The L moments are used in the three stages of regional frequency analysis: the delineation of homogeneous regions, the identification of a regional parent distribution, and the estimation of distribution parameters. Numerical analysis is conducted on 5 min to 24 hours annual rainfall extremes from 375 precipitation gaging stations in Canada. The numerical analysis concluded that Canada could be considered as a single homogeneous region in which the L skewness and L kurtosis display no significant spatial variability. Also, on the basis of mean annual precipitation (MAP), Canada can be divided into climatologically homogeneous subregions, in which the L coefficient of variation is virtually constant. The parent distribution was identified as the general extreme value (GEV), the parameters of which depend on the MAP and storm duration. A hierarchical regional approach is proposed for fitting the identified GEV distribution, where the L skewness, L coefficient of variation, and mean are estimated on a regional, subregional, and single-site basis, respectively. Monte Carlo simulations indicate that design storms estimated by the proposed hierarchical approach are substantially more accurate than those estimated by the single-site method. The simulations also demonstrate that the proposed hierarchical approach makes the estimation of design storms at ungaged sites less dependent on the availability of precipitation data. ¿ 1999 American Geophysical Union