Survival of the fittest

Computerworld - Survival analysis could help you predict that one of your best customers is about to jump ship for a competitor. Or it could help you decide whether that costly promotion is really going to be worth it. Or it could help you tailor that next catalog mailing and double your return.

The aptly named analytic technique, also called survival data mining, has been used by doctors for decades to predict the life expectancy of heart-transplant patients and by biologists to assess the probability that a cell invaded by a virus will die within 24 hours. Engineers have long used it to estimate the mean time to failure of a disk drive or a robotic welder. More recently, sociologists and psychologist have started using it to predict when certain types of people will divorce or seek help for depression.

But until recently, attempts to apply survival analysis to business problems have been mostly university projects, says Edward Malthouse, a marketing communications professor at Northwestern University. "Now it's really taking off in the database marketing world—for credit cards, hotels, airlines, catalogs and so on," he says.

Survival analysis refers to a family of "time to event" prediction techniques mathematically geared to problems with the following characteristics:

• They deal with discrete events that will occur to some but not all members of a given population. Certain patients will die, some disk drives will fail, a certain number of prescriptions for Valium will be written, and some of your best customers will desert you.

• They involve time-dependent outcomes. Is that key customer going to cut up his charge card tomorrow (better call him today), next quarter (send him mail) or not in the next five years (leave him be)?

• The outcomes of interest, or "dependent variables," aren't continuous—like income, height or IQ—but are dichotomous. Either the patient will die within six months or he won't. Your customer will leave this year or he won't.

• Outcomes often can be anticipated by trigger events, such as customer complaints.

Randy Collica, a senior analyst at Hewlett-Packard Co., says use of survival mining to understand and predict customer behavior has sprung up in the past couple of years. He says it's really just an extension of older practices in which a company would take its most recent customer data, or data from a time slice such as a quarter or year, and try to predict attrition based on that.

But survival mining puts time itself into the analysis as a variable. "It is a superset technique," Collica says. "Including time as an element is adding that much more information."