Announcement: SOA releases March 2022 Exam P passing candidate numbers.

Additional Predictive Modeling Sessions

Additional Predictive Modeling Sessions

By Vincent Kane


As Council Member of the Education & Research (E&R) Section, I coordinated two sessions on the topic of predictive modeling (PM) in 2008. The first was jointly sponsored by the Health and E&R Section Councils and was presented at the SOA Spring Health Meeting in Los Angeles on May 29th, 2008. The second upcoming session is sponsored by the E&R Section Council and is planned for Oct. 21st, 2008 at the SOA Annual Meeting in Orlando.

Spring Health Meeting—Session 45–Predictive Models: Paralyzing Puzzle or Process Panacea?

Predictive models have been implemented across health care organizations to measure provider efficiency, develop risk–adjusted payments, project costs for pricing and underwriting, and to profile high cost cases for medical management. However, different applications require different types of predictive models that were developed to answer specific questions of interest. Models vary according to underlying theory, the calibration data sample, the predicted outcome, data inputs feeding the model, and the method of estimation. Such aspects must be considered before a model is ever implemented so that its results are accepted by stakeholders. Correct model selection leads to results which facilitate, rather than paralyze, decision–making.

At the Spring Health Meeting predictive modeling session, three speakers shared their insights into uses and applications of predictive models. First, Bill Glasheen, research scientist at DxCG, presented on classic uses of predictive models and demonstrated that there is no turnkey solution for all applications. Bill discussed the main modeling considerations for several illustrative examples, from risk–adjusted capitation to economic profiling of physicians. Next up was Steve Griffiths, of the Ingenix Medical Informatics team, who presented on predictive modeling applications in medical management. He covered the use of PM for identification and stratification of the "right people" amenable to intervention, in addition to the use of propensity scoring techniques for program evaluation. Last, but certainly not least, was the E&R Section Council's own Margie Rosenberg, professor at the University of Wisconsin, Madison. Margie presented a sophisticated two–part Bayesian model for use in predicting health care spending and hospitalizations for a small longitudinal sample of Cystic Fibrosis patients. Her paper is also published in the January 2008 issue of the North American Actuarial Journal (NAAJ), entitled "Predictive Modeling of Costs for a Chronic Disease with Acute High Cost Episodes."

The session attracted over 100 attendees who sought answers to the title question of this session, and hopefully left less "paralyzed" and undoubtedly more informed regarding PM applications in health care practice.

Upcoming SOA Annual Meeting—Session 60—Predictive Models: Innovation versus Implementation

The predictive modeling session planned for the Annual Meeting will cover both life and health applications to encourage attendance from both practice areas. The focus will be on new innovative techniques in predictive modeling, and whether the more sophisticated models have experienced wide acceptance for commercial applications. These models increasingly use non–claims based sources of data as predictive inputs. Examples of new data sources include lifestyle–based data, consumer and credit data, laboratory values, health risk appraisal data, and other inputs from electronic medical records or transactional detail.

The evolution and innovation of these models has outpaced their diffusion and adoption due to a variety of factors. These include availability of high quality data, regulatory and privacy concerns, strategic buy–in, and barriers in model implementation, monitoring, and updating. More importantly, the additional predictive power to be gained using non–claims based data is the subject of heated debate.

Three speakers are planned for what promises to be a "standing room only" crowd. Jason Xue, consultant at Watson Wyatt, has applied Generalized Linear Models (GLM) to analyze annuity mortality experience for client insurance companies. This modeling approach promises to uncover details that traditional methods have overlooked, and may be used for development of mortality tables specifically tuned to a company's own book of business. Next, Chris Stehno, senior manager at Deloitte Consulting, will present on what it takes to be successful in model implementation for both life and health applications, and how to gain organizational acceptance of the newest, most innovative modeling techniques. The last speaker will be Ognian Asparouhov, chief scientist at MEDai, who will address whether the new predictive models are really better or if they are too sophisticated for practical use compared to more simple approaches. Asparouhov will also discuss whether there is a ceiling on predictive performance and whether one can quantify the return for an additional point of R–squared.