Implications of regional variations in climate change vulnerability and mitigation behaviour for social-climate dynamics

Abstract

How regional heterogeneity in social and cultural processes drive–and respond to–climate dynamics is little studied. Here we present a coupled social-climate model stratified across five world regions and parameterised with geophysical, economic and social survey data. We find that support for mitigation evolves in a highly variable fashion across regions, according to socio-economics, climate vulnerability, and feedback from changing temperatures. Social learning and social norms can amplify existing sentiment about mitigation, leading to better or worse global warming outcomes depending on the region. Moreover, mitigation in one region, as mediated by temperature dynamics, can influence other regions to act, or just sit back, thus driving cross-regional heterogeneity in mitigation opinions. Under high emissions scenarios, the peak temperature anomaly varies by several degrees Celsius depending on how these interactions unfold. Our model exemplifies an exploratory framework for studying how global geophysical processes interact with population-scale concerns to determine future sustainability outcomes.

Many climate models now include economic variables, and this study shows how also integrating powerful social and cultural forces can dramatically alter predictions of both social and climate dynamics, and their diverse manifestation in world regions.

Introduction

Anthropogenic climate change is underway and its myriad and growing effects range from frequent and severe weather events¹, to increased disease transmission², as well as conflict, violence and human migration³. The impacts will extend beyond humans to many varieties of ecosystems and their biodiversity^1,4–6. It will take a combination of both mitigation and adaptation to address the climate change crisis⁷.

The Intergovernmental Panel on Climate Change (IPCC), whose reports are used to inform policy worldwide, has curated results from a substantial number of models studying climate change⁸. These include Integrated Assessment Models (IAMs)⁹, which are nuanced in their ability to account for a variety of socio-economic and geo-political scenarios, including projections of population, income, technology costs and future government policies (most recently in the form of the Shared Socioeconomic Pathways¹⁰). On the other hand, most such models lack mechanistic representations of human behavioural dynamics. Instead, they assume humans are homogeneous economic agents^11,12 (exceptions include agent-based representations of behaviour used in refs. ^13–15, for example). In order to capture the dynamic interplay between population mitigation behaviour and climate change, a mechanistic model of population behaviour–motivated not just by economic factors but also social processes–is required. Although economic costs are an important driver of the adoption of clean energy, behaviour (including social processes) are increasingly recognised as equally important¹⁶.

The limitations of IAMs and the need to model social behaviour in the context of climate change mitigation^17–20 have led to the development of coupled social-climate models that account for two-way feedback between social processes and climate change^21–24. These models have shown that behavioural and geophysical variables have comparable influence on climate projections^21,22. Social-climate models are an example of social-ecological or coupled human-and-natural system models. In other types of natural systems, social-ecological interactions result in novel system states that are not found in ecological or social models in isolation from each other^25–30, suggesting a valuable role for coupled social-climate models.

The social processes that are relevant to coupled social-climate models are numerous and varied. Subjective individual preferences and lifestyle choices directly impact mitigation efforts, such as eating less meat³¹, using public transport³², or even decisions requiring substantial financial investments, such as installing solar panels³³. Individuals can also collectively influence policy through their votes or even through protests or other large-scale social movements (for instance, the Fridays for Future campaign³⁴). Furthermore, the success of policies can depend on the interaction between economic and social factors³⁵. Policy measures often require the support of individuals to be fully effective, as demonstrated during the COVID-19 pandemic, where control depended critically on individuals’ willingness to comply with government regulations^36,37.

Social norms are especially relevant^38–40. It has been found that when injunctive norms (social pressure) that promote sustainability clash with negative descriptive norms (the perception that others do not follow the behaviour), the effectiveness of injunctive norms diminishes⁴¹. On the other hand, social norms can influence behaviour even to the point that cost-saving environmentally friendly measures might not be adopted, if they conflict with personal or perceived social norms⁴². For example, in business travel culture, the perception of professional expectations and status associated with business air travel–even for short flights–can prevent the use of cleaner forms of transport⁴³. Similarly, although some socio-cultural contexts support plant-based diets³¹, other contexts support the consumption of costly beef products instead of less expensive vegetarian food⁴⁴. Hence, social processes can prevent the adoption of sustainable behaviour, even when that behaviour is less economically costly.

Social and cultural values and their influence on behaviour vary across the world^45,46, as do perceptions of climate change and support for mitigation^47–49, but this heterogeneity has so far not been represented in social-climate models. Here, we construct a coupled social-climate model that represents the world population broken down into the five regions defined in the Shared Socioeconomic Pathways⁵⁰: Asia (ASIA), Latin America (LAM), the Middle East and Africa (MAF), the Organization for Economic Co-operation and Development (OECD) countries and the Reforming Economies of Eastern Europe and the former Soviet Union (REF) (Fig. 1a). Our treatment of social dynamics uses the replicator equation from evolutionary game theory^51–53, accounting for perceived climate change severity, learning processes, injunctive social norms⁵⁴ and economic costs. The social model represents dynamic decision-making processes regarding support for climate change mitigation. The model output determines the extent of climate change mitigation, which then determines anthropogenic carbon emissions. These carbon emissions are treated as input to the climate component of the model, which churns out a temperature anomaly (relative to pre-industrial levels). The temperature anomaly influences human perception of the severity of climate change, which once again affects social dynamics and the population opinion on mitigation. Through this, we are able to capture regional differences in some of the factors that influence mitigation support. These include diverse social and cultural settings^{45,47–49,55}, and differences in vulnerability to climate change caused by variations in exposure to extreme weather and socioeconomic circumstances^56–60. Understanding mitigation at the regional level is also useful because global agreements on mitigation targets are influenced by interactions at this scale^61,62. Our objectives are to uncover aspects of social-climate dynamics that are unique to models with regional structure, and to provide a framework for developing geographically structured coupled social-climate models.

Fig. 1. Model diagram and calibrated scenarios: baseline, best case and worst case.

a In our model, carbon emissions from five world regions, Asia (ASIA), Latin America (LAM), the Middle East and Africa (MAF), the Organization for Economic Co-operation and Development (OECD) countries and the Reforming Economies of Eastern Europe and the former Soviet Union (REF)⁵⁰, are summed and fed into the earth system model which computes the temperature, which then goes on to affect mitigation in each region. b Model dynamics generate the anomaly in the global mean surface temperature (T(t)) (in °C, relative to pre-industrial levels), under the best case (blue, calibrated to Representative Concentration Pathway⁸⁸ (RCP) 2.6 emissions), worst case (red, calibrated to RCP 8.5 emissions) and baseline (yellow, calibrated to RCP 4.5 emissions) scenarios, with socioeconomic projections following Shared Socioeconomic Pathway SSP2¹⁰ for all three. Region-specific social dynamics are not shown here. The solid line shows median trajectories for the filtered parameter values, with 0.25 and 0.75 quantiles represented by the shading.

