3

Socioeconomic Module

In this chapter, the committee discusses preparation of the socioeconomic and emissions inputs to SC-CO₂ estimation and recommend improvements to the current IWG procedure. The chapter presents a basis for evaluating current and potential future approaches and the desired characteristics of a socioeconomic module. It also includes a survey of the resources available for the task, including scenario databases, models of the economy and emissions, means of extracting information from historical data, and expert elicitation. An illustration of an improved method for projecting population, economic activity, and emissions that could be applied in the near term, with a focus on characterizing uncertainty in the variables to be used in the climate and damages modules is provided. For the longer term, recommendations are offered for the development of a socioeconomic projection model designed to meet the special requirements of SC-CO₂ projection, noting that it is best supported by a program of research and development (R&D) on economic modeling frameworks.

BASIS FOR EVALUATION

The purpose of a socioeconomic module is to provide a set of projections of population, and gross domestic product (GDP) that drive projections of CO₂ and other relevant emissions, which are inputs to the calculation of the baseline climate trajectory. These projections take into account possible future mitigation policies and other drivers of change (see Box 2-2, in Chapter 2). The baseline emissions in turn influence

the response of the climate to a pulse of CO₂ emissions (see Chapter 4). Estimates of population and GDP, perhaps disaggregated by region and sector, are also direct inputs to the damage calculations (see Chapter 5). The trajectory of GDP per capita is also needed for the committee’s recommended discounting procedure (see Chapter 6).

The socioeconomic component of the current IWG SC-CO₂ estimation methodology is based on five scenarios of population, GDP, and emissions through 2100: they were selected from those produced by the detailed-structure integrated assessment models (IAMs) used in the EMF-22 multimodel comparison exercise of the Energy Modeling Forum (Clarke et al., 2009). Four of these scenarios are reference scenarios (no mitigation policy) that roughly span the distribution of reference fossil fuel combustion and industrial CO₂ emissions in the EMF-22 project. They entail atmospheric CO₂ concentrations between 612 and 889 ppm in 2100. One of the five scenarios involves atmospheric stabilization at a radiative forcing equivalent to 550 ppm CO₂ by 2100, and thus assumes moderately strict mitigation measures. The IWG extended each of these scenarios to 2300 to capture the persistence of climate change and its associated net damages, assuming that growth rates of population and per capita GDP in each scenario decline linearly to zero in 2300. The IWG does not offer a rationale for these growth assumptions.

The five scenarios used by the IWG do not span uncertainties in relevant variables (e.g., GDP, population, and energy) or reflect the broader scenario literature (e.g., Kopp and Mignone, 2012; Rose et al., 2014b). In estimating the SC-CO₂, these five scenarios are weighted equally, thereby treating them as equally likely. The IWG does not provide a justification for this implicit assumption. As discussed throughout this report, good scientific practice requires that key variables and associated uncertainties be clearly identified, characterized, and supported; that the methods used to produce probabilistic projections be consistent with the available peer-reviewed literature; and that the projections themselves be consistent with the main features of the historical record.

For estimating the SC-CO₂, the socioeconomic module needs to produce projections far enough into the future to capture the vast majority of discounted damages.¹ The committee recognizes that this may entail projecting GDP, population, and emissions two to three centuries into the future, which presents a significant challenge. Although projecting the impact of a change in radiative forcing on mean global temperature involves parametric uncertainty (see Chapter 4), the basic physics of the climate system are well established. In contrast, models that project

__________________

¹ Vast majority” is a deliberately vague term that signals much more than majority, but not 100 percent.

population or GDP are subject to the behavior of individuals and social systems, which are more malleable than the principles governing physical systems. Therefore, a near-term approach for a socioeconomic module that relies on projecting historical data, combined with elicitation of expert judgment is presented. The importance of conducting sensitivity analyses for the distribution of GDP, population, and emissions to investigate their impact on estimates of the SC-CO₂ is also discussed.

For any long-term projection of population and GDP, associated projections of emissions of CO₂ from fossil fuel and industrial sources and land use change, as well as other greenhouse gases and aerosols, will depend on the joint evolution of various technologies and policies aimed at mitigating emissions. Thus, it would be desirable for the socioeconomic module to explicitly take into account the likelihood of these future changes. The committee discusses a near-term approach consistent with these criteria below (“Developing a Socioeconomic Module in the Near Term”).

Two additional desirable criteria are more difficult to satisfy. The first deals with disaggregation of global totals. As discussed further below, historical experience and expert judgment provide a basis for computing a probability density function for both global average per capita GDP growth over time and for global population that are consistent with alternative economic growth projections. However, the empirically based literature on climate-related damages is typically concerned with particular regions and even particular sectors (e.g., agriculture) in each region. Unfortunately for modeling purposes, the relative contributions of different sectors and regions to global growth has varied significantly over time. For instance, in 1960 it would have been difficult to predict the rise of the Chinese economy or the fall of the Soviet Union over the following half century or the advance of computer and communications technologies and their spinoffs. In a world of many regions and many sectors, rigorous characterization of uncertainty regarding their relative contributions to global growth would require construction of a probability density function over many variables, extending far into the future—a task well beyond the current capacity of the research community. Accordingly, a less ambitious approach is recommended in the near term.

