reliable comparisons. In many circumstances, current clinical data systems and risk adjustment strategies are technically incapable of meeting those reasonable expectations.
There are many potential sources of variation in measured health outcomes (see Box 8-1), and risk adjustment methods can account only for differences in known patient factors. For example, Eddy estimates that all major factors proven to explain infant mortality rates (race, maternal alcohol consumption, maternal tobacco smoke exposure, altitude, and differences in prenatal care delivery performance) account for only about 25 percent of documented variation in patient outcomes (Eddy, 2002). The other 75 percent of outcomes are beyond the reach of risk adjustment strategies.
Geographic aggregation (e.g., variation in hospital programs and local referral patterns) can also play a defining role. A recent evaluation of hospital quality outcome measures found that most produced statistically reliable results when aggregated to the level of a metropolitan area, state, or multistate region, but only a few measures produced valid results at the level of individual hospitals (Bernard et al., 2003). The fact that an outcome mea-