Results

We use our model to explore the evolution of support for mitigation across multiple regions, and its impact on the global temperature anomaly. The model simulations explore how several crucial parameters and parameter combinations influence the proportion of mitigation supporters (‘mitigators’) in region i at time t, x_i(t), and the mean global surface temperature anomaly at time t, T(t). Our simulations reveal significant regional heterogeneity, with population opinion about mitigation evolving in a highly variable fashion across regions and scenarios. This occurs because differing social and economic circumstances generate different responses to the temporally evolving global temperature anomaly. Social learning has divergent impacts on public support for mitigation, depending on perceived cost of climate impacts and the strength of social norms. Moreover, social dynamics in one region can affect other regions through the intermediary of temperature feedback, allowing the possibility that one region can get a ‘free ride’ on the mitigation efforts of other regions. We describe these results in more depth in the following subsections.

Characterising mitigation in each region

We first give an overview of social/economic inputs and the resulting social-climate regimes in each of the five regions. Economic projections show how per capita mitigation costs are expected to decrease over time in all regions, driven by projected declines in the costs of renewable energy (Fig. 2a)⁶³. The projected decrease is largest in the OECD and REF, suggesting that the cost savings from transitioning to renewable energy are considerable in these regions. However, the OECD and REF have both lower initial support for mitigation and lower perceived vulnerability to climate change, compared to ASIA, LAM and MAF (Table 1 and Fig. 2b), despite standing to gain more (in total per capita cost savings) than the other regions by switching to renewable energy. In all regions, the initial number of mitigators is less than 50%, meaning that majority-enforcing social norms will initially impede the spread of pro-mitigation opinion and must be outweighed by declining mitigation costs and rising global temperatures. We also estimated the strength of social norms from survey data⁴⁶ and found that the OECD has somewhat weaker norms than other regions (Table 1).

Fig. 2. Mitigation and climate impact costs, and three social-climate scenarios.

a Estimated per capita costs of renewable energy relative to fossil fuels over time (β(t)). b Estimated per capita percentage loss in income as a function of temperature anomaly (f(T)). c–g Simulated mitigation support over time (x(t)) for the best case (blue), baseline (yellow) and worst case (red) scenarios, with solid lines showing median trajectories for filtered parameters and shadows outlining 0.25 and 0.75 quantiles. Cost estimates and social-climate dynamics are shown for five regions⁵⁰, Asia (ASIA), Latin America (LAM), the Middle East and Africa (MAF), the Organization for Economic Co-operation and Development (OECD) countries and the Reforming Economies of Eastern Europe and the former Soviet Union (REF). Note that a, b are model inputs and c–g are model outputs. Mitigation trajectories in c–g correspond to temperature trajectories of Fig. 1.

Table 1.

Estimated initial mitigation support, vulnerability to climate impacts and social norm strength

Region	Initial fraction of mitigators (x0)	Vulnerability (v)	Strength of social norms (δ)
ASIA	0.76	4.74	0.71
LAM	0.73	4.58	0.77
MAF	0.72	5.34	0.80
OECD	0.60	2.64	0.66
REF	0.54	3.30	0.76

See ‘Methods’ for details. Parameter estimates for five regions, Asia (ASIA), Latin America (LAM), the Middle East and Africa (MAF), the Organization for Economic Co-operation and Development (OECD) countries and the Reforming Economies of Eastern Europe and the former Soviet Union (REF), defined in ref. ⁵⁰. Vulnerability (v) lies between 0 and 10, with 10 representing the highest possible vulnerability. The strength of norms (δ) is a number between 0 and 1, with larger values corresponding to stronger norms.

These social and economic differences condition the emergent social-climate dynamics in our model. Calibrating the model with these inputs for each region to the Representative Concentration Pathways (RCP) 4.5, 2.6 and 8.5 yields three social-climate scenarios: baseline, best case and worst case. In best case and baseline scenarios, mitigation support in ASIA, LAM and MAF grows faster than in the OECD and REF (Fig. 2c–e versus f, g). Under baseline conditions, mitigation achieves majority support in ASIA, LAM and MAF by about 2030, in contrast to the OECD and REF, which reach majority support by about 2060. In all regions, social norms eventually lock the regions into a strong pro-mitigation social norm (x = 1). However, in the OECD and REF, fewer initial mitigators and less vulnerability to climate change (compared to ASIA, LAM and MAF) mean that majority-enforcing social norms initially impede the growth in support for mitigation. In the worst case scenario, mitigation support dies out in all regions except ASIA. Social norms lock the other regions into an anti-mitigation state relatively soon, despite the availability of cost-saving renewable energy sources.

Social learning and norms affect mitigation

Individuals in all regions are subject to the same processes–such as social learning and norms–but we find these same processes can generate very different outcomes in different regions, both quantitatively and qualitatively. In this section, we explore how social-climate dynamics change as we vary the parameter values for the social learning rate (κ₀), the strength of social norms relative to the cost of mitigation (‘relative social norm strength’, δ₀), and the cost of climate change impacts relative to the cost of mitigation (‘relative climate impact’, c₀), compared to the best case, baseline and worst case scenarios. Our parameter planes show that mitigation support in ASIA, LAM and MAF is less easily perturbed by parameter changes than the OECD and REF (Figs. 3 and 4, a–c versus d, e).

Fig. 3. Regional responses to social learning and climate impact costs are heterogeneous.

a–e Percentage change in population support for mitigation in 2050 relative to baseline scenario in each region, as a function of the social learning rate (κ₀) and the relative climate impact cost (c₀) varied for each region separately, and f peak temperature anomaly as a function of κ₀ and c₀ varied for all regions simultaneously. The five regions considered are: Asia (ASIA), Latin America (LAM), the Middle East and Africa (MAF), the Organization for Economic Co-operation and Development (OECD) countries and the Reforming Economies of Eastern Europe and the former Soviet Union (REF)⁵⁰. Areas in blue are parameter combinations leading to an increase in 2050 mitigation support compared to baseline and those in red indicate a decrease. On each plane, ‘A’, ‘B’ and ‘C’ marks the best case, baseline and worst case scenarios. Higher social learning generally supports mitigation, except in OECD and REF when c₀ is low.

Fig. 4. Social norms can help or hinder support for mitigation.

a–e Percentage change in population support for mitigation in 2050 relative to baseline scenario in each region, as a function of the social learning rate (κ₀) and the relative norm strength (δ₀) varied for each region separately, and (f) peak temperature anomaly as a function of κ₀ and δ₀ varied for all regions simultaneously. The five regions considered are: Asia (ASIA), Latin America (LAM), the Middle East and Africa (MAF), the Organization for Economic Co-operation and Development (OECD) countries and the Reforming Economies of Eastern Europe and the former Soviet Union (REF)⁵⁰. Areas in blue are parameter combinations leading to an increase in 2050 mitigation support compared to baseline and those in red indicate a decrease. On each plane, ‘A’, ‘B’ and ‘C’ marks the best case, baseline, and worst case scenarios. Mitigation support is hindered across a wider region of parameter space for the OECD and REF than for ASIA, MAF and LAM. Also shown (g–i) is the median time evolution of mitigation support (dashed blue) and temperature anomaly (solid red) for parameter combinations ‘G’, ‘H’ and ‘I’ in (f), shadows outline 0.25 and 0.75 quantiles.

