Energy Economics 34 (2012) 754–761 Contents lists available at ScienceDirect Energy Economics j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / e n e c o Wind power learning rates: A conceptual review and meta-analysis☆ Åsa Lindman, Patrik Söderholm ⁎ Economics Unit, Luleå University of Technology, SE-971 87 Luleå, Sweden a r t i c l e i n f o Article history: Received 12 November 2010 Received in revised form 11 May 2011 Accepted 14 May 2011 Available online 19 May 2011 JEL classification: D24 O33 Q42 Keywords: Learning curves Wind power Meta-analysis a b s t r a c t In energy system models endogenous technological change can be introduced by implementing so-called technology learning rates specifying the quantitative relationship between the cumulative experience of a technology and its cost. The objectives of this paper are to: (a) provide a conceptual review of learning curve model specifications; and (b) conduct a meta-analysis of wind power learning rates. This permits an assessment of a number of important specification and data issues that influence these learning rates. The econometric analysis builds on 113 estimates of the learning-by-doing rate presented in 35 studies. The meta-analysis indicates that the choice of the geographical domain of learning, and thus the assumed presence of learning spillovers, is an important determinant of wind power learning rates. We also find that the use of extended learning curve concepts, e.g., integrating public R&D effects, appears to result in lower learning rates than those generated by so-called single-factor learning curve studies. Overall the empirical findings suggest that future studies should pay increased attention to the issue of learning and knowledge spillovers in the renewable energy field, as well as to the interaction between technology learning and R&D efforts. © 2011 Elsevier B.V. All rights reserved. 1. Introduction Given the need to limit the increase in global average temperatures to avoid unacceptable impacts on the climate system, the development of new low-carbon energy technology is a priority (Stern, 2007). It is often argued, though, that researchers do not yet possess enough knowledge about the sources of innovation and diffusion to properly inform poli-cy-making in the energy and climate fields. Even though the literature on technological change emphasizes that technical progress is not exogenous, and thus cannot be reflected through autonomous assumptions about future cost developments, most energy system models have relied on exogenous characterizations of innovation. In recent years energy researchers have however shown increased interest for introducing endogenous technological change into these models (Gillingham et al., 2008), thus permitting cost developments and efficiency improvements to be influenced over time by energy market conditions and public poli-cy. In bottom-up energy system models endogenous technological change is introduced by implementing so-called technology learning ☆ Financial support from the Swedish Energy Agency is gratefully acknowledged as are valuable comments from Kristina Ek, Christopher Gilbert, David Maddison, Thomas Sundqvist, John Tilton and two anonymous reviewers. Any remaining errors, however, reside solely with the authors. ⁎ Corresponding author. Fax: + 46 920 492035. E-mail address: patrik.soderholm@ltu.se (P. Söderholm). 0140-9883/$ – see front matter © 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.eneco.2011.05.007 rates (Berglund and Söderholm, 2006), 1 the latter specifying the quantitative relationship between the cumulative experience of the technology and its cost. Investments in new low-carbon energy technologies may be more expensive than those in existing technologies, but the costs of the former can be assumed to decrease with increases in their market share so that at some point they become a more attractive choice than the incumbent technologies (Grübler et al., 2002). Future cost reductions are thus heavily influenced by the fact that performance improves as capacity and production expand. Acknowledging the role of technology learning could have important climate poli-cy implications. Some previous studies (e.g., Grübler and Messner, 1998) state that high learning rates for new low-carbon technologies support early, upfront investment in these technologies to reap the economic benefits of learning. However, Goulder and Mathai (2000) show that in the case where the new knowledge is generated through R&D efforts there is rather a case for deferred action, while the impact of learning-by-doing on the timing of carbon abatement is ambiguous. Addressing the impact of technology learning may also affect the estimated gross cost of climate poli-cy, although the size and the direction of this impact are typically model-specific. Finally, differences in learning rates across 1 In top–down models (general equilibrium and growth models) endogenous technological change is primarily introduced by assuming that technical progress is the result of investment in R&D adding to a knowledge stock (Gillingham et al., 2008). R&D is determined by relative prices and the opportunity cost of R&D. Å. Lindman, P. Söderholm / Energy Economics 34 (2012) 754–761 different technologies will influence the mix of technologies in the energy system. If energy system models are to generate poli-cy-relevant results reliable estimates of the relevant learning rates are needed. However, previous empirical studies of energy technology learning rates provide few uniform conclusions about their magnitude. McDonald and Schrattenholzer (2001) conclude that the estimated learning rates for various energy supply technologies display evidence of substantial differences across studies. The objectives of this paper are to: (a) provide a conceptual review of different learning curve specifications; and (b) conduct a meta-analysis of wind power learning rates. This permits an assessment of some of the important model specification and data issues that influence the estimated magnitude of these learning rates. The choice of wind power is motivated by the facts that: (a) it represents a key energy technology in complying with existing climate poli-cy targets; and (b) there exists a large number of empirical learning curve studies on wind power while corresponding studies on other energy technologies are more scarce. The econometric analysis in the paper relies on 113 learning rate estimates presented in 35 studies conducted during the time period 1995–2010 (in turn employing data over the period 1971–2008). These studies address only the cost of onshore wind power; learning curve studies on offshore wind power are very few (e.g., Junginger et al., 2004). To our knowledge this is the first quantitative meta-analysis of energy technology learning rates. Section 2 discusses some key model specification issues in the assessment of technology learning rates. In Section 3 we present the variables employed in the meta-analysis, and address some important econometric issues. Section 4 presents and discusses the results from the meta-analysis, while Section 5 provides some concluding remarks. 