The second desirable but difficult criterion is the incorporation of feedbacks from the damages and climate modules to income, population, and emissions projections. As discussed in Chapter 2, there are many potential linkages and feedbacks between modules. Identifying the most important feedbacks and incorporating them in a fully integrated socioeconomic-climate-damages framework would represent a significant advance beyond the current state of the art.

Development of such a framework might start with the climate system impacts on human and natural systems described by Working Group

II of the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC, 2014a), which identifies regions and sectors where such interactions appear to cause the most significant physical impacts. For some impacts a next step could be to incorporate recent research that assigns economic values to such impacts (e.g., Diffenbaugh et al., 2012; Reilly et al., 2012a, 2012b; Taheripour et al., 2013; Baldos and Hertel, 2014; Grogan et al., 2015; Diaz, 2016). As discussed in the final section of this chapter, such an effort would also be an important component of a longer-term research strategy.

The next section discusses the scholarly resources that are available to construct an improved socioeconomic module in an SC-CO₂ framework. The subsequent two sections cover an approach to producing improved estimates in the near-term and a recommended longer-term strategy.

LITERATURE AND METHODS

There are four resources that can be used in the construction of socioeconomic modules: detailed-structure models, scenario libraries, time-series analysis of historical data, and elicitation of expert opinion.

Detailed-Structure Models of the Economy

The models used to generate the scenarios (available in the various libraries discussed below) are significant resources available to the IWG.² Among these are detailed-structure models that attempt to model the structure of the global economy. These represent nations and aggregate regions and their interaction through international trade and disaggregate the sectors that make up the individual economies. They differ from reduced-form models like the IAMs used to produce estimates of the SC-CO₂, SC-IAMs. SC-IAMs model a single global economy or a small number of regions and include more limited economic sectoral detail than a detailed-structure model.³

These detailed-structure models differ from one another in mathematical form, but they tend to fall into two general categories, partial equilibrium and general equilibrium. Partial equilibrium formulations represent particular sectors in detail (e.g., energy, agriculture) but do not consider interactions among sectors and interactions with the macro economy. Therefore, many prices in the economy are assumed to be exogenous. Examples of this type of detailed-structure IAM include the global change assessment model (GCAM)⁴ and Prospective Outlook on Long-term Energy Systems (POLES) (Kitous, 2006). In contrast, general equilibrium formulations consider the market transactions and linkages among sectors (including capital, labor, resource markets, and international trade), and all treat prices in the economy as endogenous. Examples of this approach include the anthropogenic emission prediction and policy analysis (EPPA) model (Chen et al., 2015), MERGE (a model for estimating the regional and global effects of greenhouse gas reductions) (Blanford et al., 2014), and World Induced Technical Change Hybrid Model (WITCH) (Bosetti et al., 2006).

These types of models have been used not only for scenario construction, but also for more formal uncertainty analysis of energy and emissions (e.g., Reilly et al., 1987; Manne and Richels, 1994). Recently, an analysis by Gillingham and colleagues (2015) used the EPPA, GCAM, MERGE, and WITCH models (along with two reduced-form IAMs, DICE [Dynamic Integrated Climate-Economy model] and FUND [Framework for Uncertainty, Negotiation and Distribution model]) in a study that considered uncertainty in population and GDP. Another analysis (Bosetti et al., 2015) imposed uncertainty in the cost parameters of key technolo-

__________________

² An example is the set of models that contributed to the Fifth Assessment Report’s Working Group III scenario database (Intergovernmental Panel on Climate Change, 2014b, Annex II, Table AII.14).

³ For an overview of modeling terminology, see Box 2-1 in Chapter 2.

⁴ See http://jgcri.github.io/gcam-doc [January 2017].

gies in the GCAM and WITCH models, while holding population and GDP constant. To generate a projection of emissions for a study of uncertainty in climate, Webster and colleagues (2008, 2011) introduced both types of uncertainty in the EPPA model, considering both uncertainty in population and drivers of GDP and uncertain distributions of many input parameters, such as elasticities, resource stocks, and technology costs.

These models produce information of use in damage estimation, including both regional and sectoral detail (e.g., the role of the agricultural sector). Moreover, many are formulated to provide additional details needed for climate modeling, such as emissions of land CO₂ and non-CO₂ greenhouse gases and their geographical distribution. Because of their formulation, these models ensure consistency in the relationships among population, GDP, and emissions of CO₂ and other greenhouse gases in each nation or region and in their aggregation to global emissions. Moreover, these features are preserved in the construction of probabilistic scenarios or other representations of uncertainty. At the same time, use of a detailed-structure model does not reduce the underlying information requirement associated with projecting regional and sectoral detail far into the future.

Scenario Databases

A number of multidecade to century-scale scenarios have been developed and have been catalogued in study-specific and scientific community assessment libraries. In 1992, the Intergovernmental Panel on Climate Change (IPCC) developed a set of global greenhouse gas emissions scenarios for use in climate change policy assessments, called the integrated scenarios 1992 (Leggett et al., 1992; Pepper et al., 1992). Through a long and complex process, the IPCC updated those scenarios in its Special Report on Emissions Scenarios (Nakicenovic et al., 2000).