We also find that a higher social learning rate generally helps boost mitigation support by 2050. The higher the learning rate, the faster the strategy with higher utility spreads. Increased social learning can scale up support from 40% to 100% by 2050 in the OECD and REF if relative climate impact costs are high or norms weak (Figs. 3 and 4d, e). The corresponding gains for ASIA, LAM and MAF are more modest (Figs. 3 and 4, d, e versus a–c). This is because ASIA, LAM and MAF are already primed to support mitigation on account of social and economic circumstances, except in the worst case scenario, as already noted.

Thus, social learning usually helps increase mitigation support, but in some cases it can hurt. For instance, when climate impact costs are low, an increase in the social learning rate reduces support for mitigation in the OECD and REF (Fig. 3d, e). This can also occur for all five regions when the social norm strength is sufficiently high (Fig. 4a–e). This occurs because of the ‘lock in’ effect: the amplification of feedback from within the social system (via majority-enforcing norms, δ₀) can have almost as much of an impact as feedback from the climate system (via c₀) on global warming outcomes (Supplementary Fig. 1a). If initial support for mitigation is not very strong (x < 0.5), entrenched anti-mitigation norms make it difficult for pro-mitigation opinions to spread, which causes severe warming outcomes in extreme cases (Supplementary Fig. 1b, c). One way to avoid such outcomes is by stepping up the relative cost of climate impacts, as might be achieved through education campaigns (Supplementary Fig. 1a). However, we note that the parameter ranges for which social learning harms mitigation support are generally far from the our best case, baseline and worst case scenarios (‘A’, ‘B’, ‘C’ in the plots), so that for parameter values representing the real world, increasing social learning usually boosts mitigation support.

In all regions, a high level of initial mitigation support can help counteract future declines caused by low climate impact costs or strong norms (Supplementary Fig. 2f–o) and prevent severe warming outcomes (Supplementary Fig. 3a–c). We also observe that higher climate cost or weaker social norms both support mitigation opinion, but parameter planes for both parameters indicate some regimes of rapid change in opinion, while other regimes show diminishing returns (Supplementary Fig. 4a–e).

Social-climate feedbacks can result in complex behaviour. For instance, the parameter planes for the peak temperature anomaly show a ‘ridge’ in the lower left (Fig. 3f) or lower right (Fig. 4f), signalling a non-monotonic response of temperatures to changes in the social learning rate. This region corresponds to a lower temperature anomaly than for nearby values of the social learning rate. Current legally binding policies are expected to lead to around 2.6 °C in 2100⁶⁴ which lies between baseline and worst case outcomes, so we pick parameter set ‘H’ that approximates this in order to explore the effects of varying the learning rate. For instance, if the learning rate is high (point ‘G’ in Fig. 4f), initial pro or anti mitigation opinion spreads rapidly. This leads to rapid growth in mitigation support in ASIA, LAM and MAF, but support declines in the OECD and REF where prevailing opinion does not favour mitigation. Because of this, the temperature peaks at about 3.5 °C, despite the eventual resurgence of mitigation in the OECD and REF after 2100 (Fig. 4g). On the other hand, if the learning rate is low (point ‘I’ in Fig. 4f), populations in all five regions are too slow to change their ways, so the temperature peaks at about 3. 8 °C (Fig. 4i). However, moderate learning (on the ridge, point ‘H’) results in the lowest peak temperature of 2.8 °C because mitigation support grows fast enough in ASIA, LAM and MAF by the end of the century, but not so fast as to allow the rapid propagation of anti-mitigation sentiment in the OECD and REF (Fig. 4h). Note that, as the social learning rate grows from ‘I’ to ‘H’ to ‘G’, there is a growing gap between the trajectories of mitigation opinion in ASIA/LAM/MAF versus the OECD/REF (Fig. 4g–i).

Temperature-mediated interactions between regions

The social processes in our model are not just a function of socio-economic differences between isolated regions. Rather, social dynamics in all regions influence each another through the mediation of climate dynamics. For example, in the worst case scenario, increased climate impact cost and/or lower mitigation costs in ASIA stimulate an 80% increase in the cumulative mitigation effort in ASIA (blue areas in Fig. 5a). The resulting drop in greenhouse gas emissions allows LAM to relax its own mitigation effort by 40% (red areas in Fig. 5d). Altering the strength of norms in the ASIA has a similar effect on mitigation support in LAM (Fig. 5b, e). Temporal trajectories of temperature and support for mitigation illustrate the effect. When norms are weaker in ASIA (‘C’ in Fig. 5b, e), mitigation support spreads quickly from initially low levels, thus reducing the temperature anomaly but also slowing uptake of mitigation by LAM (Fig. 5c). But when norms are stronger in ASIA and initially serve to slow the spread of mitigation support (‘F’ in Fig. 5b, e), LAM is motivated by warming temperatures to shift to a pro-mitigation state more rapidly (Fig. 5f). Since regions can choose, in principle, to increase or decrease the cost of mitigation through taxes and subsidies, this effect could be considered as non-mitigating regions free-riding on the efforts of mitigating regions. Similar but weaker effects are observed under parameter variation in other regions (Supplementary Figs. 5, 6). Parameter planes and time series shown in Fig. 5 remain qualitatively unchanged when using standard SSP-RCP pairings to calibrate the model (Supplementary Fig. 7), suggesting that our findings are robust to alternate socioeconomic scenarios.

Fig. 5. Regional interactions mediated by climate dynamics can lead to free-riding.

Here, stronger mitigation by Asia (ASIA) leads to weaker mitigation by Latin America (LAM). Panels show the percentage change in cumulative mitigation effort in (a) ASIA and (d) LAM, as a function of the relative climate impact cost (c₀, vertical axis) and rate of decline of mitigation cost (horizontal axis) in ASIA. Cumulative mitigation is measured as the area bounded by the x(t) curve for 2020–2200 relative to the worst case scenario. A higher rate of decline of mitigation costs corresponds to lower mitigation costs, compared to estimates shown in Fig. 2a. Areas in blue (red) are parameter combinations leading to an increase (decrease) in cumulative mitigation effort. b, e show the same as (a, d), except the horizontal axis is the norm strength δ₀. Note that the colourbar scale here differs from other figures. Also shown is the median time evolution of mitigation support in ASIA and LAM (dashed blue) and temperature (solid red) for c marked points ‘C’ and (f) marked points ‘F’ from (b) and (e), with shadows outlining 0.25 and 0.75 quantiles.