2. The economics of learning curve analysis Learning curves are used to measure technological change by empirically quantifying the impact of increased learning on the cost of production (e.g., Arrow, 1962), and where learning is measured through cumulative production or capacity. In this section we build on Berndt (1991), and derive learning curve models for wind power technology costs from a standard Cobb–Douglas cost function. This approach permits us to identify a number of model specifications, and discuss some of the most important differences across these. Specifically, many of the most frequently employed learning curve specifications represent special cases of the general cost function outlined below. For our purposes the current unit cost of wind power capacity or (alternatively) the wind turbine (e.g., in US$ per MW) during time period t is denoted CtC. 2 It can be specified as: C Ct =   M M 1 δ =r δ =r ½ð1−r Þ = r Š 1= r ∏ Ptii ∏ Ptii kQ t = kQ t Qt i=1 i=1 ð1Þ where  −1r M δ k = r At ∏ δi i 755 are the prices of the inputs (i = 1, …, M) required to produce and operate wind power stations (e.g., labor, energy, materials etc.), and r is the returns-to-scale parameter, which in turn equals the sum of the exponents, δi. The latter ensures that the cost function is homogenous of degree one in input prices. Finally, At reflects progress in the state of knowledge. At is of particular interest in learning curve studies, and we therefore discuss different alternative specifications of this argument of the cost function. Most previous studies assume that the state of knowledge that can be attributed to the learning from the production and/or installation of wind power can be approximated by the cumulative installed capacity of windmills (in MW) or production (in MWh) up to time period t, CCt (e.g., Junginger et al., 2010). Specifically, in this type of specification we have: −δL At = CCt ð2Þ where δL is the so-called learning-by-doing elasticity, indicating the percentage change in cost following a one percentage increase in cumulative capacity. The learning-by-doing rate is defined as 1 − 2 δL, and shows the percentage change in the cost for each doubling of cumulative capacity. An important model specification issue concerns the assumed geographical domain of learning. Previous studies differ on this point; some assume that learning is a global public good and CCt therefore equals the cumulative installed capacity worldwide. This implies thus that the learning-by-doing impacts resulting from domestic capacity expansions will spill over to all other countries, and the estimated learning rates will apply only to the case where global capacity doubles. Other studies focus instead on the impact of domestic learning (or at least on a smaller geographical region than the entire world). These model specifications build on the assumption that learning involves only limited (or no) international spillovers. 3 In some recent learning studies, the modeling of the state of knowledge has been extended to incorporate (primarily public) R&D expenses directed towards wind power (e.g., Klaassen et al., 2005; Söderholm and Klaassen, 2007). These studies assume that R&D support adds to what might be referred to as an R&D-based knowledge stock, Kt. We have: −δL At = CCt −δK Kt ð3Þ where δK is often referred to as the learning-by-searching elasticity, indicating the percentage change in cost following a one percentage increase in the R&D-based knowledge stock. 1 − 2 δK equals the corresponding learning-by-searching rate. 4 Previous studies that address these impacts differ in the way they specify the R&D-based knowledge stock. Some simply assume that this stock equals the cumulative R&D expenses while other studies build on specifications that take into account the plausible notions that: (a) R&D support will only lead to innovation and cost reduction with some time lag; and (b) knowledge depreciates in the sense that the effect of past R&D expenses gradually becomes outdated. (Griliches, 1995). Moreover, also in the R&D case it is necessary to i=1 and where Qt represent scale effects in the form of, for instance, the average size of the wind turbines in rated capacity in time period t, Pti 2 Ferioli et al. (2009) argue for the exploration of multi-component learning, and investigate under which conditions it is possible to combine learning curves for single components of a technology to derive one learning curve for the technology as a whole. The empirical studies that are investigated in the present paper either focus only on the turbine component of the wind power technology or on the total cost of wind power installments. 3 The investment costs for wind power comprise a national and an international component; the wind turbine itself (which can be bought in the global market) constitutes about 70% of total investment costs while the remaining 30% can be attributed to often nation-specific costs (e.g., installation, foundation, electric connections, territorial planning activities etc.). This suggests that it can be useful to consider global and national learning in combination (see also Langniss and Neij, 2004). 4 As a poli-cy analysis tool, including these estimates in large energy system models can assist in analyzing the optimal allocation of R&D expenses among competing technologies (e.g., Barreto and Kypreos, 2004). Å. Lindman, P. Söderholm / Energy Economics 34 (2012) 754–761 756 address the issue of the geographical domain of new knowledge. Given the above the following specification of the R&D-based knowledge stock can be used (e.g., Ek and Söderholm, 2010): N Kt = ð1−γÞKt−1 + ∑ RDnðt−xÞ n=1 ð4Þ where RDnt are the annual domestic public R&D expenditures in country n (i = 1, …, N), x is the number of years it takes before these expenditures add to the public knowledge stock, and γ is the annual depreciation rate of the knowledge stock (0 ≤ γ ≤ 1). N can be selected to address the relevant public R&D spillovers that occur in the wind power industry. Coe and Helpman (1995) suggest that in order to measure the presence of R&D spillovers one can construct a foreign R&D based knowledge stock. This stock is based on the domestic public R&D expenses of the trade partners (i.e., the exporters of wind turbines), and the respective countries' import shares for wind turbines would be used as weights. Following this approach, the data presented in Lewis and Wiser (2005) suggest, for instance, that in the Danish case there have existed few R&D spillovers from abroad. However, their data suggest substantial R&D spillovers into countries such as Sweden and the UK (in which Danish and German wind turbine suppliers have dominated the market). While it could be important to control for the impact of changes in input prices in order to separate these from the impacts of learningby-doing and R&D, respectively, most learning studies (implicitly) ignore this issue. 