Robust Bayesian computation of the fraction of attributable risk for climate damages: hypothesis-set completeness, trend-variability interaction, sensitivity, prior handling, and extension to damage types such as hurricane losses

The fraction of attributable risk (FAR) - the share of a damaging event's probability attributable to anthropogenic climate change - is the quantity that courts, insurers, and policymakers would need in order to assign responsibility for climate damages. Jaeger, Krause, Haas, Klein and Hasselmann (2008) proposed computing it by sequential Bayesian learning over a finite set of expert hypotheses about regional temperature trend and volatility, and demonstrated the method on the 2003 Alpine heat wave using the Swiss summer-temperature record, finding an anthropogenic contribution above 90% that was robust to very different prior weightings. The authors themselves listed what must be resolved before the method can defensibly attribute climate damages: the hypothesis set omits physically important nonstationary patterns (greenhouse-gas warming coupled with aerosol-driven cooling and natural long-term mechanisms, in the sense of Mitchell 1972); the interaction between an accelerating mean trend and changing interannual variability is confounded across the demonstration's hypotheses; sensitivity to station choice and station moves, event thresholds, and model structure is unquantified; prior elicitation and the communication of prior-dependence for legal, insurance, and policy use are unsettled; and the method has not been extended to other damage types such as hurricane-related losses. Later literature has executed major parts of it by other means - ensemble-based and extreme-value event attribution, including explicit FAR-based estimates of hurricane and flood losses (about USD 67 billion of Hurricane Harvey's damages attributed to human influence) - while the 2024-2026 field assessments describe impact attribution as a still-developing area with open methodological questions, and no work completes the source method's own robustness and extension program. The problem remains open in its original generality.

Background

The fraction of attributable risk (FAR) is a concept borrowed from epidemiology, where it denotes the share of disease risk attributable to an exposure. Allen (2003) proposed it as the quantity courts and insurers would eventually need for climate change: if $P_A$ is the probability of a damaging event in the world that includes anthropogenic (human-caused) climate change and $P_N$ is its probability in the counterfactual world without that influence, then

$$F=\frac{P_A(X)-P_N(X)}{P_A(X)}$$

is the fraction of the event's risk attributable to human activity. The immediate scientific context was the summer 2003 European heat wave - at the time the hottest European summer in the instrumental record, causing tens of thousands of excess deaths - for which Stott, Stone and Allen (2004) used large climate-model ensembles to conclude that human influence had at least doubled the risk of such a heat wave (a FAR of at least 0.5). The legal context was the 2007 U.S. Supreme Court decision Massachusetts v. EPA, which held that greenhouse-gas emissions are regulable air pollutants and thereby raised the prospect of liability and insurance claims that would require defensible attributable-risk numbers rather than the output of a single best-fit climate model.

Jaeger, Krause, Haas, Klein and Hasselmann (2008) - a team including Klaus Hasselmann, awarded the 2021 Nobel Prize in Physics for climate modeling - asked how such a number can be computed when experts disagree about how the climate is changing. Their answer is a Bayesian model-averaging scheme: Bayesian statistics treats probability as degree of belief updated by data, and model averaging keeps several competing models alive with weights instead of selecting one. Competing expert assessments about regional temperature evolution are formalized as a finite set of parametric hypotheses, each specifying a full probability distribution for each year's observation. In the paper's demonstration, four hypotheses describe the Swiss mean summer-temperature record (1864-2002): an unchanged (stationary) climate; a modest natural linear warming trend; an anthropogenic linear warming with growing interannual variability (the year-to-year scatter of summer temperatures around the trend); and an anthropogenic accelerating warming with constant variability. Initial prior weights encode expert disagreement, each new observation updates the weights by Bayes' rule (sequential Bayesian learning), and the FAR of a critical event - the 2003 Alpine heat wave, defined as a summer exceeding the mean of the preceding 100 years by four standard deviations - is computed by comparing the event probability from the posterior-weighted mixture of the anthropogenic hypotheses with that from the posterior-weighted mixture of the non-anthropogenic hypotheses. Starting from two very different prior weightings, the updating concentrated the posterior on the accelerating-warming hypothesis and yielded a FAR of roughly 99.8%, i.e., an anthropogenic contribution above 90% robust to the priors.

