Extreme sea levels are traditionally estimated from extrapolation of observed annually maximum (or minimum ) sea level heights, by fitting an extreme-value distribution. This approach needs long-term observation, which is often not available for some newly developing areas. Pugh and Vassie (1978) proposed combining the probability distributions of tide and surge to obtain the probability distribution of tota (tide plus surge) sea level. In that study, tide and surge were regarded as independent. I ater, Pugh and Vassie (1980)studied the extreme sea level at Southend and concluded that the interaction between tide and surge was important for estuarine areas. In the present study the dependence of the probability distribution of surge on tidal level is examined. It is shown that for some Chinese coastal areas the statistical characteristics of surge in the case of lower tidal level can be twice as large as that for higher tidal level. This is attributted to the meteorological effects. The strong northerly monsoon in winter can cause larger surges and lower the tidal level at the same time. In order to take into account the dependence of the surge distribution on tidal level, the conditional distribution of surge versus tide is introduced. The marginal distribution of tide can be simply obtained through the statistics of the predicted tidal heights for a period of 19 years. The distributions of surge corresponding to different tid levels are considered as"similar”, that is, the normalized distributions of surge corresponding to different tidal levels are considered as identical to each other and are the same as the normalized marginal distribution of surge. Thus, the conditional distribution of surge can be calculated from the normalizedmarginal distribution and the standard deviation of surge corresponding to different tidal level. From the joint distribution of tide and surge the distribution of total sea level can be obtained. The extreme sea levels obtained from the joint probability method are compares with those obtaind from the traditional method. It is shown that general consistency is achieved if the observational sea level series are long, and the former results are much more stable than the latter if the observational sea level series are short.