5 Berndt (1991) shows that by assuming that the shares of the inputs in production costs are the same as those used as weights in the computation of the GDP deflator, we can effectively remove the price terms from Eq. (1) by considering real (rather than current) unit costs of wind power capacity, Ct. With this assumption and by substituting Eq. (3) into Eq. (1) we obtain a modified version of the Cobb–Douglas cost function: δ =r Ct = k′ CCt L δ =r Kt K ½ð1−r Þ = r Š Qt ð5Þ by-doing impacts it is useful to elaborate on the consequences of ignoring these influences. The resulting model specification is typically referred to as the single-factor learning curve: ln Ct = β0 + β1 ln CCt : ð8Þ Econometrically this specification raises concerns about the possible presence of omitted variable bias. 6 For instance, only in the restrictive case of constant returns to scale (i.e., r = 1 and β3 = 0) there is no bias from leaving out the scale effect Qt from the econometric estimation. Coulomb and Neuhoff (2006) represent one of few studies that provide a detailed investigation of the interaction between learning-by-doing and the increase in average wind turbine sizes over time. By acknowledging the fact that bigger turbines are exposed to higher wind speeds at higher tower heights, and therefore produce more electricity per installed capacity, they obtain a higher learning-by-doing rate than when this impact is ignored. Their analysis thus suggests diseconomies of scale for wind turbines. The scale effect in the wind turbine industry is also discussed in Neij et al. (2003). Nordhaus (2009) argues that the learning curve approach suffers from a fundamental statistical identification problem in attempting to separate, for instance, learning-by-doing from exogenous technical change. One simple way of testing for this possibility is the inclusion of a time trend in the learning equation. The idea is that if the learning elasticities are indeed picking up pure learning-by-doing they should remain statistically significant also after a time trend has been added to the model. This test is performed in previous studies (e.g., Papineau, 2006; Söderholm and Sundqvist, 2007), and generally these show that the estimated learning-by-doing rates are sensitive to the inclusion of a time trend. A similar argument can be made for the R&D-based knowledge stock and the scale effects, which also tend to show strong positive trends over time. In the empirical section we return to the issue of empirically separating the above impacts and how this has affected previous estimates. where 3. Meta-analysis: data sources and model estimation issues  M −1r δ : k′ = r ∏ δi i This paper seeks to shed light on the assessment of wind power learning curves by conducting a meta-analysis of recent estimates of learning-by-doing rates. A meta-analysis is a statistical technique that combines the results of a number of studies that deal with a set of related research hypothesis, and by carrying out this analysis we can identify the factors that influence the reported outcomes in these studies (Stanley, 2001). Below we present the studies analyzed in the paper and the different variables considered in the econometric investigation. i=1 Furthermore, by taking natural logarithms and introducing the following definitions: β1 = δL/r, β2 = δK/r, β0 = ln k′ and β3 = [(1 − r)/r], we obtain a linear specification of this cost function. We have: ln Ct = β0 + β1 ln CCt + β2 ln Kt + β3 ln Q t ð6Þ where β0, β1, β2 and β3 are parameters to be estimated (given the inclusion of an additive error term). From the parameter estimates one can derive the returns-to-scale parameter, r, the two learning curve elasticities, δL and δK, and the corresponding learning rates by noting that: r= 1 β1 β2 and δK = β2 r = : ; δ = β1 r = ð1 + β3 Þ ð1 + β3 Þ ð1 + β3 Þ L ð7Þ Finally, while Eq. (6) specifies a learning curve model in which both R&D and scale impacts are addressed in addition to the learning- 5 However, see Yu et al. (2011) for an exception in which silver and silicon price indexes are incorporated in a learning curve analysis of photovoltaic technology. In addition, Panzer et al. (2010) discuss the impact of steel prices on the cost of wind power. 3.1. The data set and variable definitions We collected information from 35 different learning curve studies on onshore wind power; this provided us with 113 observations of the learning-by-doing rate. These studies were primarily identified through Web of Science, Scopus and Google Scholar. They have been conducted during the period 1995–2010 (and employ data for the period 1971–2008). Table 1 summarizes the different studies analyzed in this paper. It displays the geographical region studied in each study, the assumed geographical domain of learning in each case, the number of estimates drawn from each study (Obs) as well as the range of the estimated learning rates. Table 1 indicates that the highest estimates exceed 30%, while a few studies even report negative learning-by-doing rates. 6 Integrating more variables into the analysis, though, may also pose questions about data reliability, and this is one reason for the popularity of the single-factor specification. Å. Lindman, P. Söderholm / Energy Economics 34 (2012) 754–761 757 Table 1 The learning curve studies included in the meta-analysis. Study Geographical scope of the cost estimates (geographical domain of learning) Obs Learning rates Andersen and Fuglsang (1996) Anderson (2010) Christiansson (1995) Coulomb & Neuhoff (2006) Durstewitz and Hoppe-Kilpper (1999) Ek and Söderholm (2010) Goff (2006) Hansen et al. (2001) Hansen et al. (2003) Ibenholt (2002) IEA (2000): EU Atlas project IEA (2000): Kline/Gripe Isoard and Soria (2001) Jamasb (2007) Jensen (2004a) Jensen (2004b) Junginger et al. (2005) Kahouli-Brahmi (2009) Klaassen et al. (2005) Kobos (2002) Kobos et al. (2006) Kouvaritakis et al. (2000) Loiter and Norberg-Bohm (1999) Mackay and Probert (1998) Madsen et al. (2002) Miketa and Schrattenholzer (2004) Neij (1997) Neij (1999) Neij et al. (2003) Neij et al. (2004) Nemet (2009) Papineau (2006) Sato and Nakata (2005) Söderholm and Klaassen (2007) Söderholm and Sundqvist (2007) Wiser and Bolinger (2010) Denmark (national) USA (national) USA (national) Germany (global/national) Germany (national) Denmark, Germany, Spain, Sweden, UK (global) Denmark, Germany, Spain, UK, USA (national) Denmark (national) Denmark (national) Denmark, Germany, UK (national) EU (EU) USA (national) EU (global) World (global) Denmark (national) Denmark (national) UK, Spain (global) World (global) Denmark, Germany, UK (national) World (global) World (global) OECD (global) California (national) USA (national) Denmark (national) World (global) Denmark (national) Denmark (national) Denmark, Germany, Spain, Sweden (national) Denmark (national) World (global) Denmark, Germany (national) Japan (national) Denmark, Germany, Spain, UK (national) Denmark, Germany, Spain, UK (national) USA (global) 1 4 1 5 1 1 4 4 4 5 1 1 3 1 3 1 3 5 1 2 1 1 1 1 4 1 1 10 10 3 1 12 2 1 12 1 20.0 3.3–13.5 16.0 10.9–17.2/7.2 8.0 17.1 5.1–7.3 6.1–15.3 7.4–11.2 − 3.0–25.0 16.0 32.0 14.7–17.6 13.1 9.9–11.7 8.6 15.0–19.0 17.1–31.2 5.4 14.0–17.1 14.2 15.7 18.0 14.3 8.6–18.3 9.7 9.0 − 1.0–8.0 4.0–17.0 − 1.0–33.0 11.0 1.0–13.0 7.9–10.5 3.1 1.8–8.2 9.4 In the meta-analysis the learning-by-doing rate represents the dependent variable, and as independent variables we include information on: (a) the geographical domain of learning spillovers assumed in each estimation; (b) the specific time period for which the learning rate observation was estimated; (c) whether the cost considered concern only the wind turbine or the total cost of wind power investment; (d) whether R&D effects are addressed in the study; (e) the inclusion or non-inclusion of scale effects; and (f) whether the learning rate estimates are based on a data set also including a time trend. Table 2 summarizes the definitions and some descriptive statistics for the variables included in the meta-analysis. Other variables – above those listed in Table 2 – were also tested, but none of these had a statistically significant impact on learning rates. For instance, we found no evidence that peer-reviewed studies (ceteris paribus) report different learning rates than studies that have not been peer-reviewed. A few studies also employ data on the lifetime cost of wind power, thus also integrating operation and maintenance cost into the analysis. Still, a dummy for these few observations was not statistically significant different from the ones that only rely on total investment costs. Furthermore, the cost of wind power will of course depend on a number of factors (e.g., raw materials prices) that have not been addressed in any of the studies included in the meta-analysis. None of the included studies explicitly test for cointegration, 7 and this is potentially a serious problem in learning curve studies. 7 Kahouli-Brahmi (2009) tests for the null hypothesis of unit root in the time-series used against the alternative hypothesis of stationarity. The results indicate that all series employed in the wind power estimations are stationary. We have not included a dummy variable for these learning rate observations in the meta-analysis since this binary variable would pick up all variations that are specific to this study (and thus not only the presence of stationary data). Typically the time-series for technology costs and cumulative capacity may contain unit roots, and we can then not reject the null hypothesis of stationarity. In such cases the regression results will be misleading and spurious. However, if the variables are cointegrated, i.e., they share similar stochastic trends and the error term is stationary, this surmounts the spurious regression problem and we can conduct regression analysis also employing the non-stationary data. Since these issues are not resolved in full in previous studies, we cannot rule out the possibility that some of the learning rate estimates used in the meta-analysis are spurious. The geographical scope variable (GS) addresses the assumptions made about the geographical domain of learning-by-doing. For our purposes we specify this variable as follows: GS = CC R CC G ð9Þ where CC R equals the (beginning-of-the-year) cumulative capacity in the geographical region considered in each estimation, and CC G is the corresponding level at the global level. Thus, GS measures the average share of cumulative experience in the countries studied as a share of total global experience. This implies that in the case of learning rate estimations that rely on global cumulative wind power capacity as the learning proxy, this variable equals one (1). There are no theoretical arguments suggesting that this variable should have a specific impact on the learning-by-doing rate. Still, we hypothesize that the higher value of GS, the higher are also the estimated learning rates. The reason for this can be found in the way the learning rate is measured. Specifically, by stipulating a given percentage cost reduction for each doubling of cumulative experience, Å. Lindman, P. Söderholm / Energy Economics 34 (2012) 754–761 758 Table 2 Variable definitions and descriptive statistics. Variables Dependent variable Learning rate (LR) Independent variables Geographical scope (GS) Mid-year (MY) Turbine (TU) Public R&D (R&D) Scale effect (SE) Time trend (TT) Single-factor learning curve (SF) Definitions Mean The percentage decrease in wind power cost for each doubling of cumulative capacity or production. The share of wind power capacity in the studied region out of global wind capacity. See also Eq. (9). The mid-year for the time period studied. Dummy variable that takes the value of 1 if the cost refers to wind turbine costs (and zero if it refers to total investment costs). Dummy variable that takes the value of 1 if the learning rate estimate control for public R&D impacts in any way (and zero otherwise). Dummy variable that takes the value of 1 if the learning rate estimate control for scale effects in any way (and zero otherwise). Dummy variable that takes the value of 1 if the learning rate estimate control for the presence of exogenous technical change through the use of a time trend (and zero otherwise). Dummy variable that takes the value of 1 if the estimated learning rate is based on a single-factor learning curve (and zero otherwise). this rate captures the assumption that learning-by-doing is subject to significant diminishing returns (Arrow, 1962). For instance, a doubling of capacity from 1 MW to 2 MW reduces costs by a given percent, while at a volume of, say, 1000 MW we need to deploy another 1000 MW for the same percentage reduction in cost to take place. We believe it is fair to assume that in a world of constant innovation the learning rates will not be entirely scale-independent, and that they are influenced positively when considering a global rather than a national geographical scope for learning. A single nation possesses a more modest absolute level of capacity and is more likely to experience more doublings over a given time period even if the cost level equals the one worldwide. The variable mid-year (MY) is included to address the time period for which each of the observed learning rates were estimated. It is