The literature suggests four basic "facts" about biased selection in PGPs: First, the well disproportionally enroll in PGPs. Second, for a group defined according to low expenditure, expenditure regresses toward the mean. Whether this is applicable to PGPs is unclear. Third, the well disproportionally disenroll from PGPs. Finally, although more evidence is needed, there appears to be little or no bias in PGP enrollee populations. This paper presents a framework that unifies these patterns. Based on these findings, this paper draws several conclusions that are relevant to Medicare capitation payments to HMOs; First, attempts to increase the sophistication of the AAPCC to compensate for biased selection should focus on predictors of the permanent component of expenditures instead of using past expenditures as a predictor. One can explain no more than 20 percent of the variance in such equations. Second, because of regression toward the mean, there are several interpretations of biased-enrollment estimates based on preenrollment data. The issue of these interpretations is unlikely to be resolved without a randomized experiment, although research with nonexperimental data could be useful. Third, due to the probable decay in PGPs' preferential selection, for rollovers there may be little biased selection, and for switchers the length of enrollment might be included in the AAPCC. Fourth, the adequacy of an AAPCC depends, in part, on whether biased selection results from consumer behavior or HMO behavior. Each of these conclusions suggests a research topic that should have priority: first, finding predictors of the permanent component of expenditure, thus increasing the percentage of the variance explained; second, testing the socioeconomic characteristics model of biased enrollment; third, measuring the speed at which the preferential selection of PGPs declines; fourth, investigating HMOs' ability to "select" enrollees.