Stabilising CO2 concentration as a channel for global disaster risk mitigation Authors: Saite Lu, Demosthenes Tambakis Published: 2024 ABSTRACT AbstractWe investigate the influence of anthropogenic CO2 concentration fluctuations on the likelihood of climate-related disasters. We calibrate annual incidence rates against global disasters and CO2 growth spanning from 1960 to 2022 based on a dynamic panel logit model. We also study the sensitivity of disaster incidence to stochastic carbon dynamics consistent with IPCC-projected climate outcomes for 2100. The key insight is that present and lagged CO2 growth contains valuable information about the likelihood of future disaster events. We further show that lowering carbon stock uncertainty by dampening the persistence or the variability of CO2 concentration has a first-order impact on mitigating expected disaster risk.Keywords: Climate change mitigation, CO2 concentration, Disaster incidence rate, Logit regressionSubject terms: Natural hazards, Environmental social sciences Introduction There is clear evidence that the annual incidence of climate-related disaster events, including droughts, floods, storms, heatwaves, and forest fires has risen consistently since the mid-20th century1–22. The positive trend is reflected in greater disaster severity, proxied by the number of affected individuals23, and negative impacts on output8 and productivity24. Further, it is very likely that there are tipping points in the Earth System—abrupt regime shifts to the temperature feedback from emissions driving global warming and to the deep oceans’ carbon sink capacity1–9,25–27—such that short-term disaster risk may coincide with collapsing biospheres and ecosystems, amplifying economic damage and human losses6. Climate science underscores the unequivocal anthropogenic influence of greenhouse gases-predominantly carbon dioxide (CO2)-on disaster risk28. (The main non-CO2 greenhouse gases (GHG), methane (CH4) and nitrous oxide (N2O) are also projected to grow in the 21st century1. CH4 dynamics are particularly sensitive to global surface temperature tipping points. In this paper we abstract from their contribution to disaster risk. Reference16 studies the slow decay rate of GHG concentration levels: future warming due to the long time scales of ocean-linked heat transfer could remain for millennia even if human emissions were to stop immediately.) This paper explores the predictive power of stochastic carbon stock dynamics for short-term disaster risk by modelling the sensitivity of 1-year-ahead disaster frequency to fluctuations in global average atmospheric CO2 concentration. To this end, we expand upon the logit-based framework of Refs.14,17 for the likelihood of financial crises. Our analysis incorporates the annual global disaster record spanning 63 years (1960–2022) maintained by the Centre for Research on the Epidemiology of Disasters (CRED). Here, incidence denotes the annual conditional likelihood of a climate-related disaster event. Conditional disaster incidence is driven by the growth of carbon stock and adjusted for geographic factors including countries’ absolute distance from the equator, the length of their coastlines, and countries’ urban population share as a socio-economic control variable27. Based on the logistic elasticity estimates, the binary classification power (disaster/no disaster) of our model is evaluated using the Area Under the Receiver Operating Characteristic (AUROC) curve, which exceeds 85%. Further, the cumulative impact of 1% lower CO2 concentration on certainty-equivalent (“no uncertainty”) disaster likelihood lies between ∼0.7 and 1.1%. Such estimates are consistent with the short-term attribution of climate-related disasters to anthropogenic carbon emissions, documented in Ref.29 and quantified by Ref.10,30. The paper’s contribution is twofold. First, it contributes to applied research by modeling the risk reduction potential of different decarbonisation strategies19. The research closest to our specification is27. These authors’ econometric method differs from ours in that they regress carbon emission levels to apply linear cointegration techniques, while we estimate a logit model. Our logistic approach is novel in a climate-change context because it captures the potentially nonlinear (sigmoid) relationship between disaster risk and carbon emissions. (The usefulness of the logistic approach was validated in economics literature by