AbstractINTRODUCTIONRecent genome‐wide association studies identified new dementia‐associated variants. We assessed the performance of updated polygenic risk scores (PRSs) using these variants in an independent cohort.METHODSWe used Cox models and area under the curve (AUC) to validate new PRSs (PRS‐83SNP, PRS‐SBayesR, and PRS‐CS) compared with an older PRS‐23SNP in 12,031 initially‐healthy participants ≥70 years of age. Dementia was rigorously adjudicated according to Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM‐IV) criteria.RESULTSPRS‐83SNP, PRS‐SBayesR, and PRS‐CS were associated with incident dementia, with fully adjusted (including apolipoprotein E [APOE] ε4) hazard ratios per standard deviation (SD) of 1.35 (1.23–1.47), 1.37 (1.25–1.50), and 1.42 (1.30–1.56), respectively. The AUC of a model containing conventional/non‐genetic factors and APOE was 74.7%. This was improved to 75.7% (p = 0.007), 76% (p = 0.004), and 76.1% (p = 0.003) with addition of PRS‐83SNP, PRS‐SBayesR, and PRS‐CS, respectively. The PRS‐23SNP did not improve AUC (74.7%, p = 0.95).CONCLUSIONNew PRSs for dementia significantly improve risk‐prediction performance, but still account for less risk than APOE genotype overall.