Sensitivity analysis

Our baseline model does not allow direct interactions between regions, such as might correspond to international treaties or social and cultural influences. However, a model extension shows how mitigation in all regions can benefit from cross-region interactions (Supplementary Fig. 8a–f). Strong coupling between regions could bring the peak temperature anomaly down from 2.1 °C to 1.9 °C under baseline parameters and by over 5 °C in the worst case scenario (Supplementary Fig. 8h, i). If mitigators in one region are encouraged by mitigators in other regions, this can counteract entrenched social norms within their region that deter them from switching to a pro-mitigation opinion.

Baseline and best case projections are robust to parameterising norm strength using an alternate data source⁴⁹, with differences in mitigation dynamics altering temperature projections by approximately 0.1 °C under baseline conditions (see ‘Methods’, and Supplementary Fig. 9). However, in the worst case scenario, projections are substantially different in ASIA and the MAF. For the MAF, norms are considerably weaker under the alternate parameterisation, allowing mitigation support to increase sooner (Supplementary Table 1). This reduces the temperature in 2200 by over 1 °C (Supplementary Fig. 9r).

We also tested using an alternate data source to parameterise climate change impacts⁶⁰ and found mitigation was projected to grow faster in all regions, since projected climate impacts are greater (Supplementary Figs. 10, 14c). This lowers the temperature anomaly by about 0.5 °C under baseline projections (Supplementary Fig. 10q) and by up to 4 °C in 2200 under worst case conditions (Supplementary Fig. 10r), driven by substantially different mitigation patterns in ASIA, LAM and MAF compared to using damage estimates from⁵⁹ (Supplementary Fig. 10c, f, i). However it should be noted that there are differences in what the alternate data sources measure (see ‘Methods’), so differences in outcomes could be in part due to those.

Finally, projected time series for calibration to standard SSP-RCP pairings show similar dynamics to our base calibration (Supplementary Fig. 11), except in the worst case where ASIA shows slower mitigation uptake under SSP5 compared to our assumption of SSP2 socioeconomic change (Supplementary Fig. 11c).

Discussion

Much previous research on climate change mitigation makes the implicit assumption that socio-economic differences between populations can be taken as given, and are fixed over time. In contrast, previous coupled social-climate models allow for endogenous social dynamics in response to climate dynamics, but may be taken to suggest that the same social processes have similar effects in different populations. Our regionally-resolved social-climate model suggests a third perspective: the same social processes can generate different outcomes in different regions, and these outcomes hinge partially on social processes in other regions as mediated by their shared global climate dynamics.

We see two broad patterns of mitigation emerge from our model. First, in regions like the OECD and REF, where initial mitigation support and climate change impacts are low compared to other regions, increased communication in the form of social learning usually supports pro-mitigation opinions, but not always: in some parameter regimes, social learning in the presence of entrenched anti-mitigation social norms that penalise diverging from the mainstream opinion can impede the spread of pro-mitigation opinion, even when more climate mitigation would result in cost savings. The second pattern of mitigation is what we see in ASIA, for example, where initial mitigation support is higher and climate change impacts are greater than in other regions. In such regions, increased social learning rarely works against mitigation, since concern about climate impacts outweighs social norms that would otherwise suppress mitigation. We found that the differences between these two mitigation patterns are reinforced through feedback from global temperature dynamics. Also, regions that face greater impacts shift to a pro-mitigation opinion faster, causing regions that are less vulnerable to be even less inclined to mitigate. Put simply, the feedback from the global temperature anomaly can help drive regions apart, in terms of their stance toward climate change mitigation and how quickly they shift to a pro-mitigation state.

Our model suggests that mitigation progresses more rapidly in ASIA, LAM and MAF than it does in the OECD and REF, which is consistent with empirical research^65,66. We also find that mitigation support in ASIA, LAM and MAF is stable across a broader range of parameter values than in the OECD and REF. On the other hand, ASIA, LAM and MAF could face challenges to implementing climate change mitigation measures that are not explicitly captured in our model, such as less access to clean technologies^8,67–69, or trade-offs between mitigation and development^70,71.

Our findings suggest that increased public discourse can support the spread of opinions in favour of climate change mitigation⁷² but only when climate change damages are appropriately emphasised, and anti-mitigation social norms are not too firmly entrenched. More broadly, a focus on the content of information accessible to the public could help prevent the transmission of misinformation^73,74 and also decrease pessimism about climate action^49,75, both of which could help with timely mitigation. But even then, our model shows that some populations may be very resistant to change. In a related vein, we see that mitigation trends in one region can influence mitigation in other regions purely through feedback via global warming. This could result in regions free riding or, alternately, picking up the slack for others’ lack of mitigation. This emphasises the need for binding global-scale climate change treaties⁷⁶, although starting with national or regional level treaties may be a useful stepping stone⁶¹.

There are many dimensions that we do not account for in our modelling framework, and because the model lacks sufficient behavioural realism to inform policy directly, our results should be interpreted as an exploratory exercise rather than a predictive framework. Our modelling approach assumes that support for mitigation translates into mitigation action, and we do not distinguish between different kinds of mitigation action. Future work could model this distinction, separate their impacts on emissions, and use data on correlations between different mitigation actions⁴⁹. Individuals in our model learn from others, but we do not explicitly account for information from large-scale entities such as governments or the media affecting social norms⁷⁷. Our assumption that social dynamics are local to each region does not account for the role of smaller scale social groups and networks, which may serve to break up an otherwise homogeneous and region-wide pro- or anti-mitigation norm. Our region-wide approach fails to account for marginalised, vulnerable groups, such as Indigenous populations in developed regions like the OECD⁸. Similarly, we do not include cross-regional connections through flow of information or people⁷⁸. Future work could study the role of cross-border migration attributable to climate change, and its demographic and social impacts⁷⁹. Finally, our description of social learning is very simple, and there are multiple approaches (and data sources) that could be used to parameterise social norms, which result in different parameter estimates. However given the important role it appears to play, future work might benefit from using empirical data to estimate the social learning rate, and fitting models to longitudinal survey data on mitigation support, or data on electric vehicle or renewable technology adoption as alternate measures of mitigation support. Similarly, the perception gap between actual and perceived support for climate action is important⁴⁹, and is modelled only imperfectly by a low social learning rate in our model. Many of these limitations could be remedied by building models that represent social dynamics at finer scales, such as national or sub-national. Likewise, our model implicitly assumes that social processes are sufficient to determine emission trajectories, but this assumes governments always respond to their population’s opinion about climate change. Dynamics could change if international negotiations and nonlinear feedback between different scales of social and/or political dynamics are modelled.

These limitations suggest future directions for social-climate modelling that could improve their validity and analytical power. Incorporating social dynamics into IAMs could be especially useful. The trajectory of model sophistication often increases over time in many fields: the earliest published climate models were simple one-dimensional radiative convective models, for instance⁸⁰. Social-climate models could exhibit a similar trajectory. However, simple models also have their advantages, such as transparency, ease of analysis and the ability to generate insights or suggest new hypotheses. Our regional-level model shows the crucial role that social processes and regional heterogeneity play in determining climate trajectories. It also exemplifies a framework that could be used to study how global geophysical processes interact with human-scale concerns–economic well-being, equity, and social norms–to determine future sustainability outcomes.

