Abstract Background: The PAM50, a gene expression assay to categorize breast tumors into intrinsic subtypes, has not been previously used to examine short- and long-term prognostication in a population-based cohort where treatment patterns and time of initial follow-up vary. Methods: In a stratified case–cohort design of 1,691 women from the LACE and Pathways breast cancer survivor cohorts, we used PAM50 to categorize tumors into Luminal A (LumA), Luminal B (LumB), HER2-enriched (HER2-E), Basal-like and Normal-like, and to examine risk of early and late recurrence and mortality by Cox proportional hazards regression. Results: Compared with LumA, cumulative risk of recurrence and breast cancer death was higher for LumB, HER2-E, and Basal-like tumors at 2, 5, and 10 years. However, HR of breast cancer death varied over time [<5 years (early) vs. > 5 years (late)] for both Basal-like (HR, 6.23 early vs. HR, 0.63 late) and HER2-E tumors (HR, 2.97 early vs. HR, 0.73 late) but not for LumB tumors where risk was elevated consistently (HR, 2.67 early vs. HR, 1.47 late). The contrast between LumB, HER2-E, and Basal-like compared with LumA on early recurrence was stronger when subtype was defined by PAM50 than by immunohistochemistry (IHC) markers. Conclusions: The PAM50 categorized intrinsic subtypes in a manner that more accurately predicts recurrence and survival, especially for luminal tumors, compared with commonly used methods that rely on traditional IHC clinical markers. Impact: The PAM50 is robust for use in epidemiologic studies and should be considered when archived tumor tissues are available. Cancer Epidemiol Biomarkers Prev; 23(5); 725–34. ©2014 AACR.