Ref.14, who demonstrated a nonlinear historical relationship between credit growth and the risk of financial crises, as well as in Refs.31,32. Another study, employing linear panel regressions across seven world regions, is4. Non-regression-based studies linking disaster risk to CO2 emissions include33, using a dynamic stochastic general equilibrium (DSGE) model, and Ref.34, applying non-parametric machine learning.) Further, Ref.27 finds the elasticity of disasters to CO2 emissions is ∼9, implying a disaster count doubling-time estimate of 13 years. Against that, our estimated elasticities imply that under the current trend in atmospheric CO2 concentration, the total disaster count would double from an average 1.7 to ∼3.4 events per country-year over the next 15 to 25 years. This timeframe is broadly consistent with the 40% increase in global disasters by 2030 forecast by Ref.21 and with the shorter time window available for recovery following successive disaster events22. Disparities in doubling times are important for national and global stakeholders to judge the scale of disaster adaptation and mitigation policies. Lastly, Ref.27 controls for multiple socio-economic determinants but its scope is confined to hydrological disasters. By contrast, by including all climatological disasters events between 1960 and 2022 we obtain a more comprehensive measure of global climate-related disaster risk. While in 2000–2019 climatological disasters accounted for about 16% of all disaster events, unchanged from 1980 to 1999, extreme temperature events have more than tripled, from 130 to 432, and wildfires have also risen sharply23. In recent years more frequent and longer heat waves and droughts have drawn wide international media coverage. Second, our reduced-form approach complements the fast growing research on the impact of climate risks on optimal climate policy. The results reinforce those arguing for a higher precautionary and insurance response to the volatility of carbon dynamics35. Because mean surface temperature sensitivity is concave (logarithmic) in the doubling of CO2 stocks (Arrhenius’ Law20) the Social Cost of Carbon (SCC) impact of uncertainty about carbon flows is dampened, while that of uncertainty about temperature feedbacks is not. Such uncertainty may lead to the 200-year SCC rising by over 20%36. (The SCC represents the expected present discounted value of all future marginal damages resulting from emitting a ton of CO2 today. When one ton of CO2 is emitted into the atmosphere, approximately 0.45 tonnes remain, with the remainder being absorbed by land and ocean sinks. Although this atmospheric increase may exhibit significant short- and long-term variation due to feedback effects and tipping points, we can assume the conversion factor is constant as a first-order approximation. We are grateful to Glen Peters at the Center for International Climate Research (CICERO), Norway, for this observation.) We find that scaling down the persistence and volatility of fluctuation has first-order effects on expected disaster risk. Climate science is advising that delaying mitigation efforts risks augmenting total disaster risk, with the associated short- and long-term costs29. Against this IPCC-led consensus, our key policy insight is that stabilizing carbon stock is an independent avenue for disaster risk reduction which acts to enhance any existing mitigation targets. We run logit regressions with one-way (country) and two-way (adding time) fixed effects. The logistic specification is intuitive and allows for directly testing the underlying nonlinear relationship between CO2 growth and one-period-ahead conditional disaster risk. The key role of nonlinearity is emphasised by the cumulative nature of disaster tipping points9 and the increasing interconnectedness of climate hazards6. We include lagged disaster risk on the right-hand side (see also Refs.4,37), but to minimise the risk of overfitting, we do not include interactive fixed effects (The results remain robust once interactive fixed effects are introduced. See Appendix B.), as suggested by Ref.38 in the baseline analysis. Other logistic applications to disaster risk exposure are mostly limited to a single sector or country, such as water resource management in South Korea39. To the best of our knowledge, our specification is the first to model the climate-induced disaster impact of carbon emissions on a global scale. We find the