Methods

Inputs and key mechanisms

Individuals in each region decide whether or not to support climate change mitigation. The decision is based on whether they perceive a net benefit (‘utility’) in supporting mitigation. Three inputs determine this: (1) the net mitigation cost (based on the estimated costs of renewable energy relative to fossil fuels for the region), (2) the cost of climate change impacts, which depends on the current global temperature and the region’s vulnerability, and (3) social norms that reinforce the region’s current majority opinion through social pressure. Summing these three costs determines whether an opinion that supports mitigation is desirable. Another input governs the rate at which individuals interact with others in the region (the social learning rate). If an individual interacts with someone who has a different opinion, they switch to the other person’s opinion if the net utility of that opinion is higher. The higher the net utility gain, the more likely they are to change opinion. A further input is the starting proportion of pro- and anti-mitigation individuals in the region. Our model assumes that each person is equally likely to interact with every other person within a region, and within a region, socio-economic conditions are uniform. To represent climate dynamics in our coupled model, we adapt an existing earth system model⁸¹. The model describes the carbon cycle, representing dynamics in the atmosphere, ocean, vegetation and soil, as well as changes in the global mean surface temperature, as influenced by the quantity of atmospheric carbon.

The model captures key mechanisms of social-climate feedback. A region’s carbon emissions are driven by the proportion of the population that supports mitigation. The carbon emissions from all five regions flow into the atmospheric carbon compartment of the Earth system model. The resulting impact on global surface temperature subsequently influences the perceived cost of climate impacts in each region. Changing global temperatures thus cause the region’s opinion composition to evolve over time, along with changing net mitigation costs and the landscape of social norms in the region. Higher global temperatures always increase mitigation support in a negative feedback loop. But this takes time since the global temperature responds slowly to carbon emissions. Lower net costs of mitigation (due to ever-cheaper renewable sources of energy) also increase mitigation support, which operates more quickly than the temperature feedback. In contrast, majority-enforcing social norms are a positive feedback loop that can boost mitigation support or suppress it, depending on the initial opinion composition of the population. If more than half the population supports mitigation, norms encourage the support of mitigation, otherwise they discourage it. If social pressure is strong enough, a population can become locked into a pure pro- or anti-mitigation state, regardless of the net mitigation cost or perceived climate impacts. Social norms can interact with other model processes in nontrivial ways. For example, if global temperatures rise fast enough to shift pro-mitigation to the majority opinion, then pro-mitigation social norms will ‘kick in’ through a positive feedback loop to lock in the population at a new pro-mitigation social norm. Finally, the social learning rate regulates the switching rate between opinion states. For a very low social learning rate, regions are slow to adapt to changing climate impacts, mitigation cost, or their region’s social norms.

Social dynamics

If individuals derive utility U_M from mitigating, and utility U_N from not mitigating, then the net utility gained for switching from non-mitigating to mitigating behaviour is U_M − U_N. If this quantity is positive, then mitigation support, x, will grow, else it will decrease. The specific form we assume for the net utility is presented as equation (1).

$$U_M-U_N=-\beta +f\left(T\right)+\delta \left(2x-1\right)$$ 1

β represents the net cost associated with switching to mitigating behaviour, based on the cost of renewable energy relative to fossil fuels, f(T) is the cost of climate change impacts, and δ represents the strength of social norms (in other words, the strength of social pressure that a person feels to conform to the prevailing majority opinion). The norm strength δ is multiplied by (2x − 1), so that when mitigation support is in a minority (that is, x < 0.5), this term takes away from the net utility of mitigation, and otherwise it supports it. All costs in equation (1) are perceived costs, so they are informed partly by data, but also have margins built in to account for uncertainty resulting from incomplete information. Individuals learn behaviour through interactions, which we assume take place at the social learning rate, κ, and the probability of a mitigator x encountering a non-mitigator 1 − x is given by κx(1 − x). The rate of change of the proportion of mitigators x with time is expressed as equation (2)

$$\frac{dx}{dt}=\kappa x\left(1-x\right)\left[-\beta +f\left(T\right)+\delta \left(2x-1\right)\right]$$ 2

which has the same form as the replicator equation of evolutionary game theory⁸². It is worth noting that this functional form and especially the frequency term x(1 − x) could also be taken as a representation of descriptive social norms, since it is through interactions that individuals are able to gauge the average opinion in the population.

We assume that social dynamics are specific to five world regions: Asia (ASIA), Latin America (LAM), the Middle East and Africa (MAF), the Organization for Economic Co-operation and Development (OECD) countries and the Reforming Economies of Eastern Europe and the former Soviet Union (REF). These regions are as defined in the Shared Socioeconomic Pathways (SSPs) five region aggregation⁵⁰. We use region-specific empirical estimates for β, f(T) and δ in equation (2), but we need to express these three components in the same utility measure that individuals use to make decisions. We introduce scalings $\widehat{\beta }$, $\widehat{c}$ and $\widehat{\delta }$ so that individuals can react to similar costs in different ways according to their context. This reduces the number of parameters in order to minimise overfitting and facilitate exploration of parameter space. Unscaled, β has dimensions of a cost per person, f(T) of a fraction of income, and δ, of a dimensionless quantity between 0 and 1. We divide β by per capita income (using SSP2 forecasts across all scenarios to control for socioeconomic variation) so this can be interpreted as a fraction of income and then re-write equation (2) as equation (5),

$$\frac{dx}{dt}=\kappa x\left(1-x\right)\left[-\widehat{\beta }\beta \left(t\right)+\widehat{c}f\left(T\right)+\widehat{\delta }\delta \left(2x-1\right)\right]$$ 3

$$=\kappa \widehat{\beta }x\left(1-x\right)\left[-\beta \left(t\right)+\frac{\widehat{c}}{\widehat{\beta }}f\left(T\right)+\frac{\widehat{\delta }}{\widehat{\beta }}\delta \left(2x-1\right)\right]$$ 4

$$={\kappa }_{0}x\left(1-x\right)\left[-\beta \left(t\right)+c_{0}f\left(T\right)+{\delta }_0\delta \left(2x-1\right)\right],$$ 5

where ${\kappa }_0=\kappa \widehat{\beta }$, $c_0=\frac{\widehat{c}}{\widehat{\beta }}$, and ${\delta }_0=\frac{\widehat{\delta }}{\widehat{\beta }}$.

