Predicting Non-Adherence

There is lots of buzz over yesterday’s article in the WSJ about Express Scripts being able to predict who will be adherent.  Today’s blog post on the Corporate Research Blog added some details (or further confused me).  It says that the model is 80% accurate in predicting the 10% of people who are least likely to be adherent.

Is that all it does?  For sake of this post, let’s assume it does.  That seems much less interesting and much easier to do.  In talking with a leading researcher in this area that has looked at the correlation of 9,000 variables to adherence, he told me that nothing was highly correlated, but the most correlated metric was past behavior.  Where they adherent in the past on other medications?  Did they take preventative action (e.g., get flu shots, mammograms)? 

Several people have been looking at how credit scores can be used to predict adherence.  Given errors in credit scores, this may be deceptive even if it works.

But, back to the issue.  If you know who’s least likely to be adherent, so what?  Do you give up on these people?  They aren’t likely to chance behavior.  Do you try harder or have a different strategy with these people?  If you succeed and move them to taking their medications 40% of the time (using a proxy like a 40% medication possession ratio), does it make a difference? 

I would think it’s better to focus on the people who are likely to be adherent and how to enable them to move from 40-70% MPR to >80% MPR.  We often work with clients to stratify their population and have different intervention strategies (channel, messaging, level of effort, etc.) across where they fit in the model (value, likelihood to engage, likelihood to change).