Development and validation of a two-step shared parameter model for dementia imputation in the Cardiovascular Health Study Cohort.

Abstract

BACKGROUND Large-scale dementia ascertainment for research remains challenging. We demonstrate a method to impute dementia status and onset time using data from the Cardiovascular Health Study (CHS). METHODS We used a linear mixed effects model to estimate individual cognitive trajectories and included the estimates as covariates in an accelerated failure time model used to impute time to incident dementia in the CHS Cognition Study, a sub-study with dementia ascertainment. We calibrated the model in a 60% random sample (n = 2,000) of eligible participants with CHS Cognition Study dementia classifications and validated the final model in the remaining 40% sample (n = 1,334). We then imputed dementia onset time and dementia status during follow-up in the full CHS sample, including those without (n = 1,415) CHS Cognition Study dementia classifications. RESULTS In the validation sample, relative to the CHS Cognition Study dementia classifications used as the "reference standard", specificity (98.5%), positive predictive value (81.9%), negative predictive value (92.0%) and accuracy (91.3%) were high, while sensitivity was modest (43.8%), with mean imputed onset time +/-1.5 years of classified. Performance varied by participant characteristics. Ultimately, 227 (16.0%) of participants without CHS Cognition Study classifications, and 472 (9.9%) of all CHS participants were classified as having dementia according to this approach. CONCLUSION The shared parameter approach can be implemented in samples with existing cognitive data and a validation sample with reference-standard dementia adjudication. We found high overall accuracy and higher specificity than sensitivity, similar to reported performance metrics for algorithmic approaches requiring linkage to administrative data.