¾ÅÉ«ÉçÇø

In recent years, the increased use of "big data" and statistical techniques to score potential transactions has transformed the operation of insurance and credit markets. In this paper, we observe that these widely-used scores are statistical objects that constitute a one-dimensional summary of a potentially much richer heterogeneity, some of which may be endogenous to the specific context in which they are applied. We demonstrate this point empirically using rich data from the Medicare Part D prescription drug insurance program. We show that the "risk scores", which are designed to predict an individual's drug spending and are used by Medicare to customize reimbursement rates to private insurers, do not distinguish between two different sources of spending: underlying health, and responsiveness of drug spending to the insurance contract. Naturally, however, these two determinants of spending have very different implications when trying to predict counterfactual spending under alternative contracts. As a result, we illustrate that once we enrich the theoretical framework to allow individuals to have heterogeneous behavioral responses to the contract, strategic incentives for cream skimming still exist, even in the presence of "perfect" risk scoring under a given contract.