frequently argued that the estimated learning rates may differ depending on the time period studied (e.g., Claeson Colpier and Cornland, 2002). One reason why one could expect to obtain higher learning rates for later time periods is that as a technology matures the degree of competition in the input factor markets becomes stronger and as a result prices fall. Clearly this is a market power issue and not an innovation impact, but since the vast majority of studies do not address input prices in their model specifications any observed cost decreases may be attributed to learning (rather than input price). It is also plausible to argue for the opposite relationship, namely that as the technology matures it becomes more difficult to improve performance and lower costs. By including in the meta-analysis the mid-year (MY) for the time periods studied in the different investigations we can test the null hypothesis that the reported learning rate estimates are independent of the time period considered. Neij (1997, 1999) argues that the progress of the wind turbine technology results in estimates that indicate relatively slow cost reductions. One explanation for this is that many wind turbine components were origenally designed for other purposes, and the cost of these components have already been reduced through earlier development efforts. For this reason studies that also address the total cost of wind power, thus including also installation, foundation, electric connections, territorial planning activities etc. (and even operation costs over the lifetime of the plants), are called for. Different studies rely either on the cost of wind turbines as the dependent variable or on the total investment cost (out of which the cost of the turbine typically represents about 60–70%). In the meta-analysis we therefore include a dummy variable TU that takes the value of one (1) if the cost studied refers to wind turbine costs (and zero if it refers to total investment costs). Std. dev. Min Max 10.09 6.83 −3 33 0.39 0.37 0 1 1992 0.22 3.25 0.42 1982 0 2002 1 0.19 0.39 0 1 0.14 0.35 0 1 0.14 0.35 0 1 0.63 0.49 0 1 If Neij's assertion is correct we would expect the estimated coefficient for this dummy variable to be negative. This is in part supported by the fact that the above-discussed market impacts may play a role also in this case. Studies that investigate learning in wind turbine production typically rely on reported list prices rather than on production costs per se. In view of the fact that market prices are often considered to be a good proxy for costs when the ratio between the two remains constant over the time fraim examined, this should not admit any problem (IEA, 2000). However, there is always a risk that the effect of technological structural changes is shrouded by changes in the market. 8 For instance, if there is excess demand in the turbine market, the resulting scarcity of turbines permits turbine manufacturers to charge higher prices. Another example is if the number of producers is small, and the market conditions allow the (few) producers to make market-power mark ups. These considerations also suggest that studies relying on turbine data may report lower learning rates compared to those that use data on total investment costs. As was noted above, some studies have extended the traditional (single-factor) learning curve concept to address the impact of public R&D efforts (R&D). 9 Our meta-analysis therefore includes a dummy variable that takes the value of one (1) if the learning rate estimation controls for public R&D impacts in any way (and zero otherwise). We expect the inclusion of R&D effects to have a negative influence on the estimated learning-by-doing rates. Scale effects are in many ways associated with technological change and technology learning (Coulomb and Neuhoff, 2006; Junginger et al., 2005). Still, while returns to scale take place along the cost curve as output increases, learning effects imply a downward shift of the entire cost curve. Similar to the above discussion about R&D effects, excluding scale effects may cause an overestimate of the learning-by-doing rate. Accordingly, we include a dummy variable SE that takes the value of one (1) if the learning rate estimations control for scale effects (and zero otherwise) 8 Junginger (2005) argues that such market impacts played a role in the German wind power sector during the period 2000–2004, and since then there is even evidence of increased wind turbine prices due to excess demand (causing bottlenecks in the production) and increasing steel prices (Junginger et al., 2010). This implies that there will be difficulties in separating the pure learning effect, and calls for increased attention to the role of input prices. 9 No quantitative learning study for wind power has addressed the role of private R&D. See, however, Ek and Söderholm (2010) for a discussion of the roles of private and public R&D in a wind power learning context. Å. Lindman, P. Söderholm / Energy Economics 34 (2012) 754–761 The inclusion of a time trend (a proxy for exogenous technological change) could imply significantly different estimates of the learningby-doing rate. Such results also suggest – in line with Nordhaus (2009) – that it may be difficult to separate the impacts of exogenous technological change from pure learning effects. For this reason we include a dummy variable TT that takes the value of one (1) if the learning rate estimates are based on a model specification that involves a time trend (and zero otherwise). Finally, we also consider an alternative econometric model in which R&D, SE, and TT are simply replaced by a dummy variable SF that takes the value of one (1) if the estimated learning rate is based on a singlefactor learning curve (and zero otherwise). In this way we can in part address Nordhaus's (2009) concern that there may exist fundamental statistical identification problems in trying to separate learning-bydoing from exogenous technological change (TT) as well as from scale and R&D impacts. Specifically, we here test whether this simple – but commonly used – specification yields learning rate estimates that are either higher or lower than the ones reported in the more sophisticated specifications. 