What Medicare Claims Catch & What They Miss

July 2026

Our new paper in Alzheimer’s & Dementia puts the Dementia DataHub case definitions to a hard test. The findings provide important context for interpreting the numbers on the Dementia DataHub.

The questions. When a Medicare claim indicates dementia, is the person actually living with dementia? Alternatively, can we trust that a person with no evidence of dementia on their Medicare claims actually has no dementia? Does the accuracy of Medicare claims vary by place or by patient, and if so, what drives that variation?

The design. Our team compared DDH’s claims-based dementia classification against survey-based dementia status in the 2018 Health and Retirement Study (HRS) using linked HRS-Medicare data. Dementia in HRS is classified using a logistic algorithm that draws on demographics, cognitive status, and physical health to classify dementia probabilistically. While not a clinical gold standard itself, the HRS algorithm was developed and validated against gold standard dementia diagnosis and is available for a nationally representative sample.

What we found. The DDH "Highly Likely and Likely Dementia" definition has an overall sensitivity of 50 percent and a specificity of 97 percent, for an overall accuracy rate of 91 percent. In other words, Medicare claims are only identifying half of the individuals classified as having HRS-based dementia. The broader "Any Type of Dementia" definition identifies a higher proportion of cases (59 percent sensitivity), but at the cost of more false positives.

Missed dementia cases are not random. For example, sensitivity ranged from 36 percent in the Pacific division to 60 percent in the East South Central, with urban and rural settings often diverging within divisions. Older and more impaired individuals (both physically and cognitively) were more likely to have a dementia diagnosis, indicating lower risk of under-diagnosis (false negatives) but higher risk of over-diagnosis (false positives). Women were more likely classified as having dementia in HRS, but were less likely to have it documented in their claims compared to men. Hispanic respondents had elevated false-positive rates relative to non-Hispanic whites, but contrary to prior evidence, there were no race/ethnicity differences in missed diagnosis rates among those with HRS dementia. Higher education also predicted better accuracy in both directions.

Why it matters. The DataHub is a map of diagnosed dementia rather than true underlying dementia. Differences in diagnosed prevalence across states are driven by variations in both true dementia and diagnostic accuracy. Raw comparisons of Medicare diagnosis prevalence are misleading without calibration, given the low and geographically variable sensitivity of claims data.

Findings from this paper are currently being used to inform formal small-area estimation that adjusts diagnosed prevalence toward true prevalence. Findings will be reported here on the Dementia DataHub as this work progresses. 

Read the paper: Performance of ICD-10 code-based dementia case definition in the Health and Retirement Study

Funded by the National Institute on Aging (R01AG075730). The views expressed are our own.