Statistical analyses of model ensembles for policy support

Scenario-thinking has become common in modelling for policy support, particularly through the climate change mitigation scenarios developed over the successive IPCC reports. Often, multi-model ensemble runs are performed, representing a set of possible descriptions of the modelled system to account for our limited understanding of this system. For each scenario, this results in a distribution of model outcomes that represents the range of potential system states and their probabilities. In our opinion, model outcome distributions are not used to the fullest at the moment. For example, only the median and minimum and maximum are reported, possibly because policy makers are not trained as statisticians and demand simplicity. Thereby, valuable uncertainty information is discarded. The general situation of having one or more distributions of a variable is not new: scientific domains that deal with experiments, the collection of data samples, or user studies have a developed a wide range of statistical methods to derive information from such distributions. We argue that these statistical methods can be applied to multi-model ensembles to translate model output uncertainty into simple, policy-relevant information. We aim to demonstrate this on the multi-model ensemble from the SSP/SR1.5-database. Herewith, we answer questions such as: 1) How certain can we be that we will stay below a particular GHG-emission/cost/carbon-price limit for each scenario? 2) Are the differences between scenarios significant given the uncertainty in the model ensemble? The answers show whether the designed mitigation measures are robust in the light of our limited understanding about the modelled system.