We find that incentives to a plan to devote resources to services depend on the demand for that service among the plan’s current enrollees, how well potential enrollees can forecast their demand for the service, whether the distribution of those forecasts is uniform or skewed in the population, the correlation of those forecasts with forecasts of other health care use, and on the risk-adjustment system used to pay for enrollees. We show how all these factors fit together into an index for each service the plan provides in detail.

Many papers have shown that consumers choose health plans on the basis of their anticipated spending. Medicare’s program for paying HMOs by capitation has been studied repeatedly in this regard. In a representative analysis, Hill and Brown (1990) find that individuals choosing to join HMOs for the first time were spending 23% less than those who do not choose to join and had a lower mortality rate in the period after joining. (See also Eggers and Prihoda, 1982; Garfinkel et al., 1986; Brown, Bergeron and Clement, 1993). The finding of significant adverse selection in Medicare continues to be borne out by more recent studies (Medicare Payment Advisory Commission, 1998). Numerous other studies have also found among other populations that those choosing to join HMOs are “healthier” in some ways than those not joining (Cutler and Reber, 1997; Cutler and Zeckhauser, 1997; Glied et al., 1997; Robinson, Gardner and Luft, 1993; Luft and Miller, 1988.

Risk-adjustment of payments to managed care plans are intended to counteract incentives to distort services. If plans are paid more for enrollees likely to be costly, the plan will not shun these enrollees. Individuals choose plans on the basis of what they (the individuals) can predict. A risk adjustment system that picks up the predictable part of the variance in health costs is thus able to address dangers of selection. How much of the health care cost variance individuals can anticipate is not known.

To get some idea, empirical researchers have assumed that individuals know the information contained in certain potential explanatory variables, and then investigated how much of the variance is explained by these covariates. In the most well-known of these studies, Newhouse et al. (1989) assume that individuals know the information contained in their individual time invariant contribution to the variance – in other words, they can predict the 24 percent of the variance picked up by the person fixed effects in a multiyear regression.

They regarded this as a reasonable “minimum” of what individuals could predict. Currently available risk adjusters miss a good deal of this predictable variance. Medicare’s current risk adjusters explain about 2 percent of total variance; proposed refinements improve the explanatory power considerably, but only to about 9 percent (Ellis et al., forthcoming; Weiner et al., 1996). There remains considerable room for systematic selection that would not be captured by a payment system based on existing risk adjusters.