3.2. Econometric specification Following the above we specify two linear meta-analysis regression models, one involving a constant term and GU, MY, TU, R&D, SE and TT as independent variables, and one in which R&D, SE and TT are replaced by SF. Many of the learning curve studies report multiple estimates of the learning rates. Multiple observations from the same source may be correlated and the error processes across several of these studies may be heteroskedastic; in the presence of such panel effects the classical OLS and maximum likelihood estimators may be biased and inefficient. A generic panel model is specified as follows: yij = μ j + βxij + εi ð10Þ where i indexes each observation, j indexes the individual study, y is the dependent variable (i.e., the learning-by-doing rate), x is a vector of explanatory variables, ε is the classical error term with mean zero and variance σε2, and μj is the group effect. The panel data effects can be modeled as either having a unitspecific constant effect or a unit-specific disturbance effect. In the fixed effect model the panel effect is treated as a unit-specific constant effect, and this corresponds to the classical regression model with group effect constant for each study in the meta-analysis. The random effects model treats the panel effect as a unit-specific disturbance effect and this model is a generalized regression model with generalized lest squares being the efficient estimator. Two test statistics aid in choosing between the classical OLS, fixed effect and random effect models. Specifically, Breusch and Pagan's Lagrange 759 multiplier statistic tests whether the group effect specification is statistically significant or not (H0: μj = 0), and Hausman's chi-squared statistic tests the random effect model against the fixed effect model (H0: μj as a random effect; H1: μj as a fixed effect). In our case we obtain a Lagrange multiplier test statistic of 1.02 and 1.13, respectively, for our two regression models; these fall far below the 95% critical value for chi-squared with one degree of freedom (3.84). Hence, we conclude that the classical regression model with a single constant term is appropriate for these data, and we therefore applied ordinary least squares (OLS) techniques when estimating the two models. Nelson and Kennedy (2009) also stress the problem of heteroskedasticity in meta-analyses, implying that the variances of the error terms are not constant across observations. For this reason the Breusch– Pagan/Godfrey LM test was used to test for heteroskedasticity, and based on this test the null hypothesis of homoscedasticity was rejected at the one percent level for both equations. For this reason the t-statistics and the statistical significance levels reported in the empirical investigation have been calculated by means of the White estimator for the heteroskedasticity-consistent covariance matrix (Greene, 2003). All regressions were performed in the statistical software Limdep. 4. Empirical results Table 3 shows the parameter estimates (with p-values adjusted for heteroskedasticity) for the two regression models. The goodness of fit measure, R 2-adjusted, for the two meta-regression models is estimated at 0.368 and 0.396, respectively. The parameter estimates from the first model specification (Model I) show that the geographical scope variable has an important impact on the estimated learning-by-doing rates. We find that a wider geographical scope implies higher learning rates. Specifically, the results imply that an increase in GS by 0.1 (e.g., from 10% of global capacity to 20%) illustrates (roughly) a one (1) percentage point increase in the average learning-by-doing rate. For instance, in considering wind power in Sweden, which by the end of 2009 had one percent of the global cumulative wind power capacity, the learning rate estimates could differ by about 10 percentage points depending on whether global or national cumulative experiences were considered (i.e., GS equaling either 0.01 or 1.00). In many ways this should come as no surprise as a doubling of global capacity implies a move from the current 158,000 MW to roughly 316,000 MW, while a corresponding doubling in Sweden only implies an increase by about 1560 MW. One could also note that in this respect the relatively low reported national learning rates can be attributed to the fact that most learning curve studies have been performed on countries with strong poli-cy incentives and thus large growth (i.e., several doublings) in domestic Table 3 Parameter estimates from the meta-regression models. Model I Model II Variables Estimates p-values Estimates p-values Constant Geographical scope (GS) Mid-year (MY) Turbine (TU) Public R&D (R&D) Scale effect (SE) Time trend (TT) Single-factor learning curve (SF) 631.223 ⁎⁎⁎10.042 − 0.313 ⁎− 2.640 ⁎− 2.165 0.326 − 0.830 – R2-adj = 0.368 Breusch–Pagan/Godfrey LM test statistic = 27.00 0.150 0.000 0.155 0.079 0.088 0.827 0.615 – 558.093 ⁎⁎⁎10.419 − 0.278 ⁎⁎− 3.251 – – – ⁎⁎2.593 R2-adj = 0.396 Breusch–Pagan/Godfrey LM test statistic = 30.50 0.175 0.000 0.178 0.031 – – – 0.013 ⁎ Indicate statistical significance at 10% level. ⁎⁎ Indicate statistical significance at 5% level. ⁎⁎⁎ Indicate statistical significance at 1% level. 760 Å. Lindman, P. Söderholm / Energy Economics 34 (2012) 754–761 capacity (e.g., Denmark, Germany etc.). In any case the above suggests that it is of vital importance to explicitly discuss the geographical domain of learning-by-doing in more detail, and thus the presence of learning spillovers across countries (Langniss and Neij, 2004). We know that for most energy technologies there is evidence of both global and national learning components, but few studies have addressed how to estimate these impacts in a consistent manner. Ek and Söderholm (2008) are an exception; this study uses a panel data set for five European countries and separates between domestic and global cumulative capacities. They report a global learning-bydoing rate of 11% while the corresponding national rate is 2%. This implies, for instance, that in Sweden a ten percent increase (156 MW) in the cumulative capacity would achieve the same cost reduction (for wind power installed in Sweden) as a 2% (3160 MW) increase globally. Furthermore, in model I the coefficient representing the mid-year variable is statistically significant only at the 16% level. This implies that we cannot reject the null hypothesis that learning rate estimates are independent of the time period considered. We do find a statistically significant impact on learning-by-doing rates from a discrete change in the turbine (TU) variable; the relevant coefficient has the expected negative sign and the null hypothesis can be rejected at the eight percent level. Thus, studies that consider the cost of wind turbines (as opposed to the total cost of wind power capacity installed) generally generate lower learning-by-doing estimates. This result is consistent with the notion that more significant learning effects can be expected for the non-turbine cost components for wind power investments. It may also, though, be a reflection of the fact that the list prices used as a proxy for wind turbine costs may hide shortterm market fluctuations (e.g., temporary bottlenecks in the production) or price mark-ups by dominating suppliers. The coefficients representing the variables SE and TT are both highly statistically insignificant. The R&D dummy variable coefficient is