certainty-equivalent disaster incidence rate is 6.2% annually. Separately, we estimate a stationary AR(3) process for atmospheric CO2 concentration. We detect strong (3-year) persistence of disaster likelihood to historical CO2 growth. The logit model is calibrated to carbon flows’ historical mean and standard deviation. Given the logistic transfer parameter estimates, we simulate Gaussian CO2 growth shocks and generate 1-year-ahead disaster incidence trajectories controlling for countries’ urban population share. The lowest predicted 95% confidence interval for stochastic incidence across our specifications is [6.7, 7.8]% p.a., and around 20% of total disaster risk is attributed to carbon stock uncertainty. We also find disaster risk estimates and implied elasticities are greater with time-fixed indicator variables, reflecting the higher convexity of the logit parameters under two-way fixed effects. Although they are robust to parameter uncertainty, a caveat of these baseline estimates is our assumption of reliable vulnerability data, represented here by countries’ urban population share although disaster exposure is a multifaceted process40. For sensitivity analysis we focus on the logit parameters estimated with one-way fixed effects. We calibrate the CO2 growth mean to the end-horizon (2100) CO2 stock level aligned with a specific IPCC Representative Concentration Pathway (RCP2,41), while CO2 growth variability is either set to its historical baseline or adjusted to half the baseline value. Each pathway specifies a time series of future CO2 and GHG concentrations leading to 2100. We find that certainty-equivalent disaster risk drops from ∼5.5 to 1% p.a. along with average CO2 concentration, while total disaster risk is reduced from ∼6.5 to ∼1.2% p.a. under baseline variability, with a corresponding narrowing of the uncertainty wedge. Successive negative increments to expected disaster risk exceed the continuous cumulative elasticities discussed above. Further, average risk reduction is greater (down to ∼1% p.a.) if CO2 concentration becomes more stable. This finding is supported by numerical evidence of the uncertainty wedge narrows to nearly zero. For expected disaster risk to fall, concentrations must move away from its baseline mean and/or standard deviation. Nonetheless, we find the projected uncertainty wedge under the most pessimistic pathway for 2100 (RCP8.5, representing business-as-usual emissions rising throughout the 21st century) is just ∼0.1% p.a. below baseline. By contrast, for the more ambitious RCP scenarios (the very stringent mitigation pathway consistent with RCP2.6, as well as the intermediate mitigation efforts associated with RCP4.5) maintaining the post-1960 average CO2 growth rate while halving its variability would result in a lower total disaster risk compared to baseline variability and a near-zero (<0.5% p.a.) uncertainty wedge. The key policy implication is that mitigation policies targeting lower average CO2 concentration could achieve a greater expected reduction in disaster incidence if they also aimed to stabilise carbon growth. This paper argues that aggressive decarbonisation strategies not only offer the benefits of mitigating short-term (annual) disaster risks, but also contribute to dampening the catastrophic long-term (millennial) consequences for human civilisation if business-as-usual emissions continue, as this would very likely trigger multiple irreversible climate tipping points9,13. (Critical systems that we rely on for survival-including cryosphere subsystems (the Greenland and West Antarctic ice sheets), biosphere subsystems (the Amazon Rainforest), and large-scale oceanic and atmospheric circulation patterns such as the Atlantic Meridional Overturning Circulation (AMOC)-are at risk of collapse due to climate tipping points5. This would be disastrous for human society and the economy unless drastic carbon emission reductions are urgently implemented.) In a companion paper we study the long-term implications of these insights using non-homogenous Markov chains18. In the remainder of the paper, “Data and methods” presents the data, defines disaster incidence, and derives the uncertainty wedge; “Results” outlines the empirical results; “Sensitivity analysis and discussion” contains sensitivity analysis; and “Concluding remarks” concludes. Appendix A describes the simulation methods and RCP assumptions, while Appendix B provides results from multiple robustness