Observational uncertainty is the primary source of climate uncertainty in impacts projections
by Kevin Schwarzwald, Nathan Lenssen, Thomas Bearpark, Alia Bonanno, Radley Horton, and Gernot Wagner
Fig. 1 Observational uncertainty and nonlinear impact amplification. (A) Exposure to daily temperature, fit using historical temperatures from four reanalyses (shown as population-weighted histograms over regions used in the fit). (B) Four dose-response curves for the quartic polynomial relationship between annual all-cause mortality rate and exposure to daily temperature. (C) Change in global population-weighted temperature between Global Warming Levels (GWLs) 0.61 (reflecting the late 20th century global climate) and 2 (representing a future world that has undergone additional climate change) for CanESM5, SSP2-4.5, r10i1p1f1, a randomly chosen future GCM projection, averaged across all four reanalysis bias-corrections. (D) Change in mortality rate for the same ensemble member associated with the change in temperature exposure in B, using the average of all four dose-response functions. All filled plots use 1-degree bins. Note that, since mortality data is only available for a subset of the globe, the historical in-sample climate in A is different from the baseline for B and C; the shape of the global population-weighted temperature baseline distribution at GWL0.61 is shown in a dashed line (normalized to fit within the panel) in C for reference. See fig. S2 for a version showing only in-sample changes in temperature and mortality rate.
Abstract:
Projections of future climate damages underpin environmental policy and risk management but depend on inherently uncertain climate data. Uncertainty in historical climate observations, used to train impact models and calibrate climate projections to the observed record, is rarely considered. We show this “observational uncertainty” contributes more than the spread across or within global climate models to projections of mortality, GDP per capita growth, and payroll. Most of this uncertainty stems from training impact models; even small differences in historic weather estimates can cause large differences in projected damages. Incorporating this uncertainty increases the spread in climate uncertainty by up to 300%, highlighting the importance of impact model design and the need for better constraints on observations, and implying a higher Social Cost of Carbon.
Preprint via EarthArXiv: doi.org/10.31223/X5BZ39
Presented at EGU (Vienna, Austria, 7 May 2026) under prior title "The choice of historical data product dominates climate uncertainty in projections of climate impacts in a 2-degree world" (EGU General Assembly 2026, Vienna, Austria, 7 May 2026, EGU26-11680, https://doi.org/10.5194/egusphere-egu26-11680, 2026.)
Replication data: doi.org/10.5281/zenodo.20669397
Underlying climate impact data set: BCD-ME; citation: Schwarzwald, Kevin, Nathan Lenssen, Radley Horton, and Gernot Wagner. "A bias-corrected & downscaled massive ensemble to diagnose uncertainty in climate impact projections." Scientific Data (19 October 2026).
