The current statistical methods applied in flood frequency analysis require long data records to obtain reliable estimates, particularly for long return periods. Moreover, the choice of the statistical model and the parameter estimation procedure may introduce uncertainty in the estimates. In this work, we investigate the sensitivity of flood frequency analysis to various sample sizes, statistical models, and parameter estimation methods over six major hydrological regions in the contiguous United States. Results show that flood frequency estimates based on annual maximum series approach convergence to the reference values (estimates derived from 70 years record) in terms of median for 35‐year or longer records. However, the uncertainty remains significant and a record of 35 years (20 years) is associated with ~50% (100%) larger uncertainty on the estimated 100‐year flood. The generalised extreme value distribution combined with maximum likelihood estimation method is associated with the largest uncertainty, while the log‐Pearson type III exhibits comparable bias and smaller uncertainty. Application of the partial duration series approach to 20‐year records shows no significant advantage. Our findings suggest that the hydroclimatic characteristics of the catchments exhibit limited impact on the uncertainty.