checks. Throughout, we refer interchangeably to disaster incidence and risk, as well as to CO2 stock and concentration. Data and methods The data We employ EM-DAT, the CRED database, for the occurrence and socio-economic impact of climate-related disaster events globally23,42. The primary hazards are hydro-meteorological—rainfall-induced flooding, related landslides, tropical cyclones and storms—and climatological, including droughts, wildfires, heatwaves and extreme cold events. Our analysis encompasses these extreme event categories delineated by country and year and recorded from 1959 to 2022. The total number of countries and territories ranges from 158 to 218 depending on the fixed effect specification. We exclude non-climate related geophysical (earthquakes and volcanoes) and biological events (virus outbreaks and epidemics) and include concurrent and multiple disaster events even if they overlap chronologically28. The incidence rate serving as dependent variable quantifies the conditional probability of a disaster event occurring in a specific country within a year. It is thus a disaster count unadjusted for the event’s duration. We recognise that the EM-DAT maintained by CRED has several limitations that may impact the accuracy of our analysis. The dataset often lacks granularity, particularly for disasters that span multiple countries and years, making it difficult to accurately capture localized impacts. Additionally, EM-DAT includes only events meeting specific thresholds, which may lead to underreporting of smaller-scale disasters, particularly in countries with effective mitigation systems. There is also a lack of consistent reporting standards across countries, with better-resourced nations providing more comprehensive records compared to those with limited administrative capacity. Moreover, data gaps and inconsistencies in historical records affect comparability over time and between countries. The data on atmospheric CO2 concentration is documented by the US National Oceanic and Atmospheric Administration (NOAA), measured in parts per million by volume (ppm) and adjusted for seasonality. (The growth in atmospheric CO2 concentration is an underestimate of total GHG growth; (The main non-CO2 greenhouse gases (GHG), methane (CH4) and nitrous oxide (N2O) are also projected to grow in the 21st century1. CH4 dynamics are particularly sensitive to global surface temperature tipping points. In this paper we abstract from their contribution to disaster risk. Reference16 studies the slow decay rate of GHG concentration levels: future warming due to the long time scales of ocean-linked heat transfer could remain for millennia even if human emissions were to stop immediately.). NOAA initiated monthly CH4 and N2O concentration measurements in 1983 and 2001, respectively. The estimated uncertainty in the annual CO2 mean is ≈0.12ppm, derived from the annual mean values recorded by NOAA.) Figure 1 displays the empirical histograms of annual CO2 growth rates in percent p.a. (panel (a)) and disaster incidence rates (panel (b)), superimposing smooth Gaussian and beta pdf’s fitted by maximum likelihood. Descriptive statistics. The mean incidence rate over the entire period (1960–2022) is 6.2%, with a standard deviation of 5.6% annually. Average incidence has been on an upward trajectory for shorter subperiods and its standard deviation has been declining (not reported). Disaster incidence and uncertainty wedge Disaster incidence in year t+1, pt+1, is the conditional probability of a disaster event occurring in year t+1 (binary state st+1=1) given no disaster in year t (binary state st=0). Disaster risk is often assumed to follow a Poisson arrival process in continuous time, e.g.33. By contrast, we define pt+1 to be a logistic function of the growth in global average atmospheric CO2 concentration in year t-j, ct-j=defΔlogCt-j, with j=0,...,P, and constant unconditional moments E[ct-j]=μc and var[ct-j]=σc2, all j by stationarity, and Xt=[X1t,...,XKt] to be a K×1 vector of other relevant year-t information: where hj and γk are the elasticities of pt+1 to CO2 concentration growth in lag year j and non-carbon factor k∈{1,...,K}, respectively, and h¯ is a scaling constant such that disaster incidence is bounded (pt+1∈[0,1]). Evaluating pt+1 at μc and E[Xt] yields the certainty-equivalent disaster incidence rate, denoted μ¯p: We note the average growth of CO2 concentration, μc≈1.6% p.a. from Fig. 1a