The ingredients of the open problem were already visible in the paper's own citations. Regional temperature variability itself changes under warming: Klein Tank, Können and Selten (2005) documented trends in daily temperature variance over Europe. Aerosols (airborne particles, partly human-emitted, that scatter sunlight and cool the surface) can offset greenhouse warming - the possibility analyzed by Rasool and Schneider (1971) - and natural long-term mechanisms such as the slow breakdown of the current interglacial period, which Mitchell (1972) argued human activities could intervene in, provide further physically admissible nonstationary patterns. None of these were in the demonstration's four-hypothesis set, and the paper's discussion lists what must still be done for the method to deliver defensible attributions of climate damages.

Problem Statement

The problem is the one the source paper itself poses after its demonstration: the Bayesian mixture computation of the fraction of attributable risk (FAR) for climate damages has been shown on a single event class (the 2003 Alpine heat wave, from the Swiss summer-temperature record) with a small, simple hypothesis set, and it must be made robust and extensible before its numbers can defensibly support legal, insurance, or policy claims about climate damages. The unresolved question has five interlocking components under this single objective:

(1) Hypothesis-set completeness. Bayesian learning over a finite hypothesis set can only reweight the hypotheses it is given. If physically important patterns - nonstationary combinations of long-term greenhouse-gas warming with aerosol-driven cooling and with natural mechanisms, in the sense of Mitchell (1972) - are absent from the set, the posterior FAR may misattribute risk. Can the hypothesis set be enlarged with physically motivated, expert-formulated hypotheses, and how sensitive is the posterior FAR to such enlargement? How does the posterior FAR computed under an enlarged, expert-formulated hypothesis set compare with the FAR from the demonstration's four hypotheses?

(2) Trend-variability interaction. In the demonstration, growing interannual variability and an accelerating mean trend appear in separate hypotheses, so their joint effect is confounded. A more detailed analysis of how an accelerating warming trend affects the levels of interannual variability - and of how the two jointly determine the probability of critical events - is required, and hypotheses combining both must be formulated and shown to be separable from the data. Addressing this component requires formulating at least one hypothesis in which an accelerating mean trend and changing interannual variability are modelled jointly, together with a demonstration that this hypothesis is separable on the record from the two demonstration hypotheses that carry the trend and variability changes separately, namely the anthropogenic linear warming with growing interannual variability and the anthropogenic accelerating warming with constant variability. Here "separable" means that no prior weighting of those two hypotheses - the anthropogenic linear warming with growing interannual variability and the anthropogenic accelerating warming with constant variability - reproduces the joint hypothesis's predictive distribution for the record, so that the sequential Bayesian updating - equivalently, the posterior model weights, or another explicitly stated model-comparison statistic - discriminates the joint hypothesis from them.

(3) Sensitivity. The demonstrated FAR depends on a specific aggregation of Swiss stations (with documented station moves, such as Bern's in 2006), on the chosen event threshold (a four-standard-deviation exceedance), and on the parametric forms chosen for the hypotheses. How much do defensible variations of these choices change the computed FAR?

(4) Priors in operational use. For deployment in legal, insurance, and policy contexts: how should prior weights over hypotheses be elicited from diverse expert communities and combined, and when the posterior conclusion depends materially on the prior - that is, when the data do not reconcile prior disagreement - how should that dependence be characterized and communicated? How does the variation of the posterior FAR across the admitted priors compare with its variation across the data, event-threshold, and parametric-form choices of component (3)?

(5) Extension to other damage types. Can the method be extended beyond a heat-wave indicator to other types of climate damages - for example, damages related to hurricanes (such as the losses from Hurricane Katrina, which the source paper names as its motivating example) - and do its conclusions remain robust there - to the priors and to the variations of component (3)?

