Interpreting Climate Model Projections of Extreme Weather Events for Decision Makers

The proliferation of output from climate model ensembles, such as CMIP3 and CMIP5, has greatly expanded access to future projections, but there is no accepted blueprint for how this data should be interpreted. Decision makers are thus faced with difficult questions when trying to utilize such information: How reliable are the multi-model mean projections? How should the changes simulated by outlier models be treated? How can raw projections of temperature and precipitation be translated into probabilities? The multi-model average is often regarded as the most accurate single estimate of future conditions, but higher-order moments representing the variance and skewness of the distribution of projections provide important information about uncertainty. We have analyzed a set of statistically downscaled climate model projections from the CMIP3 archive to conduct an assessment of extreme weather events at a level designed to be relevant for decision makers. Our analysis uses the distribution of 13 GCM projections to derive the inter-model standard deviation (and coefficient of variation, COV), skewness, and percentile ranges for simulated changes in extreme heat, cold, and precipitation during the middle and late 21st century for the A1B emissions scenario. These metrics help to establish the overall confidence level across the entire range of projections (via the inter-model COV), relative confidence in the simulated high-end versus low-end changes (via skewness), and probabilistic uncertainty bounds derived from a bootstrapping technique. Over our analysis domain centered on the United States Midwest, some primary findings include: (1) Greater confidence in projections of less extreme cold than more extreme heat and intense precipitation, (2) Greater confidence in the low-end than high-end projections of extreme heat, and (3) Higher spatial and temporal variability in the confidence of projected increases of heavy precipitation. In addition, our bootstrapping procedure provides a fairly easy way for decision makers to estimate the probability of future changes in the frequency of extreme weather events based on user-defined percentile thresholds.