impacts μ¯p while its volatility σc≈0.7% p.a. does not. Further, pt+1(·) is strictly convex to the left of its unique inflexion point, hence the expected incidence rate in this range exceeds μ¯p by Jensen’s inequality. We henceforth refer to this analytical property of convex transfer as Property 1: UWp is the positive uncertainty wedge between the two measures. In other words, Property 1 states that stochastic disaster risk exceeds its certainty equivalent when model-implied incidence is a convex function of annual mean CO2 growth. Further, subject to scaling and translation, the logistic (sigmoid) function is proportional to a hyperbolic tangent function3: where x is the exponential argument in Eq. (1), μ¯p is scaled to 12 wlog, and From Eqs. (4) and (5) it follows that dE[UWp]dμc and dE[UWp]dσc are both positive. In what follows we refer to this complementary analytical feature of convex transfer as Property 2. That is, the magnitude of E[UWp] grows in both average CO2 emission growth and its variability. Therefore, the uncertainty wedge widens in the persistence and volatility of CO2 growth and in the estimated sensitivity of disaster incidence to such growth. Put differently, UW_{p} narrows if either factor is dampened. (Insofar as the global economy successfully adapts to climate change-driven disasters the transfer function’s degree of convexity may weaken. The policy consensus is that there are important synergies between climate risk adaptation (CCA) and disaster risk reduction (DDR) such that CCA and DDR strategies should be unified19. Adaptation dynamics lie beyond the scope of the our study.) Figure 2 illustrates such a mean-reducing spread for the logistic function. Disaster frequency and CO2 stock variability with logistic transfer. Uncertainty is captured by the chord connecting symmetric pair of points (-a,a) and (-b,b) around a given point c0 on the x-axis, measuring CO2 growth. We note the chord shifts out with CO2 stock volatility at a given mean. Hence, more stable carbon emissions—for example, moving from b to a—result in lower expected disaster risk. This relies on the logistic (sigmoid) inflection point being highly positive such that the convex range is to its left; that is the case for all our logit parameter estimates. Indeed, it is easy to check that dE[UWp]dσc50%) of total disaster risk for the one (two)-way estimates. In other words, lower (higher) elasticity parameter estimates tend to imply lower (higher) disaster incidence after controlling for countries’ urban population share. As discussed above, relatively less (more) convex logistic transfer in Eq. (7) would indicate weaker (stronger) human interference with climate, ceteris paribus. Moreover, the average model-implied volatilities in row 5 overpredict historical variability of σ^p=5.6% p.a. under all specifications except Gaussian shocks and one-way fixed effects, and the root mean square error of model-implied p~ against trajectories generated by a beta distribution calibrated to the historical record (p~B) is considerably higher under two-way fixed effects (row 6). Lastly, Table 5 presents simulated disaster risk statistics consistent with the 2.5th (weak convexity) and 97.5th (strong convexity) percentiles of baseline elasticity point estimates, assuming the mean and variance of CO2 growth are fixed to μ^c and σ^c. The 3-year cumulative change in certainty-equivalent disaster risk from varying average CO2 stock now ranges from dlnμ¯p/dμc∼0.5 to ∼3% (rows 9 and 14). With one-way fixed effects, expected annual disaster risk in row 6 exceeds μ¯p by upto 3.1% for the 97.5 percentile elasticity estimates (Property 1). By contrast, assuming strong convexity the two-way fixed effect parameter estimates imply E[μ~p]2w≈27% p.a., of which ∼75% is the uncertainty wedge (row 11). The share of total disaster risk contributed by uncertainty ranges from ∼25% with one-way to >60% with two-way fixed effects. Table 5Baseline: robustness to convex logistic transfer.ConvexityWeakStrongCI percentile2.52.597.597.5Parameters [{h^j}j=03]2.5 [hurb=0]2.5 [{h^j}j=03]97.5 [hurb=0]97.5 μ¯p (% p.a.)6.26.26.26.2Mean E[μ~p]1w (% p.a.)