would allow us to reject some high-end estimates as unrealistic by finding clear inconsistencies between models and observations.)
Currently, we rely on global average temperature as the primary observational constraint. With greater use of all available observational data (e.g., information about regional precipitation patterns), we may be able to define additional constraints. However, our progress in constraining the high end of the distribution may remain limited until we can better quantify aerosol radiative forcing.
With climate models there is a computational trade-off because going further out on the distribution tail can require a prohibitive number of computer runs. On the other hand, simple models allow an essentially unlimited number of runs, but the output is often highly sensitive to small changes in the input values, which results in an inherent uncertainty.
There may be some important gaps in our understanding of basic physical processes in the climate system, which limits our ability to credibly project high-consequence, low-probability events. There also seem to be problems in our statistical approaches for predicting such events (which may be why we find that some areas are hit with a “500-year” flood multiple times within a few decades). Such problems deserve attention because high-risk, low-probability events are often of great interest to policy makers.
It would be useful to policy makers if we could provide other, more complex indicators of climate response that are tied more directly to regional-scale changes in temperature, precipitation, and other critical parameters. Perhaps we should consider new approaches to interpreting global climate model results that would allow us to calculate how important variables (such as precipitation and sea ice cover) change in direct relation to radiative forcing, rather than just scaling everything with the global average temperature estimated by a climate model.
Improving the understanding of how climate sensitivity evolves over time could be achieved by developing models that include more regional scale information about anthropogenically driven changes in the climate system, such as human influences on land surface characteristics and hydrological and biogeochemical cycles, as well as aerosol radiative forcing changes from biomass-burning aerosols.
Model comparison studies are valuable for helping people understand the effects of particular feedbacks on sensitivity. For instance, the recent NCAR-GFDL-Hadley intercomparison studies confirmed the essential role of cloud feedbacks. Such studies are most useful if they focus on evaluating those aspects that differ most among models, and against observations, and use this to improve the models. The goal should not be to look for one “best” modeling approach or to get the same answers from all models; if that were the case, we would have a tendency to believe the models, even if there is no real basis for doing so.
Probabilistic estimates of uncertainty will continue to have an important role in advancing understanding. The output of a model is inherently statistical, and there is no escaping the need for PDFs and for ensemble modeling. The scientific community has previously advised that an assessment requires the use of at least three models and three realizations from each model. Otherwise, you cannot know whether the difference between two models is significant relative to the difference between ensemble runs from the same model.
There is reason to hope that we can further narrow the range of sensitivity estimates through a combination of climate model results and PDFs derived from simple models. These approaches are complementary for a variety of reasons. GCMs are crucial for getting the regional details correct, but simple models are needed for probabilistic studies of global mean changes. Simple modeling approaches can indicate areas of interest that can be explored in greater detail with GCMs. Finally, PDFs from simple models could perhaps be added as a new type of routine diagnostic to understand the statistics of GCMs.
Current approaches focus primarily on the simple 2*CO2 scenario. Interpretation of model output will be most useful if it focuses more explicitly on the role of other radiative forcing agents, including the recognition that climate sensitivity may vary somewhat, depending on the type of forcing