Misclassification and selection bias when identifying Alzheimer's disease solely from Medicare claims records

BACKGROUND: Medicare claims as the basis for health condition adjustments i s becoming a method of choice in capitation reimbursement. A recent study h as found that claims-based beneficiary classification for Alzheimer's disea se produces lower prevalence estimates and higher average costs than previo us healthcare cost studies in this population. These sets of studies differ in data sources, period length, and in their specification of dementia. OBJECTIVES: Participants in the Medicare Alzheimer's Disease Demonstration (MADDE) provide a sample of persons known to have some form of dementia. Th is group is used to test the adequacy of claims data for identifying eligib le cases and any bias in expenditure differences between those flagged or n ot flagged by a claim in a given period. DESIGN: A prospective cohort design using up to 36 months of claims data. SETTING: The demonstration enrolled 4166 participants in treatment, and 394 2 in a control group in eight communities across the US. Cases were combine d in this analysis. PARTICIPANTS: Persons with available Medicare Part A Sr B claims data: thos e receiving care under fee for service reimbursement were used in the analy sis. A total of 5379 MADDE cases received fee for service care during 1991 and 1992, the period of primary interest in the analysis. MEASUREMENT: Client health and functional status inter views and Medicare P art A Sr: B claims. RESULTS: Less than 20% of MADDE participants were classified with Dementia of the Alzheimer type (DAT) from a single year of claims although 68% had a DAT diagnosis from a referring physician. Annualized expenditures were 1.7 times higher among those with DAT from claims compared with those known ot herwise to have dementia but who had not been identified with this conditio n from Medicare claims. CONCLUSION: Underclassification of dementia from claims records can be part ially remedied by increasing the period during which claims are compiled, b ut additional diagnostic sources will likely be needed to increase prevalen ce counts closer to 100% of true cases. Risk adjustment based on a single y ear of reported claims expenditures may overpay providers, at least in the short term, because payment incentives will likely increase prevalence repo rting.