1 Introduction

Climate change is one of the defining societal and policy issues of our time. As climate impacts are already felt in societies and economies, governments around the world are mobilizing enormous resources to help reduce greenhouse gases and to adapt to climate impacts. Given the complexity of the mitigation and adaptation strategies, which involve a variety of different constituencies and span a broad technology spectrum, advanced research methods can help guide policies to be most effective, efficient, and inclusive.

Computational social science has already made significant contributions to climate policy evaluation and climate impact research. This is an area of high interdisciplinarity that combines climate physical information with data representing human systems. Broadly speaking, methodologies have focused on either retrospective or prospective assessments, though mixed approaches have also emerged. Retrospective, ex-post analysis has focused on evaluating the impact of climate policies and global warming on a variety of social and economic outcomes, such as economic productivity, inequality, labour market participation, social acceptance, etc. This strand of research has employed a variety of statistical approaches such as econometrics and machine learning to data either historically observed or purposefully generated (e.g. via surveys, or experimental trials). Prospective, ex ante approaches have tackled the issue of projecting the consequences of climate change and climate strategies into the future, often far distant ones given the inertia in the climate and economic systems. Here, methodologies have focused on numerical approaches such as optimization and simulation models, often integrating different components of the human and climate systems. Prominent examples include integrated assessment models (IAMs), energy system models, computable general equilibrium models, and agent-based models.

Both approaches have had a significant policy influence in Europe and elsewhere. The increased empirical recognition of the social and economic risks of climate change has helped to make climate a policy priority. Ex-post policy evaluation has helped improve our understanding of the functioning of public interventions and to improve them. Future scenarios of emissions and the energy and economy transformations compatible with decarbonization have had a major influence in determining the outcomes of international climate negotiations such as the Paris Agreement and informing society of the possible course of actions via international panels such as the Intergovernmental Panel on Climate Change (IPCC).

This chapter provides a succinct review of the role played by both statistical and numerical methods in climate change mitigation and adaptation research. It then discussed the policy implications of the research done so far on specific policy-relevant issues. Finally, it maps possible evolutions and future contributions of computational social sciences to help the impending fight against climate change.

2 Modelling the Climate Economy

2.1 Model Paradigms

Understanding the complex relationship between climate change, social and economic factors, and the needed transition in the emitting sectors such as agriculture and energy cannot be done without complex tools. Indeed, computational approaches have become the dominant paradigm for generating scenarios of future climate and of climate-resilient strategies. One class of models that is prominent in this field goes under the name of IAMs. As evident from the title, integrated assessment modelling is a general term that captures a variety of paradigms, often of a very different nature.

A general distinction has been made between benefit-cost and detailed process models (Weyant, 2017). Both model paradigms include greenhouse gas (GHG) emissions, compute their climate consequences, and feature technologies to mitigate and adapt to climate change. They have been used for decades to inform the design of climate policies. However, they have some fundamental differences which have set them apart, despite often being equated. Benefit-cost models have a relatively aggregated representation of the mitigation component but include the feedback of climate change on the economic system. The closed-loop formulation allows doing what the name suggests: to compute costs and benefits of climate action and to optimize the trade-off between the two, suggesting courses of action which are therefore economically optimal. This class of models originates from economists’ role in early climate debates, such as the US National Academy of Science climate committee established in the early 1980s which included two future Nobel prize winners in economics, Tom Schelling and William Nordhaus. Nordhaus developed back then what has become a standard benefit-cost model, the DICE model (Nordhaus, 1994, 2008; Nordhaus & Boyer, 2000), for which he eventually won the prestigious award in 2018. DICE is a dynamic, non-linear optimization model based on the optimal growth framework of Ramsey-Cass-Koopmans, coupled with a simplified climate model and a very simple representation of emission reduction technologies. Despite the simplicity and reliance on standard, neo-classical approaches, the model has been used extensively by many scholars in different fields, thus becoming a classic still in use today even in regulatory work. Other benefit-cost IAMs are FUND (Tol 1997, www.fund-mode l.org ) and PAGE (Hope, 2006).

