"pooling" available regional flood data. A statistical characterization of the connections between ocean-atmosphere variability at interannual to decadal time scales and the frequency of the annual maximum flood in the region is needed. This relationship, coupled with "beliefs" as to scenarios for future climate derived from an analysis of the historical and paleoflood record and coupled general circulation models of the climate system, may be useful for assessing scenarios for future flood risk. A framework for formally conducting such analyses to better estimate potentially changing flood frequency distributions and their uncertainty is needed.
Second, a framework is needed for decision analysis on flood management that explicitly considers the dynamic risk and its estimation uncertainty. Clearly, such a framework needs to consider both the length of the planning period over which the projected flood risk will be used and the reliability with which the risk can be estimated from available information. Such a framework may be developed considering a "bias-variance" tradeoff or considering related explicit economic consequences. Consider first a monotonic trend in the annual maximum flood. In this case, one may be tempted to use the last 10 years of record to estimate the 100 year flood for the next 10 years (the planning period). One would reduce bias, but there would be tremendous uncertainty in risk estimates because the record is so short. If instead one had employed a 200-year period of record to project the flood risk over the next 10 years, then the bias in flood risk is likely to be larger, while the variance of flood risk estimators should be reduced. The magnitude of the expected shift (i.e. the projected bias) in the estimated 100 year flood over the next 10 years, and its economic consequences, relative to the increased uncertainty of estimate of this flood, would determine whether the shorter record is used. This answer may well be different if a 50-year planning period were considered. The bias would be larger, as would the uncertainty associated with projecting the monotonic trend into the future. This situation is complicated if quasi-periodic climate variations are considered. For instance, if a 20 year periodic climate variation were considered, using the last 10 years of record to project flood risk for the next 10 will increase both bias (as one goes from the high to the low phase of the oscillation) and variance of estimated flood risk. Explicitly conditioning the flood risk estimate on climate state has an effect similar to the selection of a subset of years of the record as discussed above. The use of such a conditional probability statement would attach higher weights to floods in years with climate state similar to the one projected and lower weights to other floods. This reduces the effective sample size used for flood risk estimation. Thus, the "conditional risk estimation" framework needed needs to consider length of record, length of planning period, the nature of the climatic non-stationarity and causal relations between the climatic factors and the floods. The utility of paleoflood and proxy climate data could be evaluated in the same framework.