[6.3, 7.1][6.7, 7.8][7.1, 8.7][7.4, 9.3] E[UWp~]1w [0.1,0.9][0.4,1.5][0.8,2.4][1.2,3.1]SD E[σp~]2w[2.6, 3.1][3.6, 4.6][5.0, 6.7][6.1, 8.3] RMSE(μ~p,μ~pB)1w 0.0100.0130.0150.021 [dlnμ¯p/dμc]1w 0.470.620.820.95Mean E[μ~p]2w[16.4, 18.8][8.2, 10.1][19.2, 26.8][14.5, 20.6] E[UWp~]2w [10.2,12.6][2.0,4.0][13.0,20.6][8.3,14.4]SD E[σp~]2w[17.0, 19.3][2.4, 4.5][22.6, 27.6][15.3, 20.6] RMSE(μ~p,μ~pB)2w 0.1390.0330.2780.167 [dlnμ¯p/dμc]2w -0.97 -0.17 3.052.26All entries are in percent p.a. rounded to 1 d.p. The 2.5% and 97.5% confidence point estimates for columns 2–5 are available upon request. For each column the logit intercept h¯ is chosen to calibrate μ¯p. μ^p is average disaster incidence from 1960 to 2022. Stochastic measure E[μp~] is computed as the average incidence rate realized over 2000 runs of 2000 years each. The expected uncertainty wedge CI, E[UWp~], is the difference between the CI for E[μ~p] and μ¯p. Simulated skewness and kurtosis values are not reported for space constraints; they are available upon request. Baseline: robustness to convex logistic transfer. All entries are in percent p.a. rounded to 1 d.p. The 2.5% and 97.5% confidence point estimates for columns 2–5 are available upon request. For each column the logit intercept h¯ is chosen to calibrate μ¯p. μ^p is average disaster incidence from 1960 to 2022. Stochastic measure E[μp~] is computed as the average incidence rate realized over 2000 runs of 2000 years each. The expected uncertainty wedge CI, E[UWp~], is the difference between the CI for E[μ~p] and μ¯p. Simulated skewness and kurtosis values are not reported for space constraints; they are available upon request. In the rest of this paper we maintain the one-way fixed effect elasticity estimates because model-implied expected disaster risk is more robust to the degree of convexity. Sensitivity analysis and discussion In the previous exercise the unconditional mean and variance of carbon stock were fixed. To explore Property 2 we now vary both based on IPCC RCP scenarios1,41. Following RCP8.5, carbon stock continues to rise through the 21st century, peaking at ∼1370ppm by 2100. At the opposite extreme, CO2 concentrations consistent with RCP2.6 are expected to peak at ∼490ppm sometime before 2100 (2024:09 was ≈422ppm), declining afterwards. (RCP8.5 implies μc=1.54% p.a. between 2023 and 2100, slightly less than the historical mean. It coincides with the baseline calibration of Ref.35 as it implies CO2 concentration growth is ≈13 ppm p.a. RCP2.6 assumes full decarbonization: μc=0. Such an ambitious outcome would require massive investment in negative emission technologies and scaled-up CO2 capture.) Figure 4a displays CO2 stock paths consistent with the trend growth rate under each RCP. Model projections 2023–2100. Table 6 presents disaster incidence statistics consistent with each pathway’s mean CO2 growth μc (row 2) coupled with baseline σ^c=0.68% p.a. (rows 5–10), and with stock levels stabilized to σcmin=0.30% p.a., i.e. a stabilized concentration scenario (rows 11–16): (We calibrate the CO2 concentration standard deviations for this stabilizing scenario to the multi-model analysis of Ref.15. These authors compare the post-2030 and post-2050 emission projections of 7 integrated assessment climate models; see Appendix A.) Table 6Sensitivity to CO2 stock uncertainty.2100 horizonRCP2.6RCP3.4RCP4.5RCP6RCP8.5μc (% p.a.)0.190.340.560.911.54μ¯p (% p.a.)1.01.21.62.55.6Δμ¯p/Δμc−1.51.72.23.5Mean E[μ~p∣σ^c][1.1, 1.3][1.3, 1.6][1.8, 2.1][2.8, 3.3][6.0, 7.1]E[UWp~][0.1,0.3][0.1,0.4][0.2,0.5][0.2,0.7][0.4,1.5]SD E[σ~p][0.7, 0.9][0.8, 1.1][1.1, 1.5][1.6, 2.2][3.4, 4.4]Skew s~p[1.4, 3.1][1.4, 3.2][1.4, 3.1][1.4, 2.9][1.2, 2.4]Kurt κ~p[5.8, 22.8][5.7, 24.2][5.6, 22.6][5.4, 20.4][4.9, 13.6]ΔE[μ~p]/Δμc (%)−[1.9,3.2][2.1,2.9][2.5,3.2][3.7,4.5]Mean E[μ~p∣σcmin][1.0, 1.1][1.2, 1.3][1.6, 1.7][2.5, 2.7][5.6, 6.0]E[UWp~][0.0,0.1][0.0,0.1][0.0,0.1][0.0,0.2][0.0,0.4]SD E[σ~p]0.3[0.3, 0.4][0.5, 0.6][0.7, 0.9][1.6, 1.9]Skew s~p[0.6, 1.3][0.6, 1.3][0.6, 1.3][0.6, 1.3][0.5, 1.1]Kurt κ~p[3.3, 6.9][3.3, 6.7][3.3, 6.6][3.2, 6.4][3.2, 5.9]ΔE[μ~p]/Δμc (%)−[1.3,1.9][1.6,1.9][2.1,2.3][3.4,3.7]All entries except rows 4 are 95% CI rounded to 1 d.p. μc=ΔCO2t is the average annual CO2 concentration growth from 2023 to 2100 consistent with the 2100 concentration level projected for each pathway. We adopt RCP2.6 as reference for ΔE[μ~p]/Δμc so the specific gradients are undefined. Stochastic p~ is the average incidence rate over 2000 runs of 2000 years each. Excess kurtosis is κ~-3. The baseline elasticity parameter estimates {h^i}i=13, h^urb, h^yt-1 are in Table 1, and h¯ is chosen to calibrate μ¯p=6.2% for each pathway. E[UWp~] is computed as in Table 5. The baseline CO2 growth st.dev. is σ^c=0.68% p.a. and the more