Regional flood frequency analysis (RFFA) is a widely adopted approach for estimating design floods in ungauged catchments. However, previous studies in Australia have reported substantial inaccuracies in flood quantile estimates for several catchments, with relative errors exceeding 50%. This study investigates the factors contributing to such poor performance in RFFA. A total of 201 gauged catchments across southeastern Australia, spanning New South Wales and Victoria, were analyzed. A log-log linear model was employed to develop prediction equations for the 1-in-20 annual exceedance probability (Q20), and model performance was evaluated using a leave-one-out (LOO) validation approach within a Quantile Regression Technique (QRT) framework. Catchments were subsequently classified into two groups based on absolute relative error (ARE) values: best-performing sites (Gr-BPS, lowest 25%) and poor-performing sites (Gr-PPS, highest 25%). Both QRT and the Index Flood Method (IFM) were then applied to these groups as well as the full dataset, and relative errors were computed for each site. Results indicate that variability in predictors and hydrological coherence exert a stronger influence on model performance than regional homogeneity alone. These findings suggest that reliable flood quantile estimation in southeastern Australia requires not only attention to regional homogeneity but also careful consideration of predictors’ coherence, with QRT demonstrating superior robustness over IFM in heterogeneous catchments.