subtracting the 1970-1984 mean (cold period) from the 1950-1964 mean (warm period). These 15-year periods were chosen on the basis of analyses of areal average SSTs in the North Atlantic (see Kushnir (1994) for details). There is a resemblance between the model and observational results, in terms of both the spatial pattern and the magnitude. It is tempting to speculate that the observed changes in SST may have arisen because of changes in the intensity of the THC in the North Atlantic. The pattern of model sea surface salinity (SSS) changes associated with fluctuations in the intensity of the thermohaline circulation, shown in Figure 4c, is somewhat similar to the pattern of model SST changes. Additional analyses have shown that the SST and SSS patterns in Figure 4 can be interpreted as the result of anomalous advection of the climatological mean SST and SSS distributions by the anomalous currents associated with variations in the THC. It should be noted that the pattern of model ocean temperature and salinity changes at 295 m depth (not shown) also resembles the observational results of Levitus (1989a), in which differences in ocean temperature and salinity were computed for two 5-year periods comprising 1955-1959 and 1970-1974.
The changes in surface air temperature associated with fluctuations in the intensity of the thermohaline circulation, computed in the same manner as for SST and SSS, are shown in Figure 5. The geographical extent of the surface
air temperature changes is not restricted to the North Atlantic, but encompasses large regions of northern Europe and the Arctic. (Note that these surface air temperature results are from the winter season only.) Over the Arctic Ocean and its vicinity, surface air temperature changes are particularly large in winter. This seasonal dependence in the high latitudes results partially from changes in sea-ice thickness associated with changes in the intensity of the thermohaline circulation. When the THC is anomalously strong, the enhanced transport of heat into high latitudes results in a reduction of mean sea-ice cover (not shown), thereby permitting an enhanced heat flux from the ocean to the atmosphere during the winter months when the mean surface air temperatures are substantially lower than that of the ocean at high latitudes. During the summer, the mean surface air temperature over the Arctic Ocean is equal to or greater than the ocean temperature at high latitudes, and there is little or no heat flux from the model ocean to the atmosphere. However, the enhanced surface absorption of summer insolation due to the reduced coverage of sea ice also contributes to the reduction of sea ice thickness and the increase of surface air temperature in winter.
The above results highlight the potential importance to climate of such fluctuations in the thermohaline circulation over the North Atlantic, Europe, and the Arctic. These variations arise in the absence of any external forcing at this time scale.
It is desirable to examine in greater detail the spatial and temporal relationships between the irregular oscillations in the intensity of the thermohaline circulation and variations in the three-dimensional fields of salinity, temperature, and density. These relations are quantified by computing, at each grid point, linear regressions between the time series of temperature, salinity, and density versus the time series of the thermohaline circulation. The regressions were computed as:
where y(t) can be salinity, temperature, or density at each grid point, t is time, a is the slope of the regression line, x(t+ t) is the time series of the THC, t is the lag, and b is the intercept of the regression line. These regressions were computed at various lags in order to provide a three-dimensional picture of the evolution of the oceanic state as the thermohaline circulation fluctuates. Prior to the regression analyses, all time series were first detrended and filtered in order to effectively remove fluctuations with time scales less than approximately 10 years. The analyses were also performed without filtering; they yielded phase relationships similar to those described below. The slope of the regression line (referred to hereafter as the regression coefficient) esti-