Using self-reported data to predict expenditures for the health care of older people

OBJECTIVES: To create and test a method for using self-reported data to predict future expenditures for the health care of older people. DESIGN: A two-stage regression model of the relationship between self-reported data and Medicare expenditures during the following year was constructed from a randomly selected (derivation) half of a cohort of fee-for-service Medicare beneficiaries. For the other (validation) half of the cohort, two sets of predictions of 12-month Medicare expenditures were generated, one using the new two-stage model and the other using the principal inpatient diagnostic cost group (PIP-DCG) method now used to risk-adjust capitation payments to Medicare + Choice health plans. Both sets of predictions were compared with Medicare's actual 12-month expenditures for the validation cohort. SETTING: Ramsey County, Minnesota. PARTICIPANTS: Community-dwelling Medicare beneficiaries aged 70 and older (N = 13,682) who responded to a mailed survey. MEASUREMENTS: Predicted-to-observed ratio (PTOR) of Medicare expenditures. RESULTS: For the validation cohort, Medicare's actual 12-month expenditures totaled $26.5 million. The two-stage model predicted Medicare expenditures of $26.4 million (PTOR = 1.00); the PIP-DCG method predicted $31.2 million (PTOR = 1.18). Within subpopulations of healthy and ill beneficiaries, the two-stage model's predictions remained considerably more accurate than the PIPDCG predictions. CONCLUSION: Self-reported data may predict future Medicare expenditures more accurately than administrative data about beneficiaries' demographic characteristics, and previous hospitalizations.