Abstract Background Many Alzheimer’s disease (AD) genetic association studies disregard age or incorrectly account for it, hampering variant discovery. Methods Using simulated data, we compared the statistical power of several models: logistic regression on AD diagnosis adjusted and not adjusted for age; linear regression on a score integrating case-control status and age; and multivariate Cox regression on age-at-onset. We applied these models to real exome-wide data of 11,127 sequenced individuals (54% cases) and replicated suggestive associations in 21,631 genotype-imputed individuals (51% cases). Results Modeling variable AD risk across age results in 5–10% statistical power gain compared to logistic regression without age adjustment, while incorrect age adjustment leads to critical power loss. Applying our novel AD-age score and/or Cox regression, we discovered and replicated novel variants associated with AD on KIF21B, USH2A, RAB10, RIN3, and TAOK2 genes. Conclusion Our AD-age score provides a simple means for statistical power gain and is recommended for future AD studies.