Climate models are useful scientific tools, but they are also objects of study in their own right. When scientists want to compare the results from two different models, or ensure that their model is realistically capturing climate behavior, they have to perform tests. There are many approaches to testing climate models, but a key approach relies on significance testing, which allows scientists to draw conclusions with statistical certainty. Researchers supported by CPO’s Climate Variability & Predictability (CVP) program have developed a specific significance test that offers simplification in estimation and interpretation of climate model applications. Dr. Timothy DelSole and Dr. Michael K. Tippett published their new approach in Advances in Statistical Climatology, Meteorology and Oceanography.
DelSole and Tippett’s approach centers on detecting and describing differences in certain types of time series data, described statistically as stationary processes, such as the Atlantic Meridional Overturning Circulation (AMOC). Their statistical test includes important considerations for working with climate data, such as working with small sample sizes. Alongside their new approach, DelSole and Tippett present a method for measuring and interpreting the distance between two processes based on classification. When applied, this method can illustrate a graphical summary of the similarities and differences between time series generated by different models.
They demonstrate their proposed approach by using it to compare simulated AMOC output from ten different global climate models. Looking at the output from ten different models, the casual observer would conclude the simulations are different--but DelSole and Tippett are interested in if these differences are statistically significant. Their method detected no significant differences for AMOC simulations from the same model, but detected differences in most cases in simulations from different models. They then use their graphical clustering approach to allow easy visualization of how different the models are from each other. In this case, the models cluster into four groups.
DelSole and Tippet are further developing their method for use in comparing multivariate time series and to compare model time series with observations.
Read the study »