In parallel to the development of benefit-cost models, a different approach emerged. This built on the work done in the 1970s to model energy systems in response to the oil price shocks, for example, by the Energy Modeling Forum, as well as by the establishment in the late 1980s of the IPCC. By the early 1990s, several detailed process models had been developed, and even a model comparison project on the economic costs of climate control was completed (Gaskins & Weyant, 1993), at the same time when structured model comparisons emerged in climate science (Smith et al., 2015). Disaggregated process-based models represent the underlying processes more explicitly than aggregate models: for example, mitigation technologies are represented in much greater detail, the climate components are based on intermediate complexity models calibrated upon large-scale climate models, and the economic sectors might be represented at higher granularity. This class of models includes simulation and optimization approaches but tends to focus on the evaluation of policies such as emission reduction ones rather than finding the optimal climate conditions for the economy. Over time, dozens of such models have been developed, whether global tools for understanding international climate policies such as the ones envisioned in the Paris Agreement or national or subnational tools to simulate local policies. An association has been established 14 years ago with the purpose of (Emmerling & Tavoni, 2021; Gambhir et al., 2019; Weyant, 2017) creating a community of scholars and practitioners focused on integrated modelling for climate change.Footnote 1

In addition to these two broad classes of integrated assessment models, additional numerical approaches have been developed over the years. A large number of computable general equilibrium (CGE) models are now available, though not always classified as IAMs (Parrado, 2010; Rausch et al., 2011). These models, alongside dynamic stochastic general equilibrium (DSGE) ones, have a detailed representation of economic sectors and of their interaction. They are used for policy evaluation and not optimization, thus belonging to the detailed process category. For example, the European Commission regularly employs a CGE and a DSGE for the impact assessment of its policy proposals, including the ambitious Fit-for-55 policy package. Model paradigms that do not enforce equilibrium are also available and also used for policy appraisal. These include macro-econometric approaches as well as agent-based models applied to climate change (Keppo et al., 2021; Lamperti et al., 2018; Ma & Nakamori, 2009).

2.2 Modelling Relevance for Climate Policy

It is hard to underestimate the contribution of computational models to the climate change policy debate, whether it is about policy impact assessment or international negotiations. The reliance of scientific bodies such as the IPCC on model-generated scenarios and numbers is clear evidence of this process: from the less than 200 scenarios in the 4th assessment report of the IPCC, scenarios have grown to well over 1000 in most recent ones. The Paris Agreement agreed upon in 2015 was, for example, heavily influenced by the fifth assessment report and in particular by the results of integrated assessment models which simulated the implications of stabilizing temperature below 2 °C. The climate neutrality pledges recently announced by several major economies can be partly attributed to a sentence in the IPCC 1.5 special report which is the outcome of model-based evaluations: ‘In model pathways with no or limited overshoot of 1.5°C, global net anthropogenic CO2 emissions decline by about 45% from 2010 levels by 2030 (40–60% interquartile range), reaching net zero around 2050 (2045–2055 interquartile range)’.

Models have not just provided the timing of climate neutrality, which has become such a focal point for international climate policies. They have also depicted transformation pathways for the economic, energy, and land systems compatible with climate stabilization: most of the scenario work has indeed focused on cost-effective pathways meeting given climate targets. These constraints have been taken as given by policy, including temperature targets but progressively carbon budgets, which have emerged as a reliable climate metric from the climate science community (Allen et al., 2009). The integration of process-based models with climate science speaks of the multidisciplinary nature of mathematical modelling.

Models have also laid out different technological and behavioural pathways to net-zero emissions: though in all climate stabilization scenarios fossil fuels are phased out quite rapidly and replaced by renewable sources and energy demand measures, the combination of different technologies and behavioural changes can change substantially across pathways consistent with the Paris Agreement. For example, the same temperature goal can be achieved with different usage of CO2 removal strategies. Although all scenarios compatible with 1.5 °C envisage some negative emission technologies, the timing and extent of removals vary across models and scenarios and have been the subject of intense academic debate (Fuss et al., 2014; Tavoni & Socolow, 2013). The extent of CO2 removals is driven as much by techno-economic assumptions made in the models about the technologies as by normative hypothesis and scenario design. For example, the choice of the intertemporal discount rate is a well-known key parameter in integrated assessment modelling but mostly for benefit-cost optimization rather than cost-effective analysis of a given temperature target. The introduction into IAMs of negative emission strategies, however, has made this normative assumption relevant also for cost-effectiveness: by sharing the burden towards future generations, scenarios with high discount rates are characterized by a higher reliance on CO2 removals (Emmerling et al., 2019). This example shows how normative judgments, often implicit in model formulations, matter for climate change (Saltelli et al., 2020). The consequences of these choices matter not just for academic purposes but also for policy design: the extent and need of negative emission technologies are now discussed in international policy such as in the revision of the Nationally Determined Contributions, as well as in national policies such as the EU Green Deal where a separate accounting of removals from standard emission reductions and their management into the Emission Trading Scheme is now debated (Rickels et al., 2021).