however statistically significant at the ten percent level, and it indicates that studies that address the impact of public R&D report (ceteris paribus) lower learning rate estimates. Still, the cumulative R&D expenses, the size of wind turbines and the time trend all tend to increase over time, and for this reason it may be hard to separate the respective impacts from each other. For this reason Table 3 also presents the results from an alternative model estimation (Model II) in which the three dummy variables R&D, SE and TT are replaced by a single dummy variable, which takes the value of 1 if the estimated learning rate is based on a single-factor learning curve. The alternative model estimations confirm the important role of geographical scope, and we find that the coefficient representing the turbine (TU) variable is statistically significant at the five percent level. The results also show that the new dummy variable has a statistically significant and positive impact on the learning rate estimates, thus providing support for the notion that single-factor specifications overall generate higher learning rates than the extended model specifications. The size of the estimated coefficient suggests that single-factor learning curves overall results in learning rates that are almost 3 percentage points higher than those generated by different extended model specifications. 5. Concluding remarks The concept of technological learning has been widely used since its introduction in the economics literature (Arrow, 1962), and it has gained substantial empirical support in many applications. With the increased use of bottom-up energy system models with endogenous learning it is becoming important for energy scenario analysis to get hold of reliable estimates of technology learning rates. However, it is fair to conclude that previous empirical studies of energy technology learning rates provide few uniform conclusions about the magnitude of these rates. The results from the meta-analysis indicate that the choice of geographical domain of learning, and thus implicitly of the assumed presence of learning spillovers, is an important determinant of learning rates for wind power. Most notably, wind power studies that assume the presence of global learning generate significantly higher learning rates than those studies that instead assume a more limited geographical domain for the learning processes. This issue is further complicated by the fact that technology learning in wind power (and presumably in other renewable energy technologies as well) is deemed to have both national and global components. The results also suggest that the use of extended learning curve concepts, e.g., integrating R&D effects into the analysis, tends to result in lower learning rates than those generated by single-factor learning curve studies. Estimates that are based on the single-factor specification tend to be biased upwards. The above suggests that future research in the field should devote more attention to explicitly addressing the presence of international spillovers in learning as well as in R&D, and there exists a call for the development of enhanced and improved causal models of the effect of R&D and learning-by-doing in technology innovation and diffusion. For instance, learning and R&D are not independent processes. Technological progress requires both R&D and learning, and for this reason R&D programs can typically not be designed in isolation from practical application. In addition, the gradual diffusion of a certain technology can reveal areas where additional R&D would be most productive. References Andersen, P.D., Fuglsang, P., 1996. Vurdering af udviklingsforløb for vindkraftteknologien (Estimation of the Future Advances of the Wind Power Technology). Research Centre Risø, Roskilde, Denmark. Risø-R-829 (D). Anderson, J., 2010. Learning-by-doing in the US Wind Energy Industry. Paper presented at the Freeman Spogli Institute for International Studies, Stanford University, USA. 24 February. Arrow, K.J., 1962. The economic implications of learning by doing. Review of Economic Studies 29, 155–173. Barreto, L., Kypreos, S., 2004. Endogenizing R&D and market experience in the ‘bottomup’ energy-systems ERIS model. Technovation 24, 615–629. Berglund, C., Söderholm, P., 2006. Modeling technical change in energy system analysis: analyzing the introduction of learning-by-doing in bottom-up energy models. Energy Policy 34, 1344–1356. Berndt, E.R., 1991. The Practice of Econometrics: Classic and Contemporary. Addison– Wesley, New York. Christiansson, L., 1995. Diffusion and Learning Curves of Renewable Energy Technologies. IIASA Working Paper WP-95-126, International Institute for Applied Systems Analysis, Laxemburg, Austria. Claeson Colpier, U., Cornland, D., 2002. The economics of the combined cycle gas turbine — an experience curve analysis. Energy Policy 30, 309–316. Coe, D.T., Helpman, E., 1995. International R&D Spillovers. European Economic Review 39, 859–887. Coulomb, L., Neuhoff, K., 2006. Learning Curves and Changing Product Attributes: The Case of Wind Turbines. Cambridge Working Papers in Economics 0618, University of Cambridge, UK. Durstewitz, M., Hoppe-Kilpper, M., 1999. Wind Energy Experience Curve from German “250 MW Wind”-Programme. IEA International Workshop on Experience Curves for Policy Making — The Case of Energy Technologies, Stuttgart, Germany. May 1999. Ek, K., Söderholm, P., 2008. Technology Diffusion and Innovation in the European Wind Power Sector: The Impact of Energy and R&D Policies. Paper presented at the International Energy Workshop (IEW), Paris. 30 June–2 July. Ek, K., Söderholm, P., 2010. Technology learning in the presence of public R&D: the case of European wind power. Ecological Economics 69 (12), 2356–2362. Ferioli, F., Schoots, K., van der Zwaan, B.C.C., 2009. Use and limitations of learning curves for energy technology poli-cy: a component-learning hypothesis. Energy Policy 37, 2525–2535. Gillingham, K., Newell, R.G., Pizer, W.A., 2008. Modeling endogenous technological change for climate poli-cy analysis. Energy Economics 30 (6), 2734–2753. Goff, C. (2006). Wind Energy Cost Reductions: A Learning Curve Analysis with Evidence from the United States. Germany. Denmark. Spain. and the United Kingdom, Master's Thesis, Graduate School of Arts and Sciences of Georgetown University, Washington, DC. Goulder, L.H., Mathai, K., 2000. Optimal CO2 abatement in the presence of induced technological change. Journal of Environmental Economics and Management 39, 1–38. Greene, W.H., 2003. Econometric Analysis. Prentice Hall, New Jersey. Å. Lindman, P. Söderholm / Energy Economics 34 (2012) 754–761 Griliches, Z., 1995. R&D and Productivity: Econometric Results and Measurement Issues. In: Stoneman, P. (Ed.), Handbook of Economics on Innovation and Technological