demarcated, homogenous cells. Without enough observations to use cell experience for setting probabilities, we need to say something about how coarse cells compromise the goal that we want and that could be achieved with a larger data set.
Most of what has preceded in this discussion works under three assumptions: (1) there is sufficient information on family structure and income and on the health and details of insurance coverage of each of the members of the family; (2) a health shock is not enough in its own right to change the nature of one’s coverage—in other words, one cannot be eligible for some insurance policy or public program by the nature of the health event (counterexamples include blindness, pregnancy if low income, and renal failure and dialysis); and (3) there are enough sample cases of families with the same family structure, income, coverage, and health status that one can identify equivalent or very nearly equivalent families to form a risk cell sufficient to use the observed distribution as a source of the risk cell specific distribution.
As a practical matter, the assumptions regarding data are not likely to hold with current large-scale data collection efforts (see Chapter 5). Most data sources do not include detailed information on eligibility and coverage provisions for either public or private insurance. Most data sets that would have sufficient information for common risk adjustment methods are small relative to the number of combinations of age, family structure, and health of families that would be needed.
Another concern is that most of the common risk adjusters are statements about expected amounts of spending, not the distribution of that spending, based on small or modest sample sizes, using multiple waves of the data. Thus, any risk classification system needs to be able to handle coarse risk cells for individuals and to find a method for combining data on individuals with varying risks internally within families. This may require the health equivalent of the family composition algorithm used in the calculations of the SPM thresholds. Then the question becomes: Is the family-equivalent health risk a weighted average of the individual risks that reflect measures of how well baseline health predicts subsequent expenditures? How well does such a measure forecast both family means and variances? If one were to address only single-person families, how well does current expected experience also affect the variability in that number?
ISSUES CONSIDERED BUT NOT ADDRESSED IN THE MCER PROPOSED APPROACH
Four issues that the panel considered but decided not to address in its proposed approach for developing an MCER measure are summarized below.