IPCC Fourth Assessment Report: Climate Change 2007
Climate Change 2007: Working Group III: Mitigation of Climate Change

11.5.3 The costs of mitigation with and without endogenous technological change

Modellers have pursued two broad approaches to endogenizing technological change, usually independently of each other: explicit modelling of R&D activities that contribute to a knowledge stock and reduce costs, and the accumulation of knowledge through learning-by-doing. Sijm (2004) and Edenhofer et al. (2006b) provide detailed comparative assessments of different implementations of both approaches with a focus on mitigation costs when endogenous technology effects are ‘switched on’. Their syntheses provide a useful window for understanding the variation in results and how policies might induce technological change.

Table 11.15: Treatment of endogenous technological change (ETC) in some global top-down integrated assessment models

Study Model ETC channel Number of production sectors Number of regions Major results (impact of ETC) Comments Focus of analysis 

Barker et al., 2006b


E3MG, econometric


LBD and R&D

41 20 

Cumulative investments and R&D spending determine energy demand via a technology index. Learning curves for energy technologies (electricity generation). Cumulative investments and R&D spending determine exports via a technology index.


Econometric model. Investments beyond baseline levels trigger a Keynesian multiplier effect.

Sectoral R&D intensities stay constant overtime


Long-term costs of stabilization

Income and production losses


Bollen, 2004


WorldScan CGE


R&D R&D (and occasionally LBD)

12 12 

ETC magnifies income losses.


Includes international spillovers.

No crowding-out effect


Compliance costs of Kyoto protocol



et al., 2003


FEEM-RICE optimal growth


R&D and LBD


Direct abatement costs are lower, but total costs are higher.

ET ceilings have adverse effects on equity and efficiency.


Factor substitution in Cobb-Douglas production.


Impact of emissions trading (+ restrictions)


Bosetti et al., 2006






An index of energy technological change increases elasticity of substitution. Learning-by-doing in abatement and R&D investments raise the index. Energy technological change explicitly decreases carbon intensity.



Experimental model exploring high inertia.


Crassous et al., 2006




R&D and LBD


Cumulative investments drive energy efficiency. Fuel prices drive energy efficiency in transportation and residential sector. Learning curves for energy technologies (electricity generation).


Endogenous labour productivity, capital deepening.




et al., 2006a



Optimal growth




R&D investments improve energy efficiency. Factor substitution in a constant-elasticity-of-substitution (CES) production function. Carbon-free energy from backstop technologies (renewables) and CCS. Learning-by-doing for renewable energy. R&D investments in labour productivity. Learning-by-doing in resource extraction




Gerlagh and Van der Zwaan, 2003



Optimal growth




Costs are significantly lower. Transition to carbon-free energy. Lower tax profile. Early abatement


Results are sensitive to elasticity of substitution between technologies as well as to the learning rate for non-carbon energy


Optimal tax profile

Optimal abatement profile

Abatement costs


Gerlagh, 2006






Factor substitution in CES production. Carbon-free energy from renewables and CCS. Learning-by-doing for both and for fossil fuels.




Goulder and Mathai, 2000


Partial cost-function model with central planner





Lower time profile of optimal carbon taxes. Impact on optimal abatement varies depending on ETC channel. Impact on overall costs and cumulative abatement varies, but may be quite large



One instrument

High aggregation

Weak database


Optimal carbon tax profile

Optimal abatement profile


Goulder and Schneider, 1999


CGE multisectoral model




Gross costs increase due to R&D crowding-out effect.

Net benefits decrease.


Lack of empirical calibration

Focus on U.S.

Full ‘crowding-out’ effect


Abatement costs and benefits


Kverndokk et al., 2004


CGE model for a small open economy




Innovation subsidy is more important in the short term than a carbon tax.

Innovation subsidy may lead to ‘picking a winner’ and ‘lock in’


Numerical illustrative model


Optimal timing and mixture of policy instruments

Welfare effects of technology subsidies


Masui et al., 2006


AIM/Dynamic - Global




Factor substitution in CES production. Investments in energy conservation capital increase energy efficiency for coal, oil, gas and electricity. Carbon-free energy from backstop technology (nuclear/renewables).



Focus on energy efficiency with limited supply-side substitution.


Nordhaus, 2002


R&DICE optimal growth




ETC impact is lower than substitution impact and quite modest in early decades.



Full ‘crowding-out’ of R&D

High aggregation


Factor substitution versus ETC

Carbon intensity

Optimal carbon tax


Popp, 2004


ENTICE, optimal growth




Impact on cost is significant. Impact on emissions and global temperature is small


Partial crowding-out effect


Welfare costs

Sensitivity analysis of R&D parameters


Popp, 2006a






Factor substitution in Cobb-Douglas production. R&D investments in energy efficiency knowledge stock. Carbon-free energy from generic backstop technology


R&D investments lower price of energy from backstop technology.



