2016 Analytics 50 Honorees

Independence Blue Cross Analytics 50 Submission

Ashley Masterson, Manager, Medical Cost and Business Analytics
Independence Blue Cross
Philadelphia, PA
Industry: Healthcare

Business Challenge

Independence Blue Cross experienced the challenge of individually targeting Medicare Advantage (MA) marketing materials in an evolving and increasingly competitive healthcare landscape. MA products are very stratified, and individuals select products for a range of reasons. This means that to market MA plans effectively, a successful campaign must contain an appropriate product and tailored message for each individual. Given budget and development constraints, how could Independence segment the prospect list so that each individual received the most relevant material?

Analytics Solution

A multistep process centered around the k-means clustering algorithm with product prediction inputs was used to estimate an individual’s product choice and content preferences. The data set consisted primarily of third-party prospect data purchased by Independence. While some of these records could be linked to existing Independence members, the majority could not. The predictive model was trained using the Independence members’ product selection information. The models were deployed to output the probability that a prospect would select a given plan. Then k-means clustering was used to segment the prospect list.

Impact

Overall, the MA annual enrollment period was a success. Independence’s net MA membership increased by over 7,000 new members. Given confounding factors like new product offerings, benefit changes and rate adjustments, campaign success cannot be solely attributed to the segmentation efforts. However, since the inception of the use of segmentation, a reduction in cost to acquire as well as member churn has been realized. For model improvement, analyses to measure accuracy of predicted plan selection within each cluster were performed. The predicted choice was compared to actual plan selection. For five of the six clusters, the model’s outputs were highly predictive of plan choice. This year modifications were made to the cluster for which actual plan enrollment did not reflect the model predictions.