Pharmacogenetic dosing algorithms help predict warfarin maintenance doses, but their predictive performance differs in different populations, possibly due to unsuspected population-specific genetic variants. The objectives of this study were to quantify the effect of the VKORC1 D36Y variant (a marker of warfarin resistance previously described in 4% of Ashkenazi Jews) on warfarin maintenance doses and to examine how this variant affects the performance of the International Warfarin Pharmacogenetic Consortium (IWPC) dose prediction model. In 210 Israeli patients on chronic warfarin therapy recruited at a tertiary care centre, we applied the IWPC model and then added D36Y genotype as covariate to the model (IWPC+D36Y) and compared predicted with actual doses. Median weekly warfarin dose was 35 mg (interquartile range [IQR], 24.5 to 52.5 mg). Among 16 heterozygous D36Y carriers (minor allele frequency = 3.8%), warfarin weekly dose was increased by a median of 43.7 mg (IQR, 40.5 to 47.2 mg) compared to non-carriers after adjustment for all IWPC parameters, a greater than two-fold dose increase. The IWPC model performed suboptimally (coefficient of determination R²=27.0%; mean absolute error (MAE), 14.4 ± 16.2 mg/week). Accounting for D36Y genotype using the IWPC+D36Y model resulted in a significantly better model performance (R²=47.2%, MAE=12.6 ± 12.4 mg/week). In conclusion, even at low frequencies, variants with a strong impact on warfarin dose may greatly decrease the performance of a commonly used dose prediction model. Unexpected discrepancies of the performance of universal prediction models in subpopulations should prompt searching for unsuspected confounders, including rare genetic variants.