stable scenario is σcmin=0.30% p.a. Sensitivity to CO2 stock uncertainty. All entries except rows 4 are 95% CI rounded to 1 d.p. μc=ΔCO2t is the average annual CO2 concentration growth from 2023 to 2100 consistent with the 2100 concentration level projected for each pathway. We adopt RCP2.6 as reference for ΔE[μ~p]/Δμc so the specific gradients are undefined. Stochastic p~ is the average incidence rate over 2000 runs of 2000 years each. Excess kurtosis is κ~-3. The baseline elasticity parameter estimates {h^i}i=13, h^urb, h^yt-1 are in Table 1, and h¯ is chosen to calibrate μ¯p=6.2% for each pathway. E[UWp~] is computed as in Table 5. The baseline CO2 growth st.dev. is σ^c=0.68% p.a. and the more stable scenario is σcmin=0.30% p.a. We make three observations. First, certainty-equivalent disaster incidence μ¯p (row 3) declines sharply with μc, from 5.6% p.a. under RCP8.5 to 1% p.a. under RCP2.6. Mitigated disaster risk reflects lower trend CO2 growth, and the same negative trend is seen for E[μ~p∣σ^c] in row 5 (Property 1). Second, while μ¯p is only impacted by μc, total disaster risk is also sensitive to σc. In particular, E[μ~p] shifts left from E[μ~p∣σ^c] to E[μ~p∣σcmin] (row 11). The shift’s magnitude from RCP8.5 to RCP2.6 grows with the ambition of decarbonization targets for 2100. Moreover, for any given μ¯p, the average uncertainty wedge E[UWp~] (rows 6, 12) narrows with more stable emissions (Property 2). For example, at the higher end of each confidence interval the expected disaster risk share attributed to uncertainty, E[UWp~]/E[μ~p] is about 25% for σ^c but falls to under 10% for σcmin. We also note that average incidence rate volatility drops with μc, and more stable carbon emissions result in less fat-tailed and less skewed incidence distributions, all else equal (rows 14–15). (By focusing on Gaussian shocks to CO2 stock we have implicitly ruled out catastrophic risk. We conjecture that Properties 1 and 2 extend to extreme value shock probability density functions in the sense of51 provided they are symmetric.) Third, observe that the discrete analogs to response coefficient dlnμ¯pdμc are Δμ¯pΔμc for certainty-equivalent disaster risk, and ΔE[μ~p]Δμc for expected disaster risk. To compute these, we evaluate the response of μ¯p and E[μ~p] to a 1% decline in μc from RCP8.5 through to RCP3.4 with base value RCP2.6. The resulting gradients Δμ¯pΔμc (ΔE[μ~p]Δμc) shown in row 4 (10 and 16) may be viewed as “decarbonisation dividends” in terms of disaster risk reduction. For certainty-equivalent risk they range from 3.5% lower annual disaster likelihood under RCP8.5, to a 1.5% p.a. reduction under RCP3.4. For total expected disaster risk and baseline σ^c, the corresponding mitigation is greater even at the 2.5% end of the ΔE[μ~p]Δμc confidence intervals (3.7 and 1.9%, row 10), while for σcmin the declines are comparable (Property 1). The implication is that pursuing decarbonisation policies at decadal frequency pays short-term dividends by reducing annual conditional disaster likelihood, as emphasized by1. We also note the discrete gradients exceed the continuous response coefficients reported in Table 1, row 11. Focusing on Property 2, expected disaster risk tends to fall with the mean and variance of carbon stock growth. Figure 4b sheds light on this key result. The top row displays simulated disaster risk histograms consistent with each RCP trajectory to 2100 from Fig. 4a. It indicates the fitted beta pdf’s β parameter rising with the mitigation target; the historical incidence β from Fig. 1b stands between RCP8.5 and RCP6.0. The middle (lower) rows show smoothed kernel density functions of the expected disaster risk consistent with each pathway for baseline σ^c (σcmin) and Gaussian CO2 growth shocks. Moving across columns 1 to 4 highlights that E[μ~p] clearly falls with μc. Then, comparing rows 2 and 3 for a given scenario illustrates the uncertainty wedge narrowing as CO2 stock becomes more stable. For our stylized comparison, expected disaster risk is more sensitive to μc than σc; the incremental decline in total annual disaster risk starting from RCP8.5 is ∼3% across and <1% down, shrinking progressively for lower μc. Therefore, lowering σc will complement any reduction in E[μ~p]. It follows that RCP1.9, the only pathway limiting global warming to below 1.5∘C (the aspirational goal of the 2015 Paris Agreement) would also benefit from sharply lower disaster risk. Overall, the results suggest