Close Close

Predictive Models: Math You Can Use

Your article was successfully shared with the contacts you provided.


For many employers and benefits brokers, predictive modeling seems to be a jumble of confusing math.

But learning a little about how predictive modeling works may provide inspiration for new ways to use these tools, which improve on traditional forecasting techniques that rely entirely on demographic information and financial factors.

Modern predictive risk models differ from the old models in 2 important ways.

First, todays predictive risk models make heavy use of clinical information drawn from health care claims. The expanded breadth of detail makes a difference in the quality of a forecast: key utilization measures are diagnosis, procedure and pharmacy codes as well as dates and place of service, and basic patient demographic information.

Second, the new predictive risk models attempt to identify future high-cost individuals, not simply count cases that already have generated huge bills. This represents an important departure from experience-rating mechanisms that simply adjust prior period claims to reflect anticipated cost and utilization trends.

The basic strategy for applying a predictive risk model is easy to understand: You classify plan members by health status according to clinical indicators, then you use assign factors called risk weights that reflect each classifications anticipated resource utilization. Although anticipated resource utilization often refers to cost, it might also refer to inpatient days, office visits, surgeries or prescriptions, for example.

A model might, for example, handle weighting by assigning employees with uncontrolled diabetes a high number of debits and employees in perfect health a low number of debits. The average number of debits is calculated across the entire group and then compared to the expected average for all groups. In practice, the schedule of debits should also reflect the fact the predictive risk model cannot account for all of the anticipated variability in cost.

Consultants might develop separate risk weights to come up with concurrent estimates for what will happen during the current plan year and prospective estimates for what will happen in future years.

An effective predictive risk model can help you forecast claim costs for an entire plan, such as a small group looking to renew its coverage. It also can help plan managers identify individuals who need special programs to keep conditions such as uncontrolled diabetes from leading to catastrophic claims.

Keep in mind that some models work better for some purposes than for others. The selection of an appropriate model depends both on the particular line of business being assessed and on the target user of the models output.

A model might produce accurate claim cost projections given inputs of detailed health care claims data; however, a broker helping a small group to change carriers would not typically have access to such detailed data. Instead, the only data available might be employees answers to health questionnaires, which can still be very helpful in identifying potential trouble points. Of course, detailed data would be available to the carrier and useful in renewal of its small group cases.

There are some limits on the models predictions. More accurately, these may actually qualify as elements to adjust to and be aware of, not necessarily static limitations. One issue is that models may not include variables for hard-to-predict catastrophes, such as terrible accidents or the births of premature babies.

Another problem is that consultants may need risk weights tailored to fit a particular employer to get useful results. Demographic factors, regional variations in prices and treatment practices, and negotiated provider payment rates could throw off the off-the-shelf risk weights.

Finally, never underestimate the impact of lag time: Many medical claims take months to creep in. It may be obvious but it bears emphasizing: any individuals identified as high-risk may have already incurred significant claims costs before anything can be done to intervene. Newer models incorporate drug claims data because drug claims come in so much faster than medical claims.

, FSA, MAAA, is chief actuary at Apex Management Group, Princeton, N.J., a division of Gallagher Benefits Services Inc. Apex is a health care, insurance services and actuarial consultancy.

Reproduced from National Underwriter Edition, April 15, 2005. Copyright 2005 by The National Underwriter Company in the serial publication. All rights reserved.Copyright in this article as an independent work may be held by the author.