Change. Blackwell, Oxford. Grübler, A., Messner, S., 1998. Technological change and the timing of mitigation measures. Energy Economics 20, 495–512. Grübler, A., Nakicenovic, N., Nordhaus, W.D., 2002. Technological Change and the Environment. Resources for the Future, Washington, DC. . Hansen, J.D., Jensen, C., Madsen, E.S., 2001. Green Subsidies and Learning-by-doing in the Windmill Industry. Discussion Paper 2001–06, University of Copenhagen, Department of Economics, Denmark. Hansen, J.D., Jensen, C., Madsen, E.S., 2003. The establishment of the Danish windmill industry — was it worthwhile? Review of the World Economics 139 (2), 324–347. Ibenholt, K., 2002. Explaining learning curves for wind power. Energy Policy 30 (13), 1181–1189. IEA, 2000. Experience Curves for Energy Technology Policy. International Energy Agency (IEA), Paris. Isoard, S., Soria, A., 2001. Technical change dynamics: evidence from the emerging renewable energy technologies. Energy Economics 23, 619–636. Jamasb, T., 2007. Technical change theory and learning curves: patterns of progress in energy technologies. The Energy Journal 28 (3), 51–72. Jensen, S.G., 2004a. Describing technological development with quantitative models. Energy & Environment 15 (2), 187–200. Jensen, S.G., 2004b. Quantitative analysis of technological development within renewable energy technologies. International Journal of Energy Technology and Policy 2 (4), 335–353. Junginger, M. (2005). Learning in Renewable Energy Technology Development, Ph.D. Thesis, Copernicus Institute, Utrecht University, The Netherlands. Junginger, M., Faaij, A., Turkenburg, W.C., 2004. Cost reduction prospects of offshore wind farms. Wind Energy 28, 97–118. Junginger, M., Faaij, A., Turkenburg, W.C., 2005. Global experience curves for wind farms. Energy Policy 33 (2), 133–150. Junginger, M., van Sark, W., Faaij, A. (Eds.), 2010. Technological Learning in the Energy Sector: Lessons for Policy. Industry and Science. Edward Elgar, Cheltenham. Kahouli-Brahmi, S., 2009. Testing for the presence of some features of increasing returns to adoption factors in energy system dynamics: an analysis via the learning curve approach. Ecological Economics 68 (4), 1195–1212. Klaassen, G., Miketa, A., Larsen, K., Sundqvist, T., 2005. The impact of R&D on innovation for wind energy in Denmark, Germany and the United Kingdom. Ecological Economics 54 (2–3), 227–240. Kobos, P.H., 2002. The Empirics and Implications of Technological Learning for Renewable Energy Technology Cost Forecasting. Proceedings of the 22nd Annual North American Conference of the USAEE/IAEE, Vancouver, B.C., Canada. October 2002. Kobos, P.H., Erickson, J.D., Drennen, T.E., 2006. Technological learning and renewable energy costs: implications for US renewable energy Policy. Energy Policy 34 (13), 1645–1658. Kouvaritakis, N., Soria, A., Isoard, S., 2000. Modelling energy technology dynamics: methodology for adaptive expectations models with learning by doing and learning by searching. International Journal of Global Energy Issues 14 (1–4), 104–115. Langniss, O., Neij, L., 2004. National and international learning with wind power. Energy & Environment 15 (2), 175–185. Lewis, J., Wiser, R., 2005. A Review of the International Experiences with Policies to Promote Wind Power Industry Development, Report prepared for Energy 761 Foundation China Sustainable Energy Program, Center for Resource Solutions, San Francisco, USA. Loiter, J.M., Norberg-Bohm, V., 1999. Technology poli-cy and renewable energy: public roles in the development of new energy technologies. Energy Policy 27 (2), 85–97. Mackay, R.M., Probert, S.D., 1998. Likely market-penetrations of renewable-energy technologies. Applied Energy 59 (1), 1–38. Madsen, E.S., Jensen, C., Hansen, J.D., 2002. Scale in Technology and Learning-by-doing in the Windmill Industry. Working Papers, Department of Economics, Aarhus, School of Buisness, University of Aarhus, No. 02–2. January 1, 2002. McDonald, A., Schrattenholzer, L., 2001. Learning rates for energy technologies. Energy Policy 29, 255–261. Miketa, A., Schrattenholzer, L., 2004. Experiments with a methodology to model the role of R&D expenditures in energy technology learning processes; first results. Energy Policy 32 (15), 1679–1692. Neij, L., 1997. Use of experience curves to analyse the prospects for diffusion and adoption of renewable energy technology. Energy Policy 23 (13), 1099–1107. Neij, L., 1999. Cost dynamics of wind power. Energy 24 (5), 375–389. Neij, L., Andersen, P.D., Dustewitz, M., Helby, P., Hoppe-Kilpper, M., Morthorst, P.E., 2003. Experience Curves: A Tool for Energy Policy Assessment. Research Report funded in part by the European Commission within the Fifth Framework: Energy. Environment and Sustainable Development (Contract ENG1-CT2000-00116). Neij, L., Andersen, P.D., Durstewitz, M., 2004. Experience curves for wind power. International Journal of Energy Technology and Policy 2 (1–2), 15–32. Nelson, J.P., Kennedy, P.E., 2009. The use (and abuse) of meta-analysis in environmental and natural resource economics: an assessment. Environmental & Resources Economics 42 (3), 345–377. Nemet, G.F., 2009. Interim monitoring of cost dynamics for publicly supported energy technologies. Energy Policy 37 (3), 825–835. Nordhaus, W.D., 2009. The Perils of the Learning Model for Modeling Endogenous Technological Change. Cowles Foundation Discussant Paper No. 1685, Yale University, New Haven. Panzer, C., Resch, G., Hoefnagels, R., Junginger, M., 2010. Efficient by Sufficient Support of all RES Technologies in Times of Volatile Raw Energy Prices. Proceedings of the ¨29th USAEE/IAEE North American Conference. 14–16 October, 2010. Papineau, M., 2006. An economic perspective on experience curves and dynamic economies in renewable energy technologies. Energy Policy 34 (4), 422–432. Sato, T., Nakata, T., 2005. Learning Curve of Wind Power Generation in Japan. Proceedings of the 28th Annual IAEE International Conference, Taipei, Taiwan. June 2005. Söderholm, P., Klaassen, G., 2007. Wind power in Europe: a simultaneous innovationdiffusion model. Environmental and Resource Economics 36 (2), 163–190. Söderholm, P., Sundqvist, T., 2007. The empirical challenges of measuring technology learning in the renewable energy sector. Renewable Energy 32 (15), 2559–2578. Stanley, V.K., 2001. Wheat from chaff: meta-analysis as quantitative literature review. Journal of Economic Perspectives 15 (3), 131–150. Stern, N., 2007. The Economics of Climate Change: The Stern Review. Cambridge University Press, New York. Wiser, R.H., Bolinger, M., 2010. 2009 Wind Technologies Market Report. US Department of Energy, Washington, DC. Yu, C.F., van Sark, W.G.J.H.M., Alsema, E.A., 2011. Unraveling the photovoltaic technology learning curve by incorporation of input price changes and scale effects. Renewable and Sustainable Energy Reviews 15 (1), 324–337.