References

1. Jaeger C.C., Krause J., Haas A., Klein R., Hasselmann K. (2008). A Method for Computing the Fraction of Attributable Risk Related to Climate Damages. Risk Analysis 28(4):815-823. DOI 10.1111/j.1539-6924.2008.01070.x. https://doi.org/10.1111/j.1539-6924.2008.01070.x
2. Allen M.R. (2003). Liability for climate change. Nature 421(6926):891-892. DOI 10.1038/421891a. https://doi.org/10.1038/421891a
3. Stott P.A., Stone D.A., Allen M.R. (2004). Human contribution to the European heatwave of 2003. Nature 432(7017):610-614. DOI 10.1038/nature03089. https://doi.org/10.1038/nature03089
4. Mitchell J.M. (1972). The Natural Breakdown of the Present Interglacial and its Possible Intervention by Human Activities. Quaternary Research 2(3):436-445. DOI 10.1016/0033-5894(72)90069-5. https://doi.org/10.1016/0033-5894(72)90069-5
5. Rasool S.I., Schneider S.H. (1971). Atmospheric Carbon Dioxide and Aerosols: Effects of Large Increases on Global Climate. Science 173(3992):138-141. DOI 10.1126/science.173.3992.138. https://doi.org/10.1126/science.173.3992.138
6. Klein Tank A.M.G., Können G.P., Selten F.M. (2005). Signals of anthropogenic influence on European warming as seen in the trend patterns of daily temperature variance. International Journal of Climatology 25(1):1-16. DOI 10.1002/joc.1087. https://doi.org/10.1002/joc.1087
7. Siliverstovs B., Ötsch R., Kemfert C., Jaeger C.C., Haas A., Kremers H. (2010). Climate change and modelling of extreme temperatures in Switzerland. Stochastic Environmental Research and Risk Assessment 24(2):311-326. DOI 10.1007/s00477-009-0321-3. https://doi.org/10.1007/s00477-009-0321-3
8. Pall P., Aina T., Stone D.A., Stott P.A., Nozawa T., Hilberts A.G.J., Lohmann D., Allen M.R. (2011). Anthropogenic greenhouse gas contribution to flood risk in England and Wales in autumn 2000. Nature 470(7334):382-385. DOI 10.1038/nature09762. https://doi.org/10.1038/nature09762
9. Herring S.C., Christidis N., Hoell A., Hoerling M.P., Stott P.A. (2020). Explaining Extreme Events of 2018 from a Climate Perspective. Bulletin of the American Meteorological Society 101(1):S1-S140. DOI 10.1175/BAMS-EXPLAININGEXTREMEEVENTS2018.1. https://doi.org/10.1175/BAMS-EXPLAININGEXTREMEEVENTS2018.1
10. National Academies of Sciences, Engineering, and Medicine (2016). Attribution of Extreme Weather Events in the Context of Climate Change. Washington, DC: The National Academies Press. DOI 10.17226/21852. https://doi.org/10.17226/21852
11. Murakami H., Vecchi G.A., Delworth T.L., Paffendorf K., Jia L., Gudgel R., Zeng F. (2015). Investigating the Influence of Anthropogenic Forcing and Natural Variability on the 2014 Hawaiian Hurricane Season. Bulletin of the American Meteorological Society 96(12):S115-S119. DOI 10.1175/BAMS-D-15-00119.1. https://doi.org/10.1175/BAMS-D-15-00119.1
12. Risser M.D., Wehner M.F. (2017). Attributable Human-Induced Changes in the Likelihood and Magnitude of the Observed Extreme Precipitation during Hurricane Harvey. Geophysical Research Letters 44(24). DOI 10.1002/2017GL075888. https://doi.org/10.1002/2017GL075888
13. van Oldenborgh G.J., van der Wiel K., Sebastian A., Singh R., Arrighi J., Otto F., Haustein K., Li S., Vecchi G., Cullen H. (2017). Attribution of extreme rainfall from Hurricane Harvey, August 2017. Environmental Research Letters 12(12):124009. DOI 10.1088/1748-9326/aa9ef2. https://doi.org/10.1088/1748-9326/aa9ef2