Computational impact assessments have also examined the social and economic consequences of climate policies. For example, the European Commission legislative proposals—including the Fit-for-55 and the mid-century strategies—have been vetted by a series of climate-energy-economy models, which have computed the repercussions for economic activity, employment, and other social dimensions. Table 14.1 (EC, ‘Policy scenarios for delivering the European Green Deal’) reports the latest estimates for the increased emission reduction ambition recently announced by the European Commission for the GDP of Europe, as computed by three climate-energy-economy models (JRC-GEM-E3, E3ME, and E-QUEST). Although all models tend to agree on relatively small macroeconomic impacts of decarbonizing the European economy, it is worth noting that different models produce estimates of different signs, as well as that the results will depend on the details of the policy formulation. For example, the GEM-E3 model of the JRC suggests that the economy will slightly contract, whereas the E3ME by Cambridge Econometrics foresees a policy-induced economic expansion. The reason for this discrepancy is the underlying economic framework assumed in each model: GEM-E3 is a computational general equilibrium model which embeds assumptions about relatively well-functioning markets. E3ME is a macro-econometric model which does not assume optimizing behaviour and full utilization of resources but is rather based on a simulation approach based on economic accounting matrices and historical relations which include, for example, voluntary and involuntary unemployment. E-QUEST is a micro-founded dynamic stochastic general equilibrium model. The choice of the European Commission to employ three models of different nature highlights the fact that when it comes to economic and social repercussions, the model paradigm choice is essential and it is hard to discriminate between good and bad models, contrary to physical models such as Global Circulation Models. Importantly for policy evaluation, different models can simulate different policy provisions: Table 14.1 shows just how relevant is the type of policies which will be implemented, for example, how carbon tax revenues will be used.

Table 14.1 Macroeconomic implications in terms of EU GDP variations of implementing an emission reduction of 55% by 2030. Source: EC, ‘Policy scenarios for delivering the European Green Deal’
Full size table

If economic policy consequences are hard to predict, social impacts are even more complicated and yet increasingly important in the objective of achieving a just transition. Economic and social inequalities, for example, are a major driver of policy acceptance and a crucial policy objective. Traditionally, integrated assessment models have not focused on inequality (Emmerling & Tavoni, 2021): however, this has now become a policy focus, and new work to expand models in order to address this request is ongoing (Gazzotti et al., 2021). The use of computational models for understanding behavioural responses has proven more difficult. As a result, models have prioritized technological, supply-side solutions over demand-side ones (Creutzig et al., 2018). This is because of the difficulties of portraying human behaviour into tractable mathematical formulations, and the traditional paucity of empirical evidence on how households respond to economic and behavioural interventions. As we will discuss later on, the empirical evidence has accumulated in recent years thanks to more robust statistical approaches and data with higher resolutions. This has opened up the possibility of using models, such as agent-based models, to better capture the behavioural responses to climate policies. Even standard integrated assessment models have developed and can now account for lifestyle changes necessary to achieve the low carbon transformation (van den Berg et al., 2019).

Besides climate mitigation policies, computational models have been used to compute the impacts of climate change and to design adaptation strategies. One policy-relevant application of IAMs, for example, has been to compute the social cost of carbon (SCC)—the monetized damages associated with an incremental increase in carbon emissions. The SCC is used to evaluate the cost-effectiveness of climate policies in the USA and has been traditionally computed using three benefit-cost IAMs. The economic valuation of climate impacts is important also in climate negotiations on the discussion of loss and damages. One major driving factor in the wide range of estimates of the SCC is the formulation and parametrization of the damage functions. The damage function originally used in simple benefit-cost models such as DICE has been criticized for lack of empirical basis and for recommending insufficient climate ambition (or equivalently for producing too low SCC). This highlights the importance of better integration of empirical and modelling approaches, a point on which we will return further in the chapter.

