AbstractExtreme high still water levels (EHSWLs) are calculated at four tide stations located along North San Francisco Bay for a range of return periods. The observed hourly still water level varying from 7 years to 122 years measurement durations and predicted tide data for the tide stations are used in the study. The conventional annual block maxima (ABM), r-largest order statistics (r-LOS), peak over threshold (POT), and convolution (joint probability) methods are used to calculate the return levels. The return level estimates by the four methods are critically compared, and the merits and limitations of the methods are investigated. Various other research questions such as the effect of the probability distribution and block size on the return level estimates, the accuracy of return level estimates with limited annual maximums, the effect of the data resolution on the return level estimates with the convolution method are investigated. The analyses show that using a limited annual maximum affects the return level estimates most. The error can go up to 0.37 m in the 100-year return level at the San Francisco station. The direct extreme value analysis (EVA) methods produce very similar results, whereas the convolution method results in significantly (0.16 m) higher return levels. The monthly and daily block sizes in the block maxima method estimate very close return levels to annual block size. The differences produced due to the probability distributions alternative to generalized extreme value (GEV) are within or very close to the 95% confidence margin of error (MOE) of ABM-GEV estimates.