An article by the team of Jens Christensens, member of WCRP Joint Scientific Committee, shows that a commonly used way of presenting statistics may results in an overestimation of uncertainty in multi-model projection on global and regional scales.
Cumulative statistics. The cumulative distribution of multimodel means of projected changes, that is, the number of models/patterns that has a mean value of change for the region in concern below or equal to the value of change. Blue curves according to regional mean ranking; red curves according to grid point ranking. The horizontal bars denote minimum to maximum ranges (light transparent color), 25th to 75th percentiles (darker transparent color) and median (50%) value (colored vertical line). (a) Global and hemispheric statistics for temperature and precipitation (land plus ocean); (b) continental scale statistics for precipitation (land only).
Multimodel ensembles are widely analyzed to estimate the range of future regional climate change projections. For an ensemble of climate models, the result is often portrayed by showing maps of the geographical distribution of the multimodel mean results and associated uncertainties represented by model spread at the grid point scale. The paper shows that grid point statistics grossly degrades available patterns of information present in multimodel climate projections and consequently inflates uncertainty estimates based on a multimodel ensemble. Similar inconsistencies occur in impact analyses relying on multimodel information extracted using statistics at the regional scale, for example, when a subset of CMIP models is selected to represent regional model spread. Such analyses may lead to costly overadaptation and hence maladaptation.
The authors advise to take caution when multimodel projections are used in impact studies, and that any selection of subensembles should be based on appropriate, well-argued, and transparent choices.
The study used 39 CMIP5 models included in the IPCC Atlas for the RCP8.5 scenario.