To compare the sensitivity, specificity, and predictive value of published versus sample-based norms to detect early dementia in the Uniform Data Set (UDS).

The UDS was administered to 526 nondemented participants from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Baseline scores were standardized using published norms and healthy control data from ADNI corrected for age, education, and sex. Subjects obtaining two scores < −1 SD (determined separately using published and sample norms) were labeled “at risk for dementia.” Both methods were compared on sensitivity, specificity, and positive/negative predictive value (PPV/NPV) for dementia at follow-up.

Risk scores derived from published data had 86.1% sensitivity, 62.0% specificity, 68.6% accuracy, 46.1% PPV, and 92.2% NPV. Those from sample norms were more sensitive (91.0%), less specific (52.9%), and less accurate (63.3%), with worse PPV (42.1%) and similar NPV (94.0%). Sample norms were better at identifying incident dementia cases with relatively lower education than those with higher education. Discrepancies between both methods were more common in women.

Sample norms are marginally more sensitive than published norms for predicting dementia, while published norms are slightly more accurate. Accuracy of risk estimates for women and those with lower education may be increased using locally generated norms.