A procedure based on a combination of descriptive and predictive ability tests, adopted for the selection of a suitable flood frequency model for use in Tanzania, is described. Based on this study the pooled annual maximum data of 53 stations in Tanzania can be equally well described by Pearson Type 3, log logistic or the general extreme value distributions, and the preferred method of parameter estimation for all three distributions is the method of probability weighted moments. A technique based on non-parametric fitting of distribution functions is suggested for systematic calculation of safety factors.