Since then, the IAM community has published a large number of scenarios, most of them generated by specific intermodel comparison studies, some with publicly available scenario libraries. In its Fourth Assessment Report and Fifth Assessment Report, Working Group III of the IPCC assembled the research community’s scenarios into large libraries in support of their respective reports (Intergovernmental Panel on Climate Change, 2007b, 2014c). The IPCC scenario libraries are a rich scientific resource with large numbers of scenarios (e.g., more than 1,000 in Intergovernmental Panel on Climate Change, 2014c), but one that needs to be used with care (see discussion below). Riahi and colleagues (2016) describe the vast amount of scenario work that has been completed, providing useful information to support future scenario construction.

Since the 2000 IPCC compilation, two specific sets of scenarios have been produced—representative concentration pathways and shared socioeconomic pathways. The former were designed to provide consistent, standardized radiative forcing information for the purpose of coordinated experiments for climate modeling (Moss et al., 2010; van Vuuren et al., 2011). The latter were designed to complement the former with additional information beyond radiative forcing to support studies of climate change impacts, adaptation, and vulnerability. Specifically, the shared socioeconomic pathway scenarios provide macro socioeconomic information (O’Neill et al., 2014), such as population structure, education levels, extent of urban development, and income distributions.⁵

Although these processes have provided much needed benchmark scenarios for coordinating the work of the various global change research communities, neither the representative concentration pathway nor the shared socioeconomic pathway scenarios was designed with SC-CO₂ computation in mind. More specifically, neither was formulated to characterize climate change or socioeconomic uncertainty. The broader existing scenario libraries do reflect some degree of both model and parametric uncertainty because a substantial number of modeling groups participated in these efforts. However, the libraries are problematic as the basis for developing probability distributions of population, income, and emissions because they do not formally consider parametric uncertainty or uncertainty over a full range of model input assumptions. In addition, in order to be useful, oversampling would need to be addressed in some fashion, with some models and studies represented more than others, and a variety of vintages of single models sometimes included. Meaningful statistics cannot be readily derived from these libraries without attention to these issues, even though attempts are regularly made to do so (e.g., by Working Group III of the IPCC [Intergovernmental Panel on Climate Change, 2014c]). Furthermore, scenario libraries, including the shared socioeconomic pathways, do not provide sectoral disaggregation, and they also typically extend only to 2100, even though projections beyond 2100 are important determinants of current SC-CO₂ estimates.

Another missing element in current scenario libraries is the effect of mitigation policies. As noted above, projections of emissions conditional on population and GDP logically need to account for the effect of future changes in mitigation policies in the United States and abroad, and such changes are themselves uncertain. Historical observations and scenario

__________________

⁵ Standardized policy assumptions, also referred to as shared policy assumptions, were also developed to represent ways that countries might reduce greenhouse gas emissions and move from a shared socioeconomic pathway baseline to a combination shared economic and representative pathway scenario (Kriegler et al., 2014).

libraries do not on their own provide a basis for attaching probabilities to future policies. Finally, preliminary work with historical data on the global economy (discussed below) indicates that the range of economic growth rates in existing scenario libraries is too narrow to properly reflect historical experience.

In short, largely because they were not designed specifically to facilitate the computation of the SC-CO₂ or characterize the global-level of uncertainty in that computation, existing scenarios are not well suited for this purpose. However, as we discuss below, they may be helpful in disaggregating projections of global population and GDP to regional or sectoral scale.

Using Historical Data and Expert Judgment for Long-Term Economic Projections

Scenarios are intended to provide an internally consistent description of a potential future, conditional on initial conditions and structural assumptions about economic system dynamics. In contrast, forecasts describe the likely future of one or more quantitative variables, often implicitly or explicitly probabilistic, based on empirical modeling.

As noted above, the IWG’s analysis indicates that projections to around 2300 may be necessary to adequately represent the damages expected to result from a pulse of CO₂ emissions. Unfortunately, the literature contains only a few examples of projections of population, GDP, and emissions of any sort beyond 2100 and provides little discussion of how to construct them (see further discussion below). In fact, the scenario libraries do not necessarily span even the range of historical experience. For example, among the IPCC baseline scenarios that extend to 2100 and were used by Working Group III in the Fifth Assessment Report (Intergovernmental Panel on Climate Change, 2014c), the range of GDP growth rates is 1.1-2.5 percent (with only 1 of 263 below 1.2% and only 2 out of 263 above 2.4%). Yet the historical data show that a set of representative rates would span a significantly wider range.

A study by Mueller and Watson (2016) provides a mathematically rigorous method for using historical data to construct probabilistic growth forecasts over future time horizons that are a large fraction (or even a multiple) of the length of the historical record. This method, like any based on historical data, rests on the assumption that the stochastic process of future growth will be the same as in the past. In addition to this assumption, such methods cannot detect or incorporate fluctuations that occur over periods longer than the historical record.

The key insight of Mueller and Watson is that low-frequency, persistent variation in historical data is the most relevant information for

understanding long-term uncertainty. In contrast, high-frequency, idiosyncratic variation in growth rates—for example, idiosyncratic shocks that arise each year—will average out over long horizons and will thus contribute little variability in the long run. Isolating these low-frequency variations and transforming the estimates of low-frequency contributions to growth back to the original sample space allowed the researchers to produce a representation of the low-frequency, persistent variation and use it to project a distribution of long-term average growth rates. In their work, Mueller and Watson looked at the solvency of the U.S. Social Security Trust Fund, with forecast example horizons of 75 years, using U.S. datasets as short as 67 years.

