The Surveillance, Epidemiology, and End Results (SEER) program is the most comprehensive and reliable source of population-based information on cancer incidence and survival in the US. The large SEER dataset is an outstanding resource for researching and explaining aspects of cancers. While SEER data-based research has great potential utility in improving survival outcomes and reducing cancer burden in the US, individuals in this dataset can have a systematically heterogeneous hazard risk of mortality. Patients in the SEER dataset are diagnosed over many years. Risk factors may change over time, and many important variables may not be available in SEER data. These together impose a heterogeneous mortality risk. Hundreds of published SEER data-based studies used a proportional hazards model, and using this model without accounting for heterogeneity may cause biased estimates. This study used SEER data from 1973-2011 to evaluate the prognostic impact of race-ethnicity on pediatric acute myeloid leukemia (AML) survival. It detected a serious bias in the estimation of the mortality risk using proportional hazards model. The main cause of the bias is heterogeneity in the baseline hazard risk due to the concurrence of the trend in declining death rate and altering composition of race-ethnicity over time. The frailty model improved the accuracy of estimation because it accounted for the heterogeneity in the baseline hazard risk among patients in SEER data.