14. Emanuel K. (2017). Assessing the present and future probability of Hurricane Harvey's rainfall. Proceedings of the National Academy of Sciences 114(48):12681-12684. DOI 10.1073/pnas.1716222114. https://doi.org/10.1073/pnas.1716222114
15. Frame D.J., Wehner M.F., Noy I., Rosier S.M. (2020). The economic costs of Hurricane Harvey attributable to climate change. Climatic Change 160(2):271-281. DOI 10.1007/s10584-020-02692-8. https://doi.org/10.1007/s10584-020-02692-8
16. Frame D.J., Rosier S.M., Noy I., Harrington L.J., Carey-Smith T., Sparrow S.N., Stone D.A., Dean S.M. (2020). Climate change attribution and the economic costs of extreme weather events: a study on damages from extreme rainfall and drought in New Zealand. Climatic Change 162(2):781-797. DOI 10.1007/s10584-020-02729-y. https://doi.org/10.1007/s10584-020-02729-y
17. Noy I., Stone D., Uher T. (2024). Extreme events impact attribution: A state of the art. Cell Reports Sustainability 1(5):100101. DOI 10.1016/j.crsus.2024.100101. https://doi.org/10.1016/j.crsus.2024.100101
18. Panwar V., Noy I., Wilkinson E. (2025). Calculating loss and damage from extreme weather events in Small Island Developing States. Regional Environmental Change 25(2):73. DOI 10.1007/s10113-025-02408-7. https://doi.org/10.1007/s10113-025-02408-7
19. National Academies of Sciences, Engineering, and Medicine (2026). Attribution of Extreme Weather and Climate Events and Their Impacts. Washington, DC: The National Academies Press. DOI 10.17226/28590. https://doi.org/10.17226/28590
20. Rahmstorf S., Coumou D. (2011). Increase of extreme events in a warming world. Proceedings of the National Academy of Sciences 108(44):17905-17909. DOI 10.1073/pnas.1101766108. https://doi.org/10.1073/pnas.1101766108
21. Hansen J., Sato M., Ruedy R. (2012). Perception of climate change. Proceedings of the National Academy of Sciences 109(37):E2415-E2423. DOI 10.1073/pnas.1205276109. https://doi.org/10.1073/pnas.1205276109
22. Christiansen B. (2015). The Role of the Selection Problem and Non-Gaussianity in Attribution of Single Events to Climate Change. Journal of Climate 28(24):9873-9891. DOI 10.1175/JCLI-D-15-0318.1. https://doi.org/10.1175/JCLI-D-15-0318.1
23. Jeon S., Paciorek C.J., Wehner M.F. (2016). Quantile-based bias correction and uncertainty quantification of extreme event attribution statements. Weather and Climate Extremes 12:24-32. DOI 10.1016/j.wace.2016.02.001. https://doi.org/10.1016/j.wace.2016.02.001
24. Lass W., Haas A., Hinkel J., Jaeger C. (2011). Avoiding the avoidable: Towards a European heat waves risk governance. International Journal of Disaster Risk Science 2(1):1-14. DOI 10.1007/S13753-011-0001-Z. https://doi.org/10.1007/S13753-011-0001-Z
25. Sprinz D.F., von Bünau S. (2013). The Compensation Fund for Climate Impacts. Weather, Climate, and Society 5(3):210-220. DOI 10.1175/WCAS-D-12-00010.1. https://doi.org/10.1175/WCAS-D-12-00010.1
26. Jaeger C.C. (2022). Klaus Hasselmann and Economics. Journal of Physics: Complexity 3(4):041001. DOI 10.1088/2632-072X/ac956e. https://doi.org/10.1088/2632-072X/ac956e
27. Stecher M., Baumgärtner S. (2024). Quantifying agents' causal responsibility in dynamical systems. Ecological Economics 217:108086. DOI 10.1016/j.ecolecon.2023.108086. https://doi.org/10.1016/j.ecolecon.2023.108086