Hiring Company
Age Wave IMPACT
Background
As the capitation amount for Medicare HMOs began to fall behind the actual average costs HMOs were incurring they began to look for ways to save money in their marketing. One way was to use a predictive model of buyers to limit mailing. The hope was that an HMO could mail to 50% of the prospects but still get 80% of the members it would if it mailed everyone.
Some HMOs were also interested in delivering different messages to different segments of customers.
Others wanted the ad agency to develop attrition models to guide pre-attrition communications and offers.
Description of Work
Advance Analysis developed predictive buying models for Medicare HMOs in Maine, New York, Pennsylvania, New Jersey, Maryland, Georgia, Wisconsin, California, and elsewhere…a total of 14 state BCBS companies. Cluster analysis was developed for two companies to allow more customized creative approaches.
We were lucky that we were able to get access to the results from the modeled direct mailing in most cases. We found that some models did well while others did not. New Jersey’s model was the first, developed by an outside vendor of the ad agency, and did not do well because of data problems. Wisconsin’s was the last and their market was changing too fast for the model to be at all valid. More stable companies such as BCBS Georgia benefited significantly from the models and saw response rates and joining rates jump substantially.
In addition to learning the lesson that it doesn’t make a lot of sense to mail in one market situation with a model developed from lists mailed in another, it became clear that certain variables were consistently highly predictive of behavior while other expected ones were not at all. For example, in predicting who would join Medicare Supplemental plans the number of people in the household was often the strongest predictor: aging parents who had support from adult children (lived in larger households), it appeared, were much less likely to buy a Supplemental plan, once income was held constant.
The other valuable lessons that came from this experience were: (a) how important it is to understand all of the data elements fully, (b) now important it was to factor in the number of mailings each recipient got, and (c) how valuable breaking down the process into logical steps could be. In this case it was very useful to develop separate models of "responders" and "joiners". A model representing the difference between those who joined and those who didn’t even respond was often minimally predictive:
- For example, one BCBS plan had a premium brand connoting trust and quality, but their Medicare Supplemental product was actually the least expensive and carried the fewest benefits in the market. So, the responders were on average high income and highly educated, but most didn’t join because when they learned about the product they realized that they could afford something better.
- The joiners were the lower income subset of the responders, resembling the full list pretty well and thus very difficult to identify through modeling.
Although the two models weren’t very useful for improving acquisition, understanding the responding and the joining processes separately was incredibly valuable as a diagnostic tool and led to a change in how marketing was conducted at every level.