Rao et al., 2006






Carbon-free energy from backstop technologies (renewables, carbon scrubbing & sequestration). Learning curves for electricity generation and renewable hydrogen production


Factor substitution in CES production in MACRO.



Rosendahl, 2004


Builds on Goulder and Mathai (2000)




Restrictions on emissions trading are cost-effective. Optimal carbon tax in Annex I region is increased with external spillovers


Outcomes are sensitive to learning rate, discount rate and slope of abatement curve


Optimal carbon tax (or permit price) over time in two regions

Optimal emissions trading +restrictions


Note: See sources for details of models.

Sources: The table is derived from Sijm (2004) and Edenhofer et al. (2006b).

In his review, Sijm (2004) distinguishes top-down models that mostly focus on explicit R&D effects, and bottom-up models that focus mostly on LBD effects. Among the top-down models, which are described in Table 11.15, he finds considerable variation in the effect of including Endogenous Technological Change (ETC). While some models find a large reduction in mitigation costs (e.g. Popp, 2006a), some find small impacts (e.g. Nordhaus, 2002). These differences can be attributed to:

  • the extent of substitution allowed of low-carbon fuels for high-carbon fuels. When this factor is included, the reduction in costs is more pronounced, and the higher it is, the greater the reduction.
  • the degree of ‘crowding-out’ associated with energy R&D expenditures. If new energy R&D is assumed to be in addition to existing R&D, this will generate larger reductions in mitigation than if new energy R&D is assumed to lead to a reduction in R&D elsewhere.
  • the approach to spillover. In addition to justifying higher rates of return from R&D, spillover implies that the market outcome with too little investment could be improved by policy intervention.
  • the degree of differentiation among R&D activities, the assumed rates of return from those activities, and the capacity of R&D activities to lower costs for low-carbon technologies.
  • the rate of learning if LBD is included. Higher rates imply larger reductions in mitigation costs with ETC included.

The first point is that the way low-carbon and high-carbon energy are treated in the models –whether as complements or substitutes – is critical is determining the flexibility of the model to low-carbon innovation and costs of mitigation. Models that do not allow high levels of substitution between low-carbon and high-carbon energy (Goulder and Mathai, 2000; Nordhaus, 2002; Popp, 2006b) indicate that R&D has less impact than those that do, e.g. by introducing a carbon-free backstop technology (Gerlagh and Lise, 2005; Popp, 2006b). Similar results are found more widely for LBD and R&D models: the more substitution possibilities allowed in the models, the lower the costs (Edenhofer et al., 2006a, p.104).

When providing evidence to support the second point – the studies of induced R&D effects via the stock of knowledge – Goulder and Schneider (1999), Goulder and Mathai (2000), Nordhaus (2002), Buonanno et al. (2003) and Popp (2004) differ considerably about the extent of crowding-out. In other words, does R&D have an above-average rate of return and does an increase in R&D to support the carbon-saving technologies come from ordinary production activities (no crowding-out), or equally valuable R&D in other areas (crowding-out)? Nordhaus (2002) assumes complete crowding-out in which carbon-saving R&D has a social rate of return that is four times the private rate of return but, because it is assumed that it replaces other equally valuable R&D activities, it costs four times as much as conventional investment. At the other extreme, Buonanno et al. (2003) consider spillovers that lead to similarly high social rates of return, but without the higher opportunity costs. Not surprisingly, Nordhaus finds very modest mitigation cost savings and Buonanno et al. find enormous savings. In general, induced technological change in a general-equilibrium framework has its own opportunity costs, which may reduce the potential for cost reduction in CGE models substantially.

Popp (2006b), in turn, suggests on the basis of the empirical evidence that half of the R&D spending on energy in the 1970s and 1980s took place at the expense of other R&D. Something between full and partial crowding-out appears more recently in Gerlagh and Lise (2005). Goulder and Matthai (2000) provide an example of the importance for cost reduction of parameters describing returns from R&D and capacity for innovation. They compare both R&D as new knowledge and learning-by-doing (LBD), finding a 29% reduction in the marginal costs with R&D by 2050 and 39% with LBD. As they note, however, this reflects the calibration of their model to a 30% cost saving based on Manne and Richels’ assumptions (1992). The model results simply reflect the choice of calibrated parameter values.

By contrast with the results for top-down models, Sijm (2004) finds considerably more consistency among bottom-up models, where the effects of learning-by-doing typically reduce costs by 20% to 40% over the next half century, and by 60% to 80% over the next century. Importantly, however, these numbers are relative to a static technology alternative. To demonstrate the influence of this assumption, van Vuuren et al. (2004) run their model without a carbon constraint, but with learning, to identify a baseline level of technological change. Their approach roughly halves the estimated effect of ETC on mitigation.