One area where computational models have indirectly contributed to policy assessment, both for mitigation and adaptation, is the generation of counterfactual emission scenarios. Evaluating policies ex ante requires first defining a world in which those policies are absent, as a reference over which to calculate the policy repercussions. This is a notoriously difficult task, given the uncertainties of predicting future outcomes but also the challenges of defining which trends and policies to include. For global climate issues, the workhorse of counterfactual emission scenarios is that of the Shared Socio-economic Pathways [SSPs (Riahi et al., 2017)]. The SSPs depict five possible scenarios of the future, with different demographic, economic, and technological trajectories, and consequent challenges for climate mitigation and adaptation challenges (O’Neill et al., 2013). The narratives span different evolutions of prosperity, inequality, and environmental degradation: they are characterized by both quantitative elements, such as population and GDP growth, and qualitative elements such as technological narratives. The SSPs have been simulated by five IAMs, which have produced the resulting emission trajectories, and consequent climate outcomes. These have been used by several other scientific communities, most notably the climate science and climate impact ones. They also have had policy repercussions, for example, on the social cost of carbon.

2.3 Challenges in Using Integrated Assessment Models to Inform Societal Change

So far, we have highlighted the growing relevance of computational mathematical models for the prospective evaluation of climate policies. Climate strategies’ repercussions for societies, households, and businesses are now routinely quantified using numerical models. Although this speaks of the growing importance of computational sciences in the climate domain, the increased reliance on structured approaches has not come without problems. For example, quantitative approaches have been condemned for exploring only a narrow set of possible futures and not keeping track of the rapid evolution of the climate technology and policy context. Most models implicitly represent value judgments and social preference and have been criticized for not exploring these normative assumptions, which are also plagued by uncertainties (MacAskill, 2016).

The modelling community has responded in various ways to the challenges of underplaying or not representing the full range of uncertainties. For example, global sensitivity analysis (Razavi et al., 2021) is a well-proven approach to ensure model-generated results are robust. It has been used in the context of large-scale climate-energy models only to a limited extent (Butler et al., 2014; Marangoni et al., 2017). An additional strategy to deal with model uncertainty has been a coordinated community response based on multi-model ensembles. Multi-model ensembles provide a range of plausible outcomes from a set of harmonized assumptions (i.e. given carbon budgets or temperature targets, still within this century). The ensemble spread is typically used to quantify model uncertainty. Although this process appears unambiguous, it can deceptively be so, since it is based on the assumption that models constitute independent estimates (Abramowitz et al., 2019; Merrifield et al., 2020). Selection and availability biases determine the typology and number of models involved in the comparison project and model dependencies inherent in the community work. Uncertainty is influenced by choices made in the model comparison project construction (Knutti et al., 2010). For example, ensemble members are not independent: they have historically shared code, use similar parametrization, and—an issue that is especially important for models examining socio-techno-economic transitions—belong to similar paradigms. The implicit normativity of climate-energy-economy models further contributes to compounding the sources of uncertainties and model relations.

The challenges in formulating policy recommendation from a vast number of scenarios and models, which often disagree, have led the policy community to often embrace simpler approaches based on few, representative scenarios. This, for example, has become the standard approach of IPCC in presenting its scenario space in an accessible way. International organizations, such as the International Energy Agency, typically present very few scenarios that become standard ones. However, the problem of the plausibility of scenarios and of the associated uncertainties is not solved by reducing the scenario space arbitrarily, unless statistical valid approaches are employed, which is typically not the case. This has important policy ramifications: for example, impact studies typically take the SSP5-RCP8.5 as a benchmark scenario, despite this being a relatively extreme one which was judged as a low probability by the scenario community itself (Ho et al., 2019). Advancements in statistical approaches and in behavioural science can be used towards not only more robust empirical evidence of policy effectiveness but also to make scenarios credible and insightful, something we turn to in the next section.

3 Data Science for Climate Impacts and Policy

The contribution of computational social sciences to climate change is not limited to numerical modelling. Actually, some of the most important contributions of the literature with direct or indirect policy repercussions have come from empirical and statistical approaches.

