When are Your Customers Most Likely to Buy?
1st June 2015 | Mike Grigsby
Mike Grigsby Outlines Survival Analysis and How You Can Use it to Determine When Your Customers are Most Likely to Buy
The following is an excerpt from Marketing Analytics.
Survival analysis is an especially interesting and powerful technique. In terms of marketing science it is relatively new, mostly getting exposure in the last 20 years or so. It answers a very important and particular question: ‘WHEN is an event (purchase, response, churn, etc.) most likely to occur?’ I’d submit this is a more relevant question than ‘HOW LIKELY is an event (purchase, response, churn, etc.) to occur?’ That is, a customer may be VERY likely to purchase but not for 10 months. Is timing information of value? Of course it is; remember, time is money.
Beware though. Given the increase in actionable information, it should be no surprise that survival analysis is more complex than logistic regression.
Conceptual Overview of Survival Analysis
Survival analysis (via proportional hazards modelling) was essentially invented by Sir David Cox in 1972 with his seminal and oft-quoted paper, ‘Regression Models and Life Tables’ in the Journal of the Royal Statistical Society (Cox, 1972). It’s important to note this technique was specifically designed to study time until event problems. This came out of biostatistics and the event of study was typically death. That’s why it’s called ‘survival analysis’. Get it?
The general use case was in drug treatment. There would be, say, a drug study where a panel was divided into two groups; one group got the new drug and the other group did not. Every month the test subjects were called and basically asked, ‘Are you still alive?’ and their survival was tracked. There would be two curves developed, one following the treatment group and another following the non-9treatment group. If the treatment tended to work the time until event (death) was increased.
One major issue involved censored observations. It’s an easy matter to compare the average survival times of the treatment vs. the non-treatment group.
But what about those subjects that dropped out of the study because they moved away or lost contact? Or the study ended and not everyone has died yet? Each of these involves censored observations. The question about what to do with these kinds of observations is why Cox regression was created; a non-paramedic partial likelihood technique, which he called proportional hazards. It deals with censored observations, which are those patients that have an unknown time until event status. This unknown time until event can be caused by either not having the event at the time of the analysis or losing contact with the patient.
What about those subjects that died from another cause and not the cause the test drug was treating? Are there other variables (covariates) that influence (increase or decrease) the time until the event? These questions involve extensions of the general survival model. The first is about competing risks and the second is about regression involving independent variables.
For a sample business case explaining survival analysis and its applications in more detail read Marketing Analytics: A Practical Guide to Real Marketing Science.
About the author: Mike Grigsby has been involved in marketing science for over 25 years. He was marketing research director at Millward Brown and has held leadership positions at Hewlett-Packard and the Gap. With a wealth of practitioner experience at the forefront of marketing science and data analytics, he now heads up the strategic retail analysis practice at Targetbase. Mike is also known for academic work, having written articles for academic and trade journals and taught at both the graduate and undergraduate levels. He is a regular speaker at trade conventions and seminars.
Save 20% when you buy Marketing Analytics with code MKTMAB