Aging
Developing A Claims-Based Algorithm for Estimating State- and Local-level Dementia Prevalence Ruijia Chen* Ruijia Chen Chen Chen Chen Boston University School of Public Health
Background: Regional dementia prevention policy planning and monitoring requires reliable area-level estimates of dementia prevalence, which are currently lacking. The best sources of individual dementia status come from panel survey studies with systematic neuropsychological assessments, but these data are sparse at local area levels. Larger population-level datasets lack these measurements. Thus, using 2006–2020 Health and Retirement Study (HRS)–Medicare administrative-Medicare Current Beneficiary Survey (MCBS) linked data, this study developed and validated an individual-level dementia prediction algorithm.
Methods: Dementia status was defined using the Gianattasio-Power algorithm, an HRS-based approach validated against detailed neuropsychological assessments. Candidate algorithms included information on age, sex, race/ethnicity, dual Medicare/Medicaid eligibility, state of residence, Alzheimer’s disease diagnosis, diabetes, hypertension, stroke, chronic obstructive pulmonary disease, obesity, and hearing impairment. The dataset was randomly split into training/testing (70/30%) samples. We fitted pooled logistic regression and random forest models using predictors measured at time t to predict dementia at time t + 1, with inverse probability attrition weights. Model performance was then evaluated using the testing sample.
Results: Among 3,643 beneficiaries with dementia status information, the mean baseline age was 81 years (SD=7). Both models demonstrated good discrimination (AUC ~0.78), with high specificity (79–82%) and moderate sensitivity (56%) (Figure). The discrimination slope for the random forest model was 0.84 and calibration-in-the-large=-0.005.
Conclusion: A simple prediction model using available Medicare administrative data performs well with respect to a well-validated dementia algorithm using experimentally ascertained survey data. These individual-level assessments could improve area-based estimates to inform future resource allocation decisions.
