scales; the correlation between this time series and the annual mean surface air temperature over the Northern Hemisphere is about -0.7 (Groisman et al., 1993). Using this relationship, and taking into account the intrinsic coherence of annual snowfall totals in southern Canada (which encompasses the main part of the 45°N to 55°N zone) with regional maximum temperature (Groisman, 1992), we could expect a future decrease in snowfall for this zone if the global warming projected by IPCC (1990) actually occurs.
We realize that the large adjustment coefficient for zonal snowfall in the 35°N to 45°N zone (1.47), while providing a more realistic mean value, cannot be considered a satisfactory solution to the problem of snowfall changes over this zone. Unfortunately for the current concerns about potential climatic change, we still do not have a reliable source of data on the snowfall over the main part of the United States.
Figure 6 shows the unbiased annual precipitation totals for the last four decades, area-averaged over the same four zones used in Figure 5 plus the southern part of the United States (a zone from 28°N to 35°N and to the east of 105°W). The basic statistics of these time series appear in Table 3. It can be seen from this table that annual precipitation has increased during these last decades over the entire continent, but especially significantly over the southeastern United States (14 percent per 40 years) and northern Canada (17 percent per 40 years).
Table 4 (adapted from Groisman, 1992) provides the same characteristics as in Table 3 for the century-scale precipitation changes over southern Canada and the contiguous United States. This table confirms the main conclusions derived from Table 3, but shows that, as with the U.S. precipitation, it is unwise to discuss the statistical significance of precipitation trends on the basis of the 41-year record alone (on the century scale the linear trend is not statistically significant).
Given the major terms of the water-balance equation, we can assume that, in general, the variability of annual precipitation drives the variability of annual stream flow (Linsley et al., 1958). Streamflow will thus have approximately the same large-scale features and areas of coherent change as the precipitation used to estimate it. The stream flow data are independently collected and do not exhibit the errors and data problems mentioned in previous sections (they have their own problems and errors, but do not suffer from the rain-gauge problem; being natural area averages, they give representative estimates of changes in the hydrological cycle of the regions under consideration. Therefore, in our study of the spatial correlation of the precipitation field over North America, we divided the area into several homogeneous regions on the basis of the results of a principal component analysis of stream flow data for the United States made by Lins (1985). In his analysis, Lins used a dense network of small watersheds not disturbed by anthropogenic activity. He evaluated five principal components,