It has become clear after speaking with numerous health insurance carriers, both in the United States and beyond, as well as at conferences (such as NHCAA), that there is a mass movement towards the nirvana that is "predictive modeling."
Now that our industry is realizing the importance of predictive modeling – it is important to have a direct dialog about what value it brings to aberrant behavior detection in the health care industry. To start, let it be clear that predictive modeling provides the potential for detecting complex behavior different from other mathematical methods. With that said, predicative modeling is not the "silver bullet" that people seem to think it is, and there is a great deal of confusion as to what it is in the first place.
A good starting point is too distinguished between descriptive and predictive analytics. Descriptive analytics looks at current/historical state to tell you what happened or what is happening currently. Examples of this are cluster analysis, correlations, multivariate, etc. All of which are very useful is telling us what is happening – but not predicting future behavior.
When people discuss predictive analysis, the first thing out of their mouths is typically, REGRESSION, clearly the all-time favorite…. But sadly, not always the best option. In many cases, decision trees (sometimes called classification trees) are better for health care, because they handle missing data well (and health care data always seems to have gaps), and they are very easy to example.
Imagine going to court and trying to explain a ‘neural network’ to a group of lawyers (who traditionally are not statistically inclined – though, I’m sure there are exceptions), to justify why a doctor is being investigated. I do not envy anyone that experience.
However, one of the most significant issues with predictive analytics is that they require historical reference to build – simply stated, if you’ve never seen it before you can’t build it. What this means is that all new fraud will slip through the cracks, until enough events have occurred and been identified to build a predictive model.
Likewise, there seems to be a great deal of hype and misconceptions around predictive models. True, predictive methods are very useful, but only as part of larger analytics approach. Transactional rules are great for identifying simple up-coding and combine very nicely with predictive analytics. Anomaly detection (or outlier analysis) is wonderful for spying new trends. Social network analysis (or link analysis) identifies collusive behavior. But all of these methods work best collectively. Largely to identify patterns that best suited to each method, but also to work collectively to reduce false positives.
This list of options always begs the question, “is there one best solution to the problem?” I would suggest that you want to use all the tools in the box, and pick the best analytical tool for your particular problem. Start by learning about the different types of analytics today.