriminant model is run various outputs are provided including the percentage of prospects who were correctly assigned and, most importantly, a “structure matrix” which shows which variables are correlated most highly with group membership. Because we use the analysis to create lists for direct mail, or outbound telemarketing, we are primarily interested in the probability of buying the product or service we are marketing. Separate models are created for each product. By sorting records from highest probability score to lowest and converting each decile’s average probability score to an average sales rate (based on a small pilot campaign or recent history), we can estimate the number of new buyers that will result from a direct mail, e-mail or telemarketing campaign and the acquisition cost associated with each decile.

While models based on data for zip codes areas are less precise than models built on block group or household specific data, significant amounts of geo-demographic data can be downloaded FOR FREE from government sources such as the US Census bureau and NOAA (for weather data). Household specific demographic and life stage segment data from firms like Acxiom, Equifax, etc. significantly improve model accuracy but at a higher cost.

For more information, I recommend you get one of David Shepard Associates' excellent books on this subject, take one of his seminars (through DMA), or visit my site at ericnovakandassocaites.com.