For an example relevant for the SC-CO₂ estimates, the committee used the Mueller and Watson approach for time horizons of 90 and 290 years (e.g., 2010-2100 and 2010-2300) for projections of per capita GDP growth using alternate datasets of 60 and 140 years.⁶ The assumption that the stochastic process governing future growth rates will be the same as in the past is very strong, especially over such a long time ratio of projection to experience, so it would be sensible for projections produced by this or any other time-series method to be evaluated by experts before being used in SC-CO₂ analysis.

Ultimately, this approach seems most useful for informing projections of economic growth, rather than population or emissions. Population projections involve complex trends in fertility and mortality and may need to be conditioned on per capita GDP. Emissions projections without accounting for any mitigation policy can generally gain less from historical data, since there has historically been little scientific and policy attention to climate change. However, using historical emissions information to develop a no-mitigation projection might be a useful input to an expert elicitation of future emission projections, which is the subject of the next section.

Given the state of scientific knowledge and historical data, it will also be necessary to rely on expert judgment in developing a socioeconomic module. As discussed in Chapter 2 and in more detail in Appendix C, there are best practices for eliciting expert judgments about the probability distributions of uncertain quantities. As discussed below, it will be impossible to avoid reliance on expert judgment in both the near term and longer term. In most cases, the committee believes it will not be sufficient for the IWG to rely only on its own expertise. It is important to be able to draw effectively on outside experts in the relevant disciplines.

__________________

⁶ The committee’s projections involved ratios of the length-of-projection horizon to historic sample that ranged from 2.0 to 4.8, compared with 1.1 in Mueller and Watson.

DEVELOPING A SOCIOECONOMIC MODULE IN THE NEAR TERM

As discussed above, the committee does not believe it will be possible in the near term to produce a module satisfying all of the criteria in Recommendation 3-1. However, the existing literature and methods do provide a basis for overcoming several shortcomings in the current IWG procedure. This section describes and recommends an approach that the IWG could implement in the near term.

The committee’s approach is based on the assumption that important aspects of future trends will be like those in the past, with elicitation of expert opinion being the only practical way to relax that assumption. Although the ideal modeling system for SC-CO₂ analysis would include structured feedbacks from climate and damages to economic activity and possibly even population,⁷ the committee does not believe that it is possible to build such a system in the near term. Hence, our approach for a near-term strategy, as in the current IWG approach, does not include those feedbacks. This section details the four steps in the proposed approach: (1) use econometric analysis to project economic growth; (2) develop probabilistic population projections; (3) use expert elicitation to produce projections of future emissions; and (4) develop regional and sectoral projections. It is important that this process reflect judgments as to the influence of future policies on the evolution of key technologies.

This approach also reflects the committee’s view that it is advantageous to have a small number of possibly interrelated projections of population, GDP, and emissions to pass to the climate module. A small number increases transparency and facilitates expert elicitation conditional on each projection. Three values are used in this approach, which is the smallest number that both introduces variability and provides a midpoint. For example, using three projections each of population and economic growth would require the experts to generate nine probabilistic projections of emissions. If terciles are used, so that three emissions projections that can be treated as equally likely are generated for each of the nine population/GDP scenarios, this will produce 27 global-level (populations, GDP, and emissions) scenarios to be passed to the climate module and, in disaggregated form, to the damage module.

An important question is whether a single set of scenarios from the socioeconomic module should be used in the climate module or whether sensitivity analyses should be conducted by using alternate sets of scenarios. Sensitivity analysis with respect to the discount rate, for which ethical

__________________

⁷ However, the DICE model currently used by the IWG does adjust global GDP and the capital stock for aggregate monetary climate damages each time period (see Chapter 5).

and policy considerations are relevant in addition to observable rates of discount is presented in Chapter 6. Sensitivity analysis is an appropriate way to account for ethical and policy considerations, which are especially difficult to reduce to probability distributions. In contrast, economic and population growth are observable and so a probabilistic projection approach based on historical data is appropriate for them. Emissions projections fall somewhere in between, because while historic emissions are observable, future emissions are subject to considerable policy influence. Overall, given the difficulty in projecting future GDP, population, and emissions, it would be valuable to examine the impact of alternate sets of scenarios to investigate their impact on estimates of the SC-CO₂.

The approach recommended in this report nonetheless focuses on a probabilistic approach to all uncertainties other than discounting. This pragmatic approach is based on the committee’s recognition that there are a quite limited number of sensitivities that can reasonably be expected to be carried through a regulatory impact analysis, in which the SC-CO₂ is only one of many variables. In the current approach of the IWG, for example, scenarios were used for socioeconomic variables, but they were ultimately collapsed to an average by assuming equal weights on each scenario. The committee believes the recommended approach provides a better scientific basis for the assignment of probabilities to alternative scenarios.

