B. B. Cael, Gregory Britten, Francisco Calafat, Jonah Bloch-Johnson, David Stainforth, Philip Goodwin

\textbf{Abstract:} Climate projections are highly uncertain; this uncertainty is costly and impedes progress on climate policy. This uncertainty is primarily parametric (what numbers do we plug into our equations?) and structural (what equations do we use in the first place?). The former is straightforward to characterise in principle, though may be computationally intensive for complex climate models. The latter is more challenging to characterise and is therefore often ignored. We developed a Bayesian approach to quantify structural uncertainty in climate projections, using the idealised energy-balance model representations of climate physics that underpin many economists' integrated assessment models (and therefore their policy recommendations). We define a model selection parameter, which switches on one of a suite of proposed climate nonlinearities and multidecadal climate feedbacks. We find that a temperature-dependent climate feedback is most consistent with global mean surface temperature observations, but that the sign of the temperature-dependence is opposite of what Earth system models suggest. This discrepancy is likely due to the assumption that the recent pattern effect can be represented as a temperature dependence. Moreover, the most likely model is less probable than the rest of the models combined, indicating that structural uncertainty is important for climate projections. Indeed, under shared socioeconomic pathways similar to current emissions reductions targets, structural uncertainty dwarfs parametric uncertainty in temperature. Consequently, structural uncertainty dominates overall non-socioeconomic uncertainty in economic projections of climate change damages, as estimated from a simple temperature-to-damages calculation. These results indicate that considering structural uncertainty is crucial for integrated assessment models in particular, and for climate projections in general.