The equation for social dynamics for region i is then expressed as equation (6)

$$\frac{d{x}_i}{dt}={\kappa }_{0_i}{x}_i\left(1-x_i\right)\left[-{\beta }_i\left(t\right)+c_{0_i}{f}_i\left(T\right)+{\delta }_{0_i}{\delta }_i\left(2x_i-1\right)\right].$$ 6

Cross-region interaction

In our base model, we do not account for direct interactions between regions, although all regions are coupled indirectly through their climate change impacts. For a sensitivity analysis, we introduce cross-regional influence of social dynamics by modifying equation (6) to equation (7),

$$\frac{d{x}_i}{dt}={\kappa }_{0_i}{x}_i\left(1-x_i\right)\left[-{\beta }_i\left(t\right)+c_{0_i}{f}_i\left(T\right)+{\delta }_{0_i}{\delta }_i\left(2x_i-1\right)+\alpha \frac{{\sum }_{j\ne i}{G}_j{x}_j}{{\sum }_{j\ne i}{G}_j}\right]$$ 7

where x_j is the fraction of mitigators in region j, G_j the GDP of region j, and α the coupling strength governing the magnitude of influence of other regions. Through this modification, mitigation in other regions encourages mitigation in a given region, weighed by the economic power of the remaining regions.

Coupled model

Our coupled social-climate system is represented by equations (8)-(13).

$$\frac{d{x}_i}{dt}={\kappa }_{0_i}{x}_i\left(1-x_i\right)\left[-{\beta }_i\left(t\right)+c_{0_i}{f}_i\left(T\right)+{\delta }_{0_i}{\delta }_i\left(2x_i-1\right)\right]$$ 8

$$\frac{d{C}_{at}}{dt}={\sum }_i\left(1-x_i\right){ϵ}_i\left(t\right)-P\left(C_{at},T\right)+R_{veg}\left(T,C_{veg}\right)+R_{so}\left(T,C_{so}\right)-F_{oc}\left(T,C_{at},C_{oc}\right)$$ 9

$$\frac{d{C}_{oc}}{dt}=F_{oc}\left(T,C_{at},C_{oc}\right)$$ 10

$$\frac{d{C}_{veg}}{dt}=P\left(C_{at},T\right)-R_{veg}\left(T,C_{veg}\right)-L\left(C_{veg}\right)$$ 11

$$\frac{d{C}_{so}}{dt}=L\left(C_{veg}\right)-R_{so}\left(T,C_{so}\right)$$ 12

$$\frac{dT}{dt}=\frac{a_E\left[\left(F_d\left(C_{at},T\right)\right)-\sigma {\left(T+T_0\right)}^4\right]}{C}$$ 13

The index i runs over each of the five regions. Equation (8) describes the time evolution of mitigation support in region i. Equations (9)–(12) represent a box model of the carbon cycle⁸¹, simplified to represent the ocean carbon sink by a single equation⁸³. The anomaly in the global mean surface temperature is calculated in equation (13), where the net downward solar radiation, F_d is influenced by the quantity of carbon dioxide in the atmosphere, which is the balance of carbon emissions generated by anthropogenic and natural sources and absorbed by land and ocean carbon sinks. The land sink consists of vegetation and soil, and we assume a single ocean sink, not distinguishing between dynamics at different depths and latitudes²². Details of the functions used in these equations are provided below.

Equation (9) describes the change in atmospheric carbon with time. The first term in the equation represents anthropogenic carbon emissions summed over all regions, with regional emissions ϵ_i(t), weighed by the proportion of non-mitigators, (1 − x_i). The emission rate ϵ_i(t) represents the worst case annual carbon emissions (in the absence of population-driven mitigation). Social dynamics begin evolving in our simulations in 2020. Prior to 2020, ϵ(t) is fixed according to global historical emissions⁸⁴. Thereafter, ϵ_i(t) is given by equation (14).

$${ϵ}_i\left(t\right)={ϵ}_{i_{2020}}+\frac{\left(t-2020\right){ϵ}_{i_{\max }}}{\left(t-2020\right)+s}$$ 14

${ϵ}_{i_{2020}}$ represents carbon emissions in 2020, s is the half-saturating constant, and ${ϵ}_{i_{\max }}$ the limiting value for ϵ_i(t), which follows an increasing but saturating trajectory with time (Supplementary Fig. 12a). In addition to anthropogenic emissions, vegetation and soil respiration (R_veg(T, C_veg) and R_so(T, C_so), respectively) contribute to atmospheric carbon. Thus these terms are added to equation (9) and subtracted from equations (11) and (12), respectively. Carbon emitted through respiration increases with the biomass contained in vegetation and soil (C_veg and C_so, respectively). Respiration also increases with temperature, following an Arrhenius functional form. Vegetation and soil respiration are described by equations (15) and (16), respectively,

$$R_{veg}\left(T,C_{veg}\right)=k_r{C}_{veg}{k}_A{e}^{-\frac{E_a}{R\left(T+T_0\right)}}$$ 15

$$R_{so}\left(T,C_{so}\right)=k_{sr}{C}_{so}{k}_B{e}^{-\frac{308.56}{T+T_0-227.13}}$$ 16

where k_r and k_sr are rate constants, and k_A and k_B normalising constants. E_a denotes activation energy, and R, the universal gas constant.

Carbon is transferred from the atmosphere to the land sink via photosynthesis of vegetation, P(C_at, T). The loss of carbon from the atmosphere is represented by the second term in equation (9) and the absorption of this into vegetation by the first term of equation (11). Photosynthesis has the functional form presented in equation (17).

$$P\left(C_{at},T\right)=\left\{\begin{array}{cc}{k}_p{C}_{ve0}{k}_{MM}\left(\frac{p{\mathrm{CO}}_{2a}-k_c}{K_M+p{\mathrm{CO}}_{2a}-k_c}\right)\left(\frac{{\left(15+T\right)}^2\left(25-T\right)}{5625}\right)\hfill & \mathrm{when}p{\mathrm{CO}}_{2a}\ge k_c\mathrm{and}-15\le T\le 25\hfill \\ 0\hfill & \mathrm{otherwise}\hfill \end{array}\right)$$ 17

This follows Michaelis-Menten kinetics, with optimal photosynthesis at T = 2, and half-saturation constant K_M. Other constants in the function are k_c, the CO₂ compensation point at and below which photosynthesis does not consume carbon, k_p, a rate constant, and C_ve0 the pre-industrial quantity of carbon in vegetation. The mixing ratio pCO_2a is the ratio of moles of CO₂ in the atmosphere to k_a, the total moles of molecules in the atmosphere, shown in equation (18),

$$p{\mathrm{CO}}_{2a}=\frac{f_{gtm}\left(C_{at}+C_{at0}\right)}{k_a}$$ 18

with f_gtm as the conversion factor used to convert carbon from GtC to moles of carbon.

Within the land sink, carbon is transferred from vegetation to the soil upon plant decay, given by equation (19).

$$L\left(C_{veg}\right)=k_t{C}_{veg}$$ 19

This quantity scales with the biomass of vegetation, C_veg, governed by rate constant, k_t. Thus the soil component of the land sink receives carbon through vegetation, L_veg and releases carbon to the atmosphere through soil respiration, R_so, described in equation (12).