Use Econometric Analysis to Project Economic Growth

As discussed above, recent work by Mueller and Watson (2016) examined how to estimate probability distributions of long-term growth rates in economic variables from historical data. That is, by looking at a small number of low frequency cosine transformations of historic growth rates, a predictive density of average growth rates can be constructed over an arbitrary horizon. Expert elicitation can then be applied to determine how likely the historical pattern is to hold over alternative horizons. The key underlying assumption is that behavior over the observed historical sample is a valid basis for projections over the chosen horizon. Although their application focused on the United States, using 60 years of data to construct projections over 75 years, it is straightforward to apply the same approach to global data over alternative (much longer) horizons.

As an example of such an application, the committee used data from the Maddison Project⁸ to construct two time series of economic growth.

__________________

⁸ The Maddison Project, begun in 2010, promotes and supports cooperation between scholars to measure economic performance for different regions, time periods, and subtopics. For details, see: http://www.ggdc.net/maddison/maddison-project/home.htm [October 2016].

One is a measure of growth in global GDP per capita from 1950 to 2010. Prior to 1950, data are available for only a subset of the global economy, so 1950-2010 represents the only sample for which global growth is measured. For a measure of per capita GDP growth from 1870 to 2010, we used the subset of 25 countries in Barro and Ursua (2008a, 2008b): these countries collectively accounted for 63 percent of global GDP in 1950, but for only 46 percent of global GDP by 2009.

The estimation results are summarized in Table 3-1. For additional details on the data construction and the committee’s use of the Mueller-Watson approach, see Appendix D. For the 1950-2010 sample, a mean annual growth rate of 2.2 percent for real GDP per capita and a 90 percent probability interval of 0.3-4.0 percent for growth for 2010-2300 is estimated. For the 1870-2010 sample, the mean annual growth is 1.4 percent and the 90 percent probability interval is –0.8 ± 3.2 percent. The prediction intervals grow slightly, but not by much, for the longer 300-year horizon relative to 100 years.

It is unclear whether the longer series is a better basis for long-term growth projections, or the shorter series with more coverage. The longer series contains more information about long-term variation, but there are more measurement issues in the distant past so it may be less relevant for understanding behavior in the future. Even if global economic data did exist for several past centuries, for example, would one look to those data to model future uncertainty? The shorter dataset is more geographically complete, as well as more consistently measured. However, selecting the key economic jurisdictions in 1870 necessarily excludes countries that underwent transitions—through above average economic growth—into

TABLE 3-1 Estimated Annual Growth Rates Using the Mueller and Watson Procedure (in percent)

NOTE: See text and Appendix D for discussion and details.

key economies in 1950 and 2010. Both kinds of estimates could be informative in selecting or creating economic growth scenarios in the SC-CO₂ process, as well as for inputs to expert elicitation.

After developing probability distributions for average economic growth rates over one or more horizons through statistical analysis of historic data or other means, it is desirable to translate them into a small number of projections of economic activity. The committee believes this is important for both transparency and tractability. It is easier to communicate a smaller number of discrete growth rate possibilities. It is also useful for connecting economic projections with population and emissions projections that involve expert elicitation conditional on the economic projections.

The approach discussed above would be to select representative growth rates for several equally likely fractile ranges. The example below is based on the distribution underlying Table 3-1 and using the mean of each tercile. However, one could also explore matching the standard deviation or other features of the data.

Continuing with the example calculation, Figure 3-1 shows the full cumulative distribution function for the projected average growth rate over 300 years in the example discussed above, using 1950-2010 data.

Based on this result, one can identify three terciles to use as equally likely projections, formed by breaking the cumulative distribution into three parts on the vertical axis: below one-third, between one-third and two-thirds, and above two-thirds. This division corresponds to growth rates below 1.95 percent, between 1.95 and 2.45 percent, and above 2.45 percent, as defined by the two vertical dashed lines. One can then com-

pute the mean growth rate in each tercile: 1.0 percent in the first, 2.2 percent in the second, and 3.3 percent in the third, indicated by boxed x’s on the cumulative distribution function. These three scenarios, defined in terms of growth rates, can then be translated into projections of per capita economic activity by applying them to an initial year value.

Though this careful examination of the historical experience provides a sound basis for projection over coming decades, it may seem implausible to assume it would hold for centuries into the future, in part because of population aging or resource constraints. Thus, it will be useful for the IWG to elicit the opinions of economists and other experts concerned with long-term trends and structural change about how the length of time that such projections can be treated as representative and equally likely and how they could best be adjusted to take account of these longer-term influences. Estimates of the extent of difference with the past experience could be elicited, and the statistically derived distribution modified accordingly.

Develop Probabilistic Population Projections

Projections of population growth can take advantage of its underlying dependence on fertility and mortality rates and the age structure of society. These rates follow patterns, and the study of demography has sought to examine how these rates and the population age structure evolve over time. The International Institute for Applied Systems Analysis (IIASA) currently provides probabilistic population projections through 2100 (Lutz et al., 2014), as does the United Nations (2015a, 2015b). Both sets of projections are based on a review of the drivers of fertility and mortality in different parts of the world and (differing) judgments of what can be expected in the future (e.g., Gerland et al., 2014). For example, IIASA’s central growth rate projection from 2015 to 2100 is 0.18 percent, with an 80 percent prediction interval of –0.18 to +0.51 percent.⁹ Neither of these two sources report complete probability densities. It would be desirable for IIASA and the United Nations to make available the underlying probabilities, from which a small number of (perhaps three) projections could be chosen to approximate the probability density functions when treated as equally likely.

