Tim Berry's otherwise excellent article (“Developing an Acquisition Modeling Strategy,” April 20) contains a minor technical flaw. He recommends multiplying the predictions from a response model with the predictions from a prospect value model to produce a “combined score” that estimates the expected value of mailing a prospect.
Mathematically, this multiplication assumes the two predictions are independent. However, these predictions may be highly and positively correlated: Prospects most likely to respond are often the prospects most likely to spend the most.
Berry's method can fail to recognize the full value of the better prospects, perhaps scoring them into lower deciles. A simple correction known as the Durbin-McFadden factor can be applied to fix this problem (see Durbin & McFadden, Econometrica, Vol. 52, No. 2).
Applying the fix is almost trivial and could significantly improve results in large mailings. At least two catalogers now use this method at my suggestion.
Alan Rimm-Kaufman, Ph.D.
Director of marketing research
Crutchfield Corp.,
Charlottesville, VA
