As more Americans choose among insurance plans, the possibility of biased selection increases in importance. Although regression toward the mean is recognized as a common problem in evaluating social programs, it has generally been ignored in studies of biased selection. Suppose that people are included in a group simply because they had expenditures in one year $100 below the mean; that is, health status or other risk factors are not part of the selection criteria. Empirically, the expected difference in the following year is about $20 and appears to fall in each subsequent year. This pattern holds for the elderly and nonelderly. Evidence of lower pre-enrollment expenditure of prepaid group practice (PGP) enrollees can be interpreted in several ways. Under one interpretation, PGP enrollees are assumed to be a random sample conditional on pre-enrollment expenditure, such that biased selection is one fifth of estimates based on 1 year of data and one half of estimates based on 4 years of data. This article cannot resolve the issue of alternative interpretations; it only raises it.