that relying solely on historical averages as best guesses for disaster incidence can lead to an underestimate of their expected stochastic frequency. This key analytical implication of convex transfer may carry significant implications for long-term climate policy as they could lead to more accurate short-term disaster projections (Property 1), and insofar as stabilizing CO2 concentrations may result in lower average disaster incidence in addition to decarbonisation policies (Property 2). More stable carbon emissions could result from avoiding boom-and-bust cycles, as recommended by Ref.14 and briefly witnessed during the COVID-19 pandemic26. Implementing such policies will likely require concerted and persistent collective action. In that respect, a key finding of Ref.33 is that optimal disaster risk adaptation requires government spending to reduce aggregate tail risk and restore the first-best solution. In turn, adaptation is most effective in mitigating risk under a central-planner equilibrium when households acquire knowledge about disaster risks, such as through public investment in early warning systems, rather than in an environment where no learning occurs. Similarly, Ref.52 recommend broadening the scope of carbon pricing schemes, encompassing a wider set of emissions sources and sinks, and extending carbon taxation to the consumption (downstream) stage. We anticipate that such expanded schemes—involving multiple stakeholders and necessitating offsetting public subsidies—have the potential to stabilize CO2 concentration levels. Concluding remarks There is great uncertainty involving the magnitude and pace of climate change. However, there is clear and urgent consensus that human-induced CO2 forcing is a prime driver of disaster events. Against this background, this paper argued that a convex risk transfer function presents policymakers and stakeholders with a strong incentive to collectively pursue more stable carbon emissions. This insight reinforces recent calls for reducing global emissions to protect humanity against irreversible long-term climate tipping risks5,13. We estimated a dynamic panel logit model linking the global disaster and atmospheric CO2 records from 1960 to 2022, and assessed the risk reduction potential of five IPCC decarbonisation scenarios, corresponding to sharply different climate policies up to 2100. The logit model’s AUROC scores were statistically significant and exceeded 0.85, indicating the model’s utility as a decision aid in one-year-ahead disaster prediction. Further, we established that uncertainty regarding the growth of CO2 concentration introduces a wedge between stochastic and certainty-equivalent disaster risks. This arises because disaster incidence is a convex function of CO2 growth, and its magnitude increases with the elasticity of logistic transfer and with the persistence of carbon concentration. Although most of our findings remain robust after controlling for the GAIN index and interactive fixed effects (IFEs), calibrating a transfer function that comprehensively accounts for socio- and techno-economic factors and exploring more heterogeneous effects across different regions would be a natural extension of the model’s application. Supplementary Information Supplementary Information 1. Supplementary Information 2. Supplementary Information 3. Acknowledgements We thank Glen Peters, Donald Robertson, Susan Solomon, and Xiyu Jiao for their helpful comments, and Nigel Stone for excellent research assistance. Author contributions D.T. initiated the research idea and conducted the Markov Chain simulations D.T. prepared Tables 3, 4, 5 and 6; Figs. 1b and 4; Appendix A S.L. prepared the disaster dataset and conducted panel logit analysis S.L. prepared Tables 1 and 2; Figs. 1a, 2 and 3; Appendix B All authors wrote the main manuscript text and reviewed the manuscript. Data availability The datasets used and/or analysed during the current study are included as supplementary files. Please contact Saite Lu (sl590@cam.ac.uk) if you have further questions. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-024-79437-5. References Associated Data Supplementary Materials Supplementary Information 1. Supplementary Information 2. Supplementary Information 3. Data Availability Statement The datasets used and/or analysed during the current study are included as supplementary files. Please contact Saite Lu (sl590@cam.ac.uk) if you have further questions.