A traditional starting point for a climate impact assessment is a climate scenario or a suite of such scenarios. Developers of climate scenarios acknowledge their limitations and have long been concerned with the associated uncertainties. However, communicating the source, magnitude, statistical significance, and impact of uncertainties to users of climate scenarios (e.g., impact analysts and stakeholders) remains problematic. In this paper we use the analysis of variance (ANOVA) approach to evaluate the relative magnitude of different sources of uncertainty, the "interaction" between the sources, and the statistical significance of the source and interaction terms. ANOVA is applied to 960 scenarios of projected change in seasonal means and standard deviations of maximum and minimum temperature for lake-modified locations near the Great Lakes. The two major sources of uncertainty contained within this particular suite of scenarios are 1) the five different downscaling methods used to develop the scenarios and 2) the four different GCM simulations to which the downscaling methodology was applied. The magnitude of the uncertainty introduced into the scenario suite from these two sources is compared to between-location, between-season, and between-predictand (i.e., maximum and minimum temperature) differences in projected changes of seasonal temperature for an approximately doubled CO₂ environment.

For this suite of climate change scenarios, the choice of GCM simulation introduced more uncertainty than the choice of downscaling methodology. In fact, the difference in group means is approximately twice as large between the different GCM simulations compared to the different downscaling methodologies. In addition, the between-model and between-method differences are considerably larger than between-location, between-season and between-predictand differences. Using the analysis of the seasonal means of maximum and minimum temperature as an example, the group (i.e., category) means differed by 3.0^oC between the four GCM simulations, 1.1^oC between the five downscaling methodologies, 0.9^oC between the four traditionally-defined seasons and 0.2^oC between the two predictands. Differences in the group means between the six locations were not significant. Interestingly, a number of two-way interaction terms were also significant including model and season, model and method, method and season, method and predictand, and predictand and season. Inspection of plots of the group means for the interaction terms led to interesting insights on the nature of uncertainty. For example, between-model differences in the projected change in the seasonal means are smallest during the summer season and largest in winter. In sum, ANOVA provides one means of communicating to users the uncertainty contained in a large suite of climate scenarios. Application of ANOVA to temperature scenarios for the Great Lakes region suggests that uncertainties introduced by the choice of downscaling methodology and the choice of GCM simulation are larger than seasonal and spatial variations in the projected changes in the mean and standard deviation, and that differences in the projected change in maximum temperature compared to minimum temperature are smaller than the uncertainties associated with the scenarios. This sobering finding is a caution to impact analysts to not over interpret climate change scenarios.