The final term in equation (9) represents absorption of carbon by the ocean sink in terms of the ocean flux, F_oc(C_at, C_oc). Equation (10)describes this flux as entering the ocean from the atmosphere, with the flux formulated as equation (20).

$$F_{oc}\left(C_{at},C_{oc}\right)=F_0\chi \left(C_{at}-\zeta \frac{C_{at0}}{C_{oc0}}{C}_{oc}\right)$$ 20

χ is the characteristic solubility of CO₂ in water and ζ is the evasion factor. The amount of carbon absorbed by the ocean is a function of existing levels of atmospheric carbon and ocean carbon. Notably, there is no direct transfer of carbon between the ocean and land sinks.

Finally, equation (13) calculates changes in the global mean surface temperature assuming a grey-atmosphere approximation, that is, assuming radiation of all wavelengths are equally absorbed by various greenhouse gases. The equation represents an energy balance model of the earth assuming blackbody radiation, with a_E representing the earth’s surface area, σ the Stefan-Boltzmann constant and C the specific heat capacity of the earth’s surface. The net downward flux of of solar radiation to the earth’s surface, F_d, accounts for albedo, A, solar flux, S, and changes in the opacities of greenhouse gases, specifically of carbon dioxide CO₂ and water vapour, H₂O_v (the opacity of methane, CH₄, is assumed not to change), and is presented in equation (21).

$$F_d=\frac{\left(1-A\right)S}{4}\left(1+\frac{3}{4}\tau \right)$$ 21

τ is the vertical opacity of the greenhouse atmosphere, and the opacities of specific greenhouse gases are given by equations (22)-(24),

$$\tau \left({\mathrm{CO}}_2\right)=1.73{\left(p{\mathrm{CO}}_2\right)}^{0.263}$$ 22

$$\tau \left({\mathrm{H}}_2\mathrm{O}\right)=0.0126{\left(H{P}_0{e}^{-\left(L/RT\right)}\right)}^{0.503}$$ 23

$$\tau \left({\mathrm{CH}}_4\right)=0.0231$$ 24

where H is the relative humidity, P₀ is the water vapour saturation constant, L is the latent heat per mole of water and R is the molar gas constant. The opacity of CO₂ scales with the quantity of atmospheric CO₂, and that of water vapour H₂O_v with temperature, T. Descriptions of and values for climate model parameters used in equations (9)–(13) are in Supplementary Table 2.

Net cost to mitigate

The net cost to mitigate, β(t), is based on empirical estimates of per capita costs of renewable energy, relative to a constant fossil fuel cost⁶³, multiplied by per capita energy demand projections⁵⁰. We use the per capita costs of the least expensive form of renewable energy available in the market which, historically, has been wind energy. However, projections show solar energy costs falling below those of wind energy around the year 2028⁶³; we account for this shift by using the lower of the two costs through time in our calculations. We use costs under a ‘fast transition’ scenario (Supplementary Fig. 13a), which assumes wind and solar energy use grow at rates at their historical rates (costs are assumed to decrease with increases in production for a given technology)⁶³. (All costs are measured as a global average of the Levelized Cost of Electricity (LCOE), in $[2020]/MWh). From these costs, we subtract a fossil fuel cost, k, which is assumed to remain constant at $100/MWh over the simulation time frame⁶³. We multiply the difference of renewable and fossil fuel costs by region-specific per capita energy demand forecasts under the SSP2 scenario⁵⁰ (Supplementary Fig. 13b), to get estimates for region-specific mitigation costs. Renewable energy costs less than fossil fuels in all regions from the outset in the simulation of social-climate dynamics.

Cost of climate change impacts

We assume that the cost f(T) associated with climate change impacts is a function of T(t), the current anomaly in the global mean surface temperature. The functional form for f(T) is based on estimated loss in global GDP⁵⁹, scaled by region-specific vulnerability v_i to climate change impacts⁵⁶. Specifically, we use estimates for the mean global GDP impact under ‘annual temperature only’ in Fig. 4a of ref. ⁵⁹. To distribute the global GDP impact over the regions in our model, we scale impacts using the INFORM Climate Change Risk Index scores⁵⁶, using baseline 2022 risk scores for each country to calculate population-weighed average risk scores for each region (‘vulnerability’, v_i), lying on a scale of 0 to 10, where 10 denotes the greatest possible vulnerability (estimates in Table 1). We assume that a global temperature anomaly of T degrees impacts the GDP of region i by f_i(T) percent, expressed as equation (25),

$$f_i\left(T\right)=s\left(T\right)\frac{v_i}{10}$$ 25

where s(T) is a function representing the response of GDP to temperature that is common to all regions, and v_i denotes the vulnerability of region i to climate change impacts. The vulnerability term indirectly captures location-specific climate change and also factors like socioeconomic vulnerability. We calculate point estimates for the function s(T) by solving equation (26),

$$l\left(T\right)={\sum }_i{p}_{i}s\left(T\right)\frac{v_i}{10}$$ 26

where l(T) represents the mean global GDP impact at T degrees warming for T = 0.84, 1.5, 2 and 3 °C from ref. ⁵⁹ and p_i is proportion of the world population in region i. We then fit function $s\left(T\right)=-100\left(a\exp \left(-bT\right)-c\right)$ to point estimates obtained by solving equation (26) using MATLAB’s lsqnonlin package (Supplementary Table 3 for fitted coefficients a, b and c, Supplementary Fig. 14a for fitted function s(T)).

To test robustness, we use impacts from ref. ⁶⁰ to fit our climate impact function. We use the method described above to estimate f_i(T), now using damage estimates for global GDP per capita impacts in 2100 under RCP 8.5 warming under ‘base’ socioeconomic growth from ref. ⁶⁰ for point estimates for s(T) (Supplementary Table 3 for fitted coefficients, Supplementary Fig. 14 for fitted function). We use estimates from the ‘short-run pooled’ impacts, to keep assumptions similar to those made in ref. ⁵⁹. Despite this, it should be noted that we cannot make an apples-to-apples comparison for the two estimates. Estimates from ref. ⁵⁹ are calculated across different Coupled Model Intercomparison Project (CMIP) scenarios, so these temperatures may occur in different years under different growth scenarios. However, the data from ref. ⁶⁰ is from the ‘base’ scenario, assuming that GDP grows at the average rate between 1980 and 2010, and temperature increases according to RCP 8.5. Further, the estimates from ref. ⁵⁹ assume 0% impacts at 0.84 °C warming, while⁶⁰ assumes 0% impacts at 0.8 °C warming. Thus ref. ⁶⁰, will tend to slightly overestimate impacts, relative ref. ⁵⁹. Finally, we note that the data from ref. ⁶⁰ measures change in GDP per capita rather than GDP, however this does not affect estimates since a percent change will remain the same across both measures.