3.1 Data-Driven Approaches for Climate Economics

The fact that climate change has been changing, in addition to the natural and large variability in weather patterns, and the fact that climate and energy policies have been slowly but gradually deployed have provided previous information to scholars interested in the causal relationship between climate and its solutions and high-stakes social and economic issues. The growth of observations not just on climate outcomes but also on social and economic ones at high spatial resolution has provided sufficient statistical power to conduct innovative empirical research.

Several approaches have been used to infer causal climate-socio-economic linkages. Panel data econometrics, for example, has been applied to understand the impacts of climate change on a large set of outcomes, as discussed below. Other econometric approaches such as difference in difference, matching, and regression discontinuity designs have been used to infer causal relationships in the absence of an exogenous variation to be exploited. Standard regression approaches have been used where counterfactual randomization was ensured, such as in randomized controlled trials. Finally, machine learning methods are increasingly used to understand and promote sustainable policies: for example, machine learning has been applied to satellite imagery, whose increased abundance and resolution can provide crucial information on sustainability in areas of the world where data is scarce (Burke et al., 2020) and where climate change impacts are also more likely to occur. Novel algorithms have also been used to better understand energy usage patterns, for example, in the residential and transportation sectors where high-frequency information is now available, and to study policies to motivate behavioural and technological changes towards a greener society.

3.2 Relevance of Empirical Methods for Climate Policy

Empirical methods are key for understanding policy effectiveness and environmental social and economic disruptions. They are also needed in order to calibrate prospective impact assessments. Over the past few years, empirical studies have greatly advanced the understanding of climate change impacts and of the policies meant to address them.

One major area has been in the quantification of climate social and economic impacts. Traditionally, the climate impact functions used in benefit-cost analysis and for the social cost of carbon were based on prospective studies which raised issues of replicability and transparency. Over the course of the past 10 years, a wealth of data-driven approaches have highlighted the relationship between historical weather variability and many outcomes (Carleton & Hsiang, 2016). For example, temperature heat induces mortality and has been connected to aggression and violence. Agriculture and crop yields are related to temperature in a strongly non-linear way, with yields dropping when temperature exceeds certain thresholds. This non-linear relationship has been documented also for energy demand (Auffhammer et al., 2017), with peak demand rising when temperatures are high.

On the economic side, temperature variability has been associated with significant macroeconomic repercussions. The identification of a non-linear relationship between temperature and economic growth has highlighted how climate impacts can persistently slow economic progress (Burke et al., 2015; Dell et al., 2012; Kalkuhl & Wenz, 2020). This view is in stark contrast to the previously assumed relations which were based on the levels and not the growth of the economy. The consequences of this new empirical evidence have been particularly prominent in the benefit-cost assessments of climate policies and in the calculations of the social cost of carbon. Once the empirically derived damage functions were plugged into the IAMs, the recommendations for policy stringency changed dramatically, and for the first time, it appeared that stabilizing climate change within the goals of the Paris Agreement made global economic sense (Gazzotti et al., 2021; Glanemann et al., 2020; Hänsel et al., 2020). Similarly, the social cost of carbon—a policy-relevant metric for setting policy in the USA—increased substantially over previously available estimates (Ricke et al., 2018).

The data science advanced on the economics of climate change impact also highlighted the major economic inequalities brought about by climate change. These economic inequalities are detectable already today (Diffenbaugh & Burke, 2019) and are forecasted to persist even in case of ambitious emission reductions, and even more in the absence of cooperation, as shown in Fig. 14.1. Although the extent of persistency of climate economic impacts is a subject of intense academic debate (Piontek et al., 2021), the accumulated evidence has shown the importance of reducing emissions as fast as possible, preparing adaptation systems, and considering additional climate interventions such as CO2 removals, to avoid temperature overshoots and consequent social and economic repercussions.

