Flood Frequency Analysis (FFA) typically relies on fitting a probability distribution to Annual Maximum Peak flows (AMPs) to estimate the frequency of various flood magnitudes. It is generally assumed that longer observational records enhance the reliability of FFA. In this study, we challenge this assumption by examining the Kickapoo watershed in the north-central United States where at-site FFA is susceptible to significant sampling errors despite a relatively long record of observations (90 years). We demonstrate that three exceptionally large events, with only a 1.7% chance to occur in a single watershed, significantly affect extreme quantiles and their associated confidence intervals. We argue that FFA using a weighted skewness coefficient, recommended by Bulletin 17C FFA guidelines, can yield more reliable flood frequency estimates than at-site methods by combining both local and regional characteristics. We also leverage a process-driven FFA approach, which integrates Stochastic Storm Transposition (SST) with Monte Carlo (MC) physics-based hydrologic modeling (SST-MC), to gain additional insights into flood frequency. We employed the WRF-Hydro hydrologic model and a process-based calibration approach with Fusion, a new high-resolution forcing dataset over the continental United States. By expanding the sample size and incorporating watershed-scale and regional information, SST-MC can effectively reduce the sensitivity of FFA to individual extreme events and provide more reliable frequency estimates. The SST-MC method also adds physical interpretations by quantifying internal variability in flood frequency. Our study highlights the benefits of integrating regional analysis and advanced physic-based hydrologic modeling techniques into traditional FFA.