How well we can predict next year's health care penditures, Y, at the individual level depends both on the distribution of Y, and the X variablesavailable for prediction. Although Y is often modeled as lognormal, its distribution is ly "worse" than that. In a largely healthy population, the mode is at 0, with the probability density function sically decreasing as Y increases. The coefficient of ariation (CV=SD/mean) is likely to be as high as 6. If the predictors include markers for serious chronic problems, such as major organ failure, these are strongly predictive but very orly "conditioned." Thus, for example, a variable that marks a problem with 1% prevalence has a CV of 10. xtremely large populations are needed to "stabilize" cost redictions in such settings, and fancy modeling may do more harm than good. We will illustrate the problems with real data and offer thoughts about solutions.
Learning Objectives: N/A
Keywords: Cost Issues, Health Care
Presenting author's disclosure statement:
Organization/institution whose products or services will be discussed: None
I do not have any significant financial interest/arrangement or affiliation with any organization/institution whose products or services are being discussed in this session.