Fig. 14.1
figure 1

GDP per capita, net of costs and impacts, population-weighted distribution in 2100 for three climate scenarios (BAU without climate impacts, global cooperation and noncooperation). From Gazzotti et al. (2021)

Full size image

Another area of data-driven research which has important ramifications for policy design is that of behavioural science and of the experimental economics literature quantifying traditional and behavioural interventions. Behavioural sciences have consistently shown how human behaviour is fraught by biases, but also that several of these can be predicted, and thus partly addressed (Ariely, 2010). Many governments around the world have promoted the use of behavioural informed public policies, including but going beyond the use of ‘nudges’ (Banerjee et al., 2021). Methodologically, disentangling the impact of policy interventions, including behavioural ones, on outcome variables is difficult. Confounding factors such as exogenous trends (e.g. in energy prices, preferences, etc.) and self-selection (e.g. environmentally sensitive households more likely to enrol in pro-environmental programmes) have traditionally made it difficult to quantify the causal impact of policies. However, the embracing of statistical approaches based on counterfactual randomization, such as laboratory, online, and field experiments, has opened up the possibility to test for causality of policy interventions. Randomized controlled trials have been now done on millions of households and have helped evaluate a variety of interventions, such as information provision, message framing, social comparisons, monetary, and symbolic incentives (Allcott, 2011; Allcott & Mullainathan, 2010; Bonan et al., 2020, 2021; Ferraro & Price, 2013; Fowlie et al., 2015). These interventions have been assessed using reliable metrics of energy usage (and consequent GHG emissions) such as actual metered electricity consumption, thus providing a reliable line of evidence. The main results of this stream of computational social science have shown that behavioural interventions, if properly designed and implemented, can lead to small but significant energy and emission reductions. However, their effectiveness is context-dependent and varies significantly across population subgroups. As such, these policy instruments should complement but not substitute traditional interventions, including infrastructural and incentive-based ones.

The potential of data science to inform climate policymaking is enormous, but it has not yet realized its full capacity and should anticipate possible critiques which might emerge in the future. In terms of potential applications, data-driven approaches can help inform local decisions and design climate-resilient infrastructures at the local level. Cities are places that abound both in data and emissions, and where well-designed infrastructural policies can promote lifestyles that are both sustainable and inclusive (Creutzig et al., 2018). Data can be used to transform mobility services and reduce congestion and pollution. Some local institutions have already begun using high-resolution data for public purposes related to sustainable planning, but only limited potential has concretized. One concern with such an extension of data-driven urban policies regards the question of data equity and privacy. The way data is handled when it comes to policymaking is as crucial as the actual policies which will derive from it: data availability is often skewed towards certain sociodemographic areas and population subgroups, and resulting policies need to ensure to go beyond pre-existing social arrangements. Furthermore, the question of privacy which has become a central element of regulatory design needs to be accounted for when relying on data-driven impact evaluation.

4 Towards an Integrated Computational Approach

Overall, we have highlighted that computational approaches—both model-based and data-driven—have played an increasingly important role in climate change policies, both for mitigation and adaptation. Computational approaches have become ubiquitous, and policymaking is now heavily dependent on them, whether it is for determining the impacts of proposed legislation or of already implemented one. However, in order to serve society well, mathematical and statistical modelling should be accompanied by an epistemic strengthening of the underlying theoretical basis, empirical validity, and scientific practices (Saltelli et al., 2020).

One focus area for climate-purposed computational approaches is that of integrating data and model-driven approaches. Traditionally, these two approaches have been used to look at retrospective and prospective policy assessment, respectively. This rigid division of labour needs to be overcome if we want to have policy appraisal which can be learned from actual experiences: the growing number of energy and climate policies being tested in real-world conditions can now provide important information for calibrating models and made them more policy-relevant. Furthermore, the growing availability of high-resolution data such as those from satellite imagery, social media, and high-frequency metered energy and environmental indicators can be harnessed to understand behavioural policy responses at a high level of granularity. Machine learning approaches can then be combined with model inputs and output to increase the understanding of the model-embedded processes and to better predict policy responses. Finally, model validation and adequate exploration of the uncertainties should become scientific practices fully integrated in mathematical modelling. Computational approaches to do that effectively are now available, and their properties are well known (Razavi et al., 2021), and yet they are often not done (Saltelli et al., 2019). This speaks of the importance of a tighter and regulated relationship between researchers and policymakers, with clear guidance from policy evaluation agencies on scientific practices and robust methodological approaches. If crafted properly in a coordinated and co-designed manner, computational social science can be of tremendous value to climate policy-making and help accelerate the climate transition and ensure it is carried out in a just and inclusive way.