The variations in the estimated effects of learning on costs in bottom-up models are driven primarily by variations in the assumed rate of learning (in other words, the extent of the reduction in costs for each doubling of installed capacity). Estimates of these rates vary, depending on whether they are assumed or econometrically estimated, and whether they derive from expert elicitation or historical studies. These learning rates vary between four leading models by as much as a factor of two for a given technology, as shown in Table 11.16.

Table 11.16: Learning rates (%) for electricity generating technologies in bottom-up energy system models

 (a) One-factor learning curves (b) Two-factor learning curves 
Advanced coal 11 
Natural gas combined cycle 10 11 11 15 24 11 
New nuclear 
Fuel cell 18 13 19 19 11 19 11 
Wind power 11 12 15 16 12 
Solar PV 18 19 19 28 25 10 19 10 


Learning rates are defined as the percentage reduction in unit cost associated with a doubling of output. The acronym LDR stands for Learning-by-Doing Rates and LSR for Learning-by-Searching Rates in two-factor learning curves. In two-factor learning curves, cumulative capacity and cumulative R&D (or ‘knowledge stock’) are used to represent market experience (learning-by-doing) and knowledge accumulated through R&D activities, respectively.

In MERGE-ETL, endogenous technological progress is applied to 8 energy technologies: six power plants (integrated coal gasification with combined cycle, gas, turbine with combined cycle, gas fuel cell, new nuclear designs, wind turbine and solar photovoltaic) and two plants producing hydrogen (from biomass and solar photovoltaic). Furthermore, compared to the original MERGE model, Bahn and Kypreos (2002; 2003a) have introduced two new power plants (using coal and gas) with CO2 capture and disposal in depleted oil and gas reservoirs. Like the MARKAL model, the ERIS model is a bottom-up energy system model. Both studies mentioned above cover six learning technologies. MESSAGE is also a bottom-up system engineering model. Like the other bottom-up energy system models, it determines how much of the available resources and technologies are used to satisfy a particular end-use demand, subject to various constraints, while minimizing total discounted energy system costs.

For a review of the literature on learning curves, including 42 learning rates of energy technologies, see McDonald and Schrattenholzer, (2001).

For a discussion and explanation of similar (and even wider) variations in estimated learning rates for wind power, see Söderholm and Sundqvist (2003) and Neij et al. (2003a; 2003b)

Sources: Sijm (2004), Messner (1997), Seebregts et al. (1999), Kypreos and Bahn (2003), and Barreto and Klaassen (2004), Barreto (2001), Barreto and Kypreos (2004), and Bahn and Kypreos (2003b).

The modelling of LBD is beset with problems. Model solutions become more complex because costs can fall indefinitely, depending on the extent of the market. Avoidance of multiple solutions typically requires the modeller to constrain the penetration of new technologies, making one element of the cost reduction effectively exogenous. Since many low-carbon technologies are compared with mature energy technologies early in the learning process, it becomes inevitable that their adoption spreads and that they eventually take over as carbon prices fall. Finally, the approach often assumes that diffusion and accompanying R&D are cost-free, although the investments required for the technologies with high learning rates may be comparable with those that are replaced.

In addition, the measurement of learning rates poses econometric problems. It is difficult to separate the effects of time trends, economies of scale and R&D from those of LBD (Isoard and Soria, 2001) and different functional forms and data periods yield different estimates, so the learning rates may be more uncertain than suggested by their treatment in the models. When controls for the effects of other variables are included, such as crowding-out effects, the influence of LBD in some models may become very small compared to the effect of R&D (Köhler et al., 2006; Popp, 2006c).

A second survey of ETC effects on aggregate mitigation costs comes from the Innovation Modelling Comparison Project (IMCP) (Edenhofer et al., 2006b). Rather than reviewing previous results, the IMCP engaged modelling teams to report results for specific concentration scenarios and, in particular, with and without ETC. Like the van Vuuren et al. (2004) study noted earlier, the IMCP creates a baseline technology path with ETC but without an explicit climate policy. This baseline technology path can then be either fixed, as autonomous technological change, or allowed to change in response to the climate policy.

Table 11.15 also summarizes the treatment of technological change in the IMCP models; in principle, the wide range of approaches provides additional confidence in the results when common patterns emerge. Like Sijm (2004), Edenhofer et al. (2006b) find that, while ETC reduces mitigation costs, there continues to be a wide range of quantitative results: some are close to zero and others generate substantial reductions in costs.