For population projections to 2300, the United Nations (2004) has published high, medium, and low projections, and Basten and colleagues (2013) have published projections under a range of assumptions about

__________________

⁹ For the total population sheet, see http://www.iiasa.ac.at/web/home/research/|researchPrograms/WorldPopulation/Reaging/2007_update_prob_world_pop_proj.html [October 2016].

fertility. Based on the more recent methodology (United Nations, 2015b), the probabilistic projections to 2100 could be extended further into the future. The IWG could explore that task with IIASA, the United Nations, and other researchers. Such extrapolation, like the economic projections beyond 2100, raise significant questions about whether the assumptions used in the model will hold over more than a century. It will be useful for the IWG to elicit the opinions of a group of expert demographers to validate and adjust probabilistic population projections beyond 2100.

There are reasons to expect that per capita income growth and population growth may be related in the long term. For example, more rapid rates of global economic growth would seem likely to hasten the demographic transition to lower birth rates in developing nations. Yet it seems unrealistic to expect a default inclusion of such relationships in any projections at this time given the dearth of academic research on integrated probabilistic projections of population and economic activity. Such projections could be included if the expert elicitation in economics or demography indicate the value of those relationships.

Combining population projections with each of the growth rates of per capita income would yield a relatively small set of projections of population and GDP that can be treated as equally likely and representative of the corresponding joint probability density function.

Use Expert Elicitation to Produce Projections of Future Emissions

The SC-CO₂ estimates are intended to be used in U.S. regulatory impact analysis (RIA) of proposed regulations and other policy initiatives. Accepted practice for benefit-cost analysis and the Office of Management and Budget (OMB) guidance for conducting RIAs establish that benefits and costs ought to be defined in comparison with a clearly stated alternative or “baseline,” with the baseline chosen to represent what the world would be if the proposed action (i.e., program, regulation, law) is not adopted. For example, OMB Circular A-4 (p. 15) states:

The committee notes that the consequences of any individual U.S. policy action affecting CO₂ emissions will take place in the context of

other actions in the United States, as well as actions by other countries. Under uncertainty, an appropriate distribution of baselines will therefore include a range of possible outcomes for these uncertain policy developments, combined with uncertain economic and technology conditions. Thus, the committee believes the IWG acted correctly in considering scenarios with alternative levels of future global CO₂ emissions mitigation, but that SC-CO₂ estimation can be improved by making such consideration more systematic.

Although knowledge of historical experience can inform judgments about the joint evolution of various technologies and of national policies to mitigate emissions, the committee believes it would be unwise to rely heavily on statistical analysis of the sort discussed above. Instead, the committee believes there is no real alternative to relying on the judgment of experts with knowledge of both political and diplomatic processes in the United States and other nations and of technical challenges to reducing emissions.

In applying expert elicitation, as discussed in Appendix C, it would be useful for expert judgments to be informed by historical data and information about the emissions trajectories associated with different levels of climate stabilization. For each scenario of population and GDP and each greenhouse gas considered, the experts could be shown several emissions projections to provide context for their own judgment. For example, they could be shown a trajectory of emissions to 2100 consistent with extrapolation of historical experience. Such a trajectory might be obtained by projecting the historical rate of decline of CO₂ emissions per dollar of real GDP, perhaps modified by the national pledges under the Paris Agreement. They could also be shown as an emissions trajectory consistent with stabilization of CO₂ concentrations at an aggressive target level.

Having seen a range of possibilities, the experts could then be asked to provide their mean emissions projections for 2100 for that scenario, along with quantities designed to enable construction of a probability distribution. A probability density function could be created by combining the experts’ judgments. And then three representative and equally likely emissions levels for 2100 could be created, and emissions trajectories could be derived from them by assuming, for instance, a constant rate of growth. Alternatively, particularly for long-lived gases such as CO₂, it may be better to work with total emissions over the period to 2100 rather than the rate of emissions at that date.

It is less straightforward to determine what useful and credible information about the period beyond 2100 could be provided to the experts. Projections of historical trends would likely be useful, although because of increasing uncertainty about technologies and policies, they are likely to be less useful than for the period to 2100. Emissions projections under

the assumption of strict abatement would also likely be useful. In eliciting judgments for both the periods before and after 2100, allowance would also need to be made for the possibility that net emissions will go to zero, with a range of uncertainty around the dates involved.

These first three steps of the four-step procedure suggested above will yield a relatively small set (e.g., on the order of 27 members) of global population/income/emissions scenarios that are representative of the underlying probability density functions. These results can then be used in the climate module (discussed in Chapter 4) to produce inputs to the damage module.

In the committee’s approach, it is essential that the socioeconomic module pass emissions projections of other climate significant forcers to the climate model. However, because asking experts to produce representative trajectories of other climate forcers for each of nine or more population/income scenarios would be unduly burdensome, simplified procedures are likely to be required. It may be sufficient to ask experts to deal with only a few of the most important forcing agents or only a few extreme scenarios and to use interpolation or other simple methods to produce the desired inputs. Whatever simplified procedures are adopted, however, it would be best if they are based on expert judgments and be clearly described and the rationale for adopting them explicitly presented.

