We investigate the validity of implicit assumptions in regional flood frequency analysis (RFFA) using Monte Carlo-style simulations of three distributed hydrological models forced with rainfall events generated using stochastic storm transposition. We test the long-standing assumption that for a set of sites within a region, physical homogeneity — defined in terms of the variability of meteorological inputs, the physics of runoff generation, and runoff routing — implies statistical homogeneity of peak flows— defined in terms of the existence of a common underlying statistical distribution with parameters that can be inferred using information from neighboring sites. Our modeling results do not support this assumption, with potentially important implications for RFFA methodologies and for the very definitions of homogeneity. We show that statistically homogeneous rainfall does not translate into predictable peak flow distribution parameters across drainage scales. Specifically, we show that changes in the coefficient of variation and skewness of peak flows cannot be inferred from upstream area alone, making popular regionalization techniques such as the index-flood method and quantile regression inadequate approximations for flood frequency estimation. Our findings are consistent across the three hydrological model formulations, lending confidence that our conclusions are not an artifact of epistemological model decisions. Finally, we argue that our methodology can serve as a framework to test new proposed empirical RFFA methods, and that it opens the door to a unified physics-informed framework for prediction of flood frequencies in ungauged basins embedded in gauged regions.