Strength of social norms

To estimate the strength of social norms, δ(t), we use survey data on perceptions of the importance of conforming to societal expectations⁸⁵, so δ represents the pressure an individual feels to conform to majority opinion. We use data from the World Values Survey (WVS 6, 2010–2014), which asks respondents of 60 countries their perception of the need to conform to societal expectations⁴⁶. WVS 6 is the most recent iteration of the survey that asks this question. Question V77 in the survey asks people to rate on a scale from 1 to 6 how important they feel it is for them to ‘behave properly; to avoid doing anything people say is wrong’, where 1 codes for this description being ‘very much like me’ and 6 codes for this being ‘not at all like me’⁴⁶. Flipping the numbers, so that 1 codes for ‘not at all like me’ and 6 codes for ‘very much like me’, we calculate a weighted mean score for each country, using the proportion of respondents with a given response as weights. This average score does not provide a meaningful description in terms of the Likert scale, but it gives an idea of the relative positioning of different countries on the basis of their responses; we would expect a country with a higher score to have stronger norms than a country with a lower score. We calculate population-weighed averages across the countries for which there is data to get regional scores between 1 and 6 for each region, which we then normalise to lie between 0 and 1. Estimated norm strengths are in Table 1.

We use survey data from ref. ⁴⁹ to perform a sensitivity analysis. Their survey covers 125 countries and was conducted in 2021/2022. We calculate the average of responses on two questions. First, the question that asks if the respondents believe if people on their country ‘should try to fight global warming’, with 1 coding to ‘yes’ and 0 coding to ‘no’, and second, the question asking people about their belief of what fraction of other people in their country are willing to contribute 1% of their income to fight global warming. We calculate population-weighted average scores for each region in our model (Supplementary Table 1). Note that this data measures of perceptions of mitigation support and beliefs about what other people should do, rather than the pressure to conform to societal expectations captured by the original parameterisation.

Initial fraction of mitigators

We use data from ref. ⁴⁹ to estimate the initial fraction of population supporting mitigation (x₀) in each region by averaging survey responses on whether people are willing to contribute 1% of their income to combat global warming (1 codes to ‘yes’ and 0 ‘codes’ to no), and if not, would they be willing to contribute <1% of their income to combat global warming (0.5 for ‘yes’), for each country. We then calculate population-weighted average scores for each region in our model (Table 1).

Energy demand

Energy demand data and forecasts⁵⁰ are used to estimate regional proportions of the total energy demand in 2020 and 2100. These proportions are then used to scale ϵ₂₀₂₀ and ${ϵ}_{\max }$, respectively, creating region-specific worst case carbon emissions trajectories, ϵ_i(t) (Supplementary Fig. 12b).

Calibration

The estimated ranges for ${\widehat{x}}_0$, κ₀, c₀ and δ₀ are chosen to include plausible values based on empirical evidence where available, and also have generous margins beyond these. We allow climate impact costs to range between half of mitigation costs to 15 times mitigation costs (based on supporting evidence⁸⁶), and assume that norms are of the same order of magnitude as mitigation costs. We also multiply the initial condition, x₀ by a scaling constant ${\widehat{x}}_0$. Ranges for these parameters are in Supplementary Table 4.

We used Approximate Bayesian Computation (ABC)⁸⁷ to calibrate ${\widehat{x}}_0$, κ₀, c₀, δ₀ to create baseline, best case, and worst case scenarios, assuming a uniform prior, selecting 2500 of 50,000 parameter sets with the least error. For our baseline scenario, the modelled carbon emissions trajectory was calibrated to the RCP 4.5 scenario, thereby generating credible intervals for the remaining parameters. For our best case scenario, we similarly calibrated the model to the RCP 2.6 scenario, and for the worst case scenario we used RCP 8.5⁸⁸ (Supplementary Table 4 for calibrated parameter point estimates and ranges). Note that the model does not converge to a steady state on the simulation time scale, since we calibrated the model to RCP scenarios. Distributions of filtered parameters for each scenario are in Supplementary Fig. 15.

One way of understanding the impact of regional socio-economic differences on mitigation pathways is outlined in the various SSP scenarios, each of which is paired with one of the RCP emissions scenarios. Here, since our goal is to understand the impact of social processes on social-climate dynamics, we adopt a different approach of using SSP2 for three different RCP scenarios, and assessing the impact of social processes by varying social parameters of the model. This allows us to isolate the effect of social processes while controlling for other variables. Other factors extrinsic to the model, such as economic and demographic changes, are based on ‘middle-of-the-road’ SSP2 projections for which data are available¹⁰, else, under assumptions closest to those of SSP2. If we were to use the standard SSP-RCP pairings, it would be difficult to know whether differences between our scenarios were attributable to social, demographic, or economic factors. We conduct a sensitivity study with the standard pairings, however. We use per capita income and energy demand forecasts from SSP1 with RCP 2.6 for the best case, and from SSP5 with RCP 8.5 for the worst case scenarios. Baseline scenario calibration continues to use SSP2 with RCP 4.5. Calibrated parameter estimates are in Supplementary Table 5.

Simulation

Our coupled model was simulated in MATLAB R2021a, using the ode45 module. Social dynamics were initiated in the year 2020, and run until 2200. Empirical estimates for renewable energy costs and energy demand give us information until 2100; for simulations beyond these, we extrapolate historical functions.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

C.T.B. and M.A. conceived the study. A.P., T.M.L. and C.T.B. developed the methodology. A.P. performed the analysis and wrote the first draft. A.P., T.M.L., C.T.B. and M.A. revised and commented on the manuscript.

Peer review

Peer review information

Nature Communications thanks the anonymous, reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

Data for the Shared Socioeconomic Pathways is available at https://tntcat.iiasa.ac.at/SspDband data for the Representative Concentration Pathways is available at https://tntcat.iiasa.ac.at/RcpDb. Country boundaries data used to generate the map in Fig. 1a are from https://www.naturalearthdata.com/downloads/10m-cultural-vectors/10m-admin-0-countries/. Historical carbon emissions data is available at https://zenodo.org/records/15016289. Energy cost forecast data are available upon request to the authors of 10.1016/j.joule.2022.08.009. Data on the effects of climate change on income are from Fig. 4a of ref. ⁵⁹ at 10.1038/s41558-024-01990-8, and data from ref. ⁶⁰ are generated using replication code at https://purl.stanford.edu/wb587wt4560/version/1. Data on vulnerability to climate impacts is from the INFORM Climate Change Risk Index, available at https://drmkc.jrc.ec.europa.eu/inform-index/INFORM-Climate-Change/Results-and-data. Survey data on the need to conform to societal expectations from the World Values Survey Wave 6 is available at 10.14281/18241.23. Survey data on public support for and perceptions of global warming can be downloaded from https://dataverse.iza.org/file.xhtml?fileId=234&version=5.0.

Code availability

Replication code is available at 10.5281/zenodo.18605304.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-73874-8.