Develop Regional and Sectoral Projections

Damage calculations are likely to require projections of population by region as well as projections of GDP by region and sector. These details will likely be needed in the calibration of aggregate climate damage functions and as inputs to regional and sectoral damage formulations. This is no small task, particularly as one would expect such disaggregated projections to depend on specific global values and be subject to considerable uncertainty.

In this section, the committee considers three approaches to using currently available models and results to develop regional and sectoral projections for the near term. Specifically, the possibility of using scenario libraries, an individual model, or the existing SC-IAMs to develop shares is discussed. Each has advantages and disadvantages, and none of the approaches enables characterization of the uncertainty in the disaggregation step itself. In the case of population, the possibility of using existing regional and national population projections is also discussed.

The first approach would be to estimate median GDP shares for each identified region, over time, using a particular scenario library. As discussed above, the committee does not recommend continuing the IWG procedure of using such scenarios as the basis for global-level projections,

but many scenario libraries do contain internally consistent projections of regional shares, generally to 2100. One can examine a collection of socioeconomic scenario results and derive the population and GDP shares over time of each consistently defined region. This analysis would produce a range of share estimates from which medians could be computed for each region and time period, although such median shares might need to be rescaled to sum to 1.0 (retaining their relative weights). These adjusted median shares could then be applied to the global population and GDP projections to construct regional population and GDP projections.

For example, suppose one of the global scenarios involves a global GDP projection of $200 trillion in 2050. If the rescaled median U.S. and Chinese shares were 20 percent and 25 percent, respectively, in 2050, one would use $40 and $50 trillion as the 2050 projections of U.S. and Chinese GDP. Depending on the breadth of the scenario library, the analysis could be broken into groups of scenarios based on different underlying global population and economic levels, with the above approach applied separately to these groups. This approach would allow the disaggregation to vary across the global projections (as well as over time) in the socioeconomic module. For projections beyond 2100, extension of share projections would be required and need to rely on additional assumptions. A simple choice would be that regional shares remain constant at their 2100 values; alternatively, trends prior to 2100 (e.g., 2080-2100) might be projected to continue in some way.

An advantage of this approach to disaggregation is that it is not tied to any particular model. The median share across models is a robust measure that remains relatively unchanged as individual models are added to or deleted from the analysis. It also provides a potential mechanism to vary disaggregation across global growth projections. However, choices about near-term damage modeling may require regional resolution beyond what is available in scenario libraries, so there is the potential need for additional disaggregation. Larger libraries in particular (e.g., IPCC) tend to have low regional resolution (e.g., five global regions), as well as the sampling issues discussed above.¹⁰ This issue highlights the need to decide which library to use. In addition, using the median share for each region does not ensure consistent shares or shares representative of a single scenario, either across regions or across population and GDP projections, in contrast to shares produced by a single model. Finally, inconsistencies might also arise if the regional shares are coupled with sectoral shares coming from another source. Scenario datasets do not currently provide sectoral disaggregation. Therefore, a different source

__________________

¹⁰ The shared socioeconomic pathways dataset is another resource with 5- and 32-region resolution for some variables, but it is based on a limited number of models and projections.

is required to provide the sectoral detail that may be needed for damage calculations.

The second approach to disaggregation would be to use the baseline projection of an existing detailed-structure economic model. A time profile of regional GDP and population shares could be derived and applied to the global aggregates. Sectoral GDP shares in each region could be constructed on the basis of value added by sector. The same type of extrapolation discussed above would likely be necessary to extend the projections beyond 2100, as such models are typically limited to a 100-year horizon. A key advantage of this approach is that a model with explicitly defined regional and sectoral economic activity ensures consistency (conditional on the model’s structure) among regional and sectoral activity. One disadvantage is that the regional and sectoral detail—while more extensive than the scenario libraries—still may not match the regions and sectors in the damage formulations. Another drawback is that the approach relies entirely on one model, although this disadvantage could be lessened by choosing a model that produces regional shares similar to those in the first approach.

The third approach is to derive shares from the sectoral GDP and/or regional population and GDP assumptions in the existing SC-IAMs or from other models used in the near-term updated damages module. This approach has the advantage of using information already at the appropriate level of disaggregation and properly defined for each damage formulation. The principal disadvantage is a dependency not only on one model, but also on one model of a specific subset of models. In addition, there is no mechanism to vary the path of shares over different global growth paths.

Disaggregated population projections could also be drawn directly from the source of the population projections (e.g., the United Nations, IIASA) as part of the global population projection process. In particular, the United Nations (2015b) provides country and regional probabilistic projections that could be used to develop regional projections consistent with the set of global projections. However, it would then make sense to extract GDP per capita by region from the source of regional economic detail—a scenario library or single model—rather than GDP shares. A time series of regional GDP per capita estimates could then be combined with regional population estimates to produce a new series of regional GDP projections. These projections could then be used to construct shares to disaggregate the global GDP projections. Importantly, this approach would preserve the relative GDP per capita across regions coming from the source economic modeling. This approach would presumably provide more credible disaggregated population projections, but it would require

a much more involved process to couple those projections with disaggregated economic estimates.