Figure 11.9 shows the effects of introducing ETC into the models for the 550 and 450 ppm CO2 stabilization scenarios 2000–2100. The reductions in carbon prices and GDP are substantial for many studies in both stabilization cases when ETC is introduced. The effects on CO2 reductions show that including ETC in the models leads to earlier reductions in emissions. It should be noted that the reduction of costs in IMCP models is not mainly driven by LBD. The assumptions about the crowding-out of conventional R&D by low-carbon R&D and the availability of mitigation options (models have different sets of options) are more important factors determining costs and mitigation profiles than LBD (Edenhofer et al., 2006b, p.101–104). One major research challenge is to test the influence of these aspects of ETC on current technologies by econometric and backcasting methods, fitting the models to historical data.

Figure 11.9

Figure 11.9: Averaged effects of including ETC on carbon tax rates, CO2 emissions and GDP: 9 global models 2000–2100 for the 450 ppm and 550 ppm CO2-only stabilization scenarios

Notes: The figures show the simple averages of results from 9 global models 2000–2100 for (a) carbon tax rates and emission permit prices in US$(2000)/tCO2, (b) changes in CO2 (% difference from base) and (c) changes in global GDP (% difference from base). The results are shown with and without endogenous technological change. The grey background lines show the range from the individual models for 450 ppm with ETC. See source for details of models.

Source: adapted from Edenhofer et al. (2006b).

Figure 11.9 emphasizes the range and the uncertainty of the results for induced technological change[16] from climate policies. The potential of ETC to reduce mitigation costs varies remarkably between different model types. For a 450 ppm CO2-only concentration stabilization level at the upper end of the range, including ETC in the model reduces mitigation costs by about 90%, but at the lower end it makes no difference (Edenhofer et al. 2006b, p. 74). The averages also somewhat exaggerate the effects of ETC because there are other assumptions that affect the costs, as evident in a meta-analysis of the macro-economic costs of mitigation undertaken for the UK Treasury’s Stern Review (Barker et al., 2006b). An example is the use of tax/permit revenues, as discussed in 11.4.4 above. This study combines the IMCP results on costs with earlier data on post-SRES scenarios (Repetto and Austin, 1997; Morita et al., 2000) so that the effects of other assumptions can be identified. The average effects of including ETC in the IMCP models by 2030 for pathways to 550 and 450 ppm CO2-only are reduced from 1.1 and 2.7% of global GDP compared to baseline, as shown in Figure 11.9, to 0.4 and 1.3% respectively using the full equation of the meta-analysis, which allows for individual model outliers, time and scenario effects as well as the approaches and assumptions adopted by the modellers. In other words, allowing for technologies to respond to climate policies reduces the GDP costs of Category III stabilization, as estimated by the IMCP models, by 1.3% by 2030. Costs across models of 2.1% without ETC, but allowing for emissions trading and backstop technologies, are reduced to 0.8% GDP by 2030 with ETC. The ETC effects become more pronounced in the reduction of costs for later years and as the stabilization targets become more stringent, partly due to the associated extra increases in the required carbon prices.

Edenhofer et al. conclude that the results for effects of ETC depend on:

  • baseline effects: baseline assumptions about the role of technology that generate relatively low emission scenarios can leave little opportunity for further ETC effects;
  • the assumption of the inefficient use of resources in the baseline (distinct from market failure associated with greenhouse gas emissions and climate change): this provides opportunities for policy to improve otherwise inefficient private decisions and may even raise welfare. Spillovers were an example of this in the Sijm (2004) discussion; some simulations also include inefficient energy investment decisions.
  • how the investment decision is modelled: recursive savings decisions, as opposed to foresight and intertemporal opti-mization, provide less opportunity for investment and R&D to expand. In the Sijm (2004) context, less responsiveness in aggregate investment and R&D implies more crowding-out.
  • the modelling of substitution towards a backstop techno-logy (such as a carbon-free energy source available at constant, albeit initially high, marginal cost): this can substantially affect the results. For example, if investment in the technology is endogenous and involves learning-by-doing, costs can fall dramatically. Popp (2006a, p.168) goes further, and shows that the addition of a backstop technology by itself can have a larger effect on mitigation costs than the addition of LBD. These results are also confirmed by the IMCP study (Edenhofer et al., 2006b, p.214). However, investment in backstop technologies requires time-consistent policies (Montgomery and Smith, 2006). It is therefore debatable to what extent the indicated potential for cost reduction can be realized under real-world conditions where a global, long-term and time-consistent climate policy has yet to be implemented.
  1. ^  When a model includes ETC, further change can generally be induced by economic policies. Hence the term ‘induced technological change’ (ITC); ITC cannot be studied within a model unless it simulates ETC. See Glossary on ‘technological change’.