It is important to note that most of the approaches discussed here do not simultaneously provide a consistent disaggregation of global GDP and population, match exactly the assumptions and level of disaggregation in the damages module, and rigorously consider how disaggregation is likely to vary over alternative global projections. Moreover, as noted at the outset, most of the approaches do not consider how to model the uncertainty associated with disaggregated results. The longer-term approach discussed below is designed to address these and related issues.

Given the several possible approaches and their various strengths and weaknesses, the IWG will need to compare the options to justify its proposed near-term approach. This involves a choice of how to balance the consistency of the disaggregation, the robustness of multiple models, the alignment with damage module aggregation, and the ability to capture variation across alternate global projections. Furthermore, given the possibility of using multiple damage formulations with different regional and sectoral levels of aggregation, the IWG may need to develop custom approaches for generating disaggregated input projections for different damage formulations.

A LONGER-TERM STRATEGY AND AGENDA FOR RESEARCH AND DEVELOPMENT

Meeting the desired features of the socioeconomic module laid out at the beginning of this chapter is a substantial challenge, though the modifications in procedure recommended in the preceding section would bring the SC-CO₂ framework closer to them. Even with these improvements, however, dependent as they are on scenario libraries and economic models developed for purposes other than SC-CO₂ estimation, shortcomings will remain.

For example, under the approaches suggested above it will be difficult to maintain consistency between regional and sectoral disaggregation of GDP and estimates of emissions, even given assumptions regarding mitigation policies. Also, it is a challenge to ensure consistency between estimates of CO₂ emissions and emissions of other greenhouse gases, such as methane and nitrous oxide. Potential feedbacks are another shortcoming. Monetary damages imply a reduction in economic activity and productive investment, reducing concurrent and future economic performance and affecting emissions net of policy as well. It is an effect illustrated in Figure 2-1 (in Chapter 2), but not considered in the current IWG procedure or in the method described in the preceding section.

In addition, understanding the net damage of climate change may require an elaboration of the four-module structure of Figure 2-1, to take more explicit account of phenomena such as climate effects on biological productivity and land use, changes in regional water availability, and the implications of human adaptation to rising temperatures and all of its associated impacts. And, most challenging, the methods and models used to prepare socioeconomic projections tend to be focused on the current century, whereas projections into subsequent centuries are required for the SC-CO₂ estimation.

In the longer term, there are many advantages to investing in the construction of a dedicated socioeconomic projection framework. Considering its unique objectives, a detailed-structure economic model designed

for the task will likely be the most effective approach in the short run. An existing detailed-structure model might be applied more or less “as is” to this task, as suggested above for near term regional or sectoral disaggregation.

However, such an approach has severe limitations for the longer term. Existing detailed-structure models were formulated to meet very different objectives than those of the IWG. Many of these models support greater sectoral and regional detail than likely is needed or desirable for the SC-CO₂ calculation, and yet they may not yield projections of the particular variables that are needed for climate damage analysis. Feedbacks of some climate impacts have been incorporated in studies using some of these models (e.g., Reilly et al., 2012a), but these were one-time studies of particular effects. The existing models have not been configured for efficient accounting of the wider set of feedbacks that may emerge from a damage module. And, as has been noted, none of the existing detailed-structure models was designed to produce projections beyond 2100, nor does any of them provide a consistent link to other projection methods for the post-2100 period. Hence, although the existing models could play a useful role in the near term, for the longer-term what is needed is a model specifically built for that purpose.

• provides internally consistent probabilistic projections, consistent with elicited expert opinion, as far beyond 2100 as required to capture the vast majority of discounted damages, taking into account the increased uncertainty regarding technology, policies, and social and economic structures in the distant future;
• provides probabilistic regional and sectoral projections consistent with requirements of the damage module, taking into account historical experience, expert judgment, and increasing uncertainty over time regarding the regional and sectoral structure of the global economy;
• captures important feedbacks from the climate and damage modules that affect capital stocks, productivity, and other determinants of socioeconomic and emissions projections.

• It should enable interactions among the modules to ensure consistency among economic growth, emissions, and their consequences; and

• is developed in conjunction with the climate and damage modules, to provide a coherent and manageable means of propagating uncertainty through the components of the SC-CO₂ estimation procedure.

Development of such a framework, designed to satisfy the long-term needs of SC-CO₂ estimation, would represent an advance in economic modeling. Though an effort to build a detailed-structure model suitable for SC-CO₂ estimation could usefully build on one or more existing models, it would be best if supported by a program of research on economic modeling frameworks and model development.

There are costs as well as benefits of the committee’s recommended approach to improved socioeconomic projections. Developing an SC-CO₂ estimation framework with a more tightly integrated socioeconomic module will take time—likely more than the 2-3 years that this report defines as the near term. Thus, some version of, or alternative to, the near-term strategy presented here will need to be used for the next revision of the SC-CO₂, and perhaps for one or more of the subsequent revisions.

In addition to initial model development, continual maintenance will be required to update underlying datasets and incorporate modifications to the SC-CO₂ procedure. Though such a dedicated model could be documented in the peer-reviewed literature, many judgments regarding its use and updating would fall to the IWG itself. It is the view of the committee that such an investment in tools to support SC-CO₂ estimation is warranted.