Energy efficient R&D investment and Aggregate Energy Demand: Evidence from OECD Countries * Karimu, Amin a Brännlund, Runar a

Size: px
Start display at page:

Download "Energy efficient R&D investment and Aggregate Energy Demand: Evidence from OECD Countries * Karimu, Amin a Brännlund, Runar a"

Transcription

1 Energy effcent R&D nvestment and Aggregate Energy Demand: Evdence from OECD Countres * Karmu, Amn a Brännlund, Runar a a Centre for Envronmental and Resource Economcs Umeå School of Busness and Economcs Umeå Unversy, Sweden July 2015 Abstract Ths paper present a dfferent perspectve n the debate on energy effcency and energy demand by classfyng the mpact of effcency measures nto drect and ndrect effect (rebound effect). It examnes the potental drect effect of energy effcent R&D capal on energy demand. Irrespectve of the rebound effect assocated wh energy effcency, s possble to have a negatve drect effect from effcency measures. Usng a sample of OECD countres, we fnd evdence n support of negatve own-r&d capal elastcy wh respect to energy demand. Further, we fnd evdence of heterogeney n the predcted mpact of energy effcent R&D capal, wh the USA havng the largest accumulated reducton n energy demand and Portugal wh the smallest. Overall, our emprcal results suggest a reasonable reducton n energy demand and assocated CO 2 emssons from an ncrease n energy effcent R&D nvestment base on the drect effect and ths vares across the countres. Keywords: Effcency Polcy, Energy Demand, Heterogeney, R&D capal, Spllovers JEL Classfcaton: C33; Q41; Q49. * Fnancal support for ths research has been provded by the Swedsh Energy Agency

2 1. Introducton After the commencement of the Kyoto protocol n 2005, there has been ncreasng attenton by polcy makers, academa and frms across varous countres, especally ndustralzed countres on a share burden n meetng the targets set by the Kyoto protocol and ther respectve natonal and regonal block s targets. Despe the global awareness on the negatve consequences of CO 2 emssons, global CO 2 emssons stll ncreased to 35.3 Ggatonnes n 2013, a 0.7 Ggatonne hgher than the 2012 fgure (about 2% ncrease n 2013 relatve to 2012). The three top ranked emters n 2013 accounts for 55% of global CO 2 emsson, and are Chna (29%), Uned States (15%) and the European Unon (EU28) (11%). In 2011, OECD countres contrbuted 38% of global CO 2 emssons and energy combustons accounts for close to 99% of the emssons (IEA, 2013). On ndvdual country level, Chna s the leadng CO 2 emtng country, followed closely by USA, whle Inda occupes the thrd sport n absolute values (IEA, 2013). Arguably, one potental channel to reducng the negatve mpact of energy on the envronment s through energy effcency. However, ths s true f the benefs from effcency measures are taxed and nvest n R&D that can produce new technologes that wll enable a shft away from fossl based energy. Ths type of polcy mx wll reduce or elmnate the potental rebound effect assocated wh effcency measures. A potental channel to ncrease energy effcency s through research and development (R&D). In most, f not all energy polcy packages for OECD countres, there s a great element or requrement for energy effcency and hence ncreasng expendure n energy effcent related R&D s. Energy effcent R&D s therefore seen as a crucal channel among polcy makers, at least n decouplng energy use from GDP growth and consequently reducng global CO 2 emssons n general, f combne wh a well desgn tax scheme. For nstance the OECD countres that are members of the European Unon (EU), under the 20/20/20 drectve are requred to contrbute ther respectve natonal targets to reducng energy use by 20% va energy effcency measures by the year 2020.Such a polcy calls for energy effcent R&D nvestment and other measures that could nfluence energy effcency among member countres. There s however very ltle emprcal studes that examne the potental contrbuton from energy effcency measures on aggregate energy demand n general, partcularly energy 1

3 effcent R&D nvestment on aggregate energy demand, at least for OECD countres. Most of the studes on R&D are rather based on s mpact on output va Grlches (1979) type of analyss that focus on a knowledge producton whn a more general producton theory framework, Hall et al. (2009) provdes a recent revew of the lerature n ths area. The few studes based on energy for nstance Metcalf and Hassett (1999), whch measures energy savngs from home mprovement nvestment, Geller et al. (2006) revews energy ntensy trends for major OECD countres from 1973 to 2002, assessng how much of the declnng trend s due to energy effcency and what share s attrbutable to structural change, Gllngham et al. (2006) provdes estmates for cumulatve savngs from energy effcency polces for the USA, Aroonruengsawat et al. (2012) examned the mpact of buldng codes on electrcy consumpton at the household level for the USA 1. Others tend to focus on assessng programs that dentfy and promote energy-effcent products as n Webber et al. (2000), whle Meyers et al. (2003) focus on energy effcency standards for applances on energy and envronmental mpacts for the USA among others. Gven the few emprcal evdence on the mpact of energy effcent R&D nvestment on energy demand couple wh the ncreasng nterest among polcy makers on the possble contrbuton of energy effcency n at least decouplng energy use from economc growth, there s the need to provde more emprcal evdence on the potental mpacts, especally from energy effcent R&D nvestment for nform polcy formulaton. The objectve of ths paper s to address the followng key questons; what s the own - energy effcent R&D capal elastcy f the usual mposon of the addve separably assumpton n the emprcal lerature to enable the dentfcaton of own/prvate R&D elastcy from spllover 2 effect s wrong? What s the potental contrbuton of energy effcent R&D nvestment on aggregate energy demand reducton f spllover effects are excluded? Is there a dmnshng return to energy effcent R&D nvestment? Whch countres n the sample are lkely to benef more from a polcy that ncrease energy effcent R&D nvestment? Answers to these questons are very mportant to clearly understand the dynamcs between energy demand and R&D nvestment n energy 1 Others studes on the mpact of effcency measures such as buldng codes on electrcy consumpton ncludes; Bon and Wllam (1991), Costa and Kahn(2011), Jocobsen and Kotchen(2013). 2 A transfer of knowledge from one actor to another n whch the recevng actor does not pay for the full cost of accessng and the use of such knowledge. It therefore nvolves unntentonal knowledge transfer from one actor to another. 2

4 effcency n order to nform polcy desgns focusng on such ssues. In answerng these questons we focus on the drect effect of energy effcent R&D capal. We classfy the effect of energy effcent R&D capal on energy demand nto two-the drect effect and the ndrect effect (rebound effect). The dea for the classfcaton s to enable us focus on only the drect effect. Our paper makes the followng contrbuton to the lerature. Frst provde the frst emprcal evdence to the best our knowledge on R&D capal elastcy wh respect to aggregate energy demand for a sample of OECD countres whout mposng the a pror addve separably assumpton between own - energy effcent R&D capal elastcy and spllover effect as well as other unobserved common factors that could conflate the elastcy estmate for the energy effcent R&D capal (henceforth called own R&D capal). It also provde the polcy effect of an ncrease n energy effcent R&D nvestment on energy demand across the countres, whch wll hghlght the countres that are lkely to acheve the greatest reducton n energy demand from such a polcy whch s lkely to be very nformatve for polcy makers. The key fndng from ths study nclude the followng; that when no efforts are made to correctly account for spllover effects and other unobserved common factors, the own R&D capal elastcy and s assocated prvate return s bas upwards, mplyng that the lnear separably assumpton usually mposed for the dentfcaton of R&D capal mght be wrong. Further, the results provded evdence of a negatve mpact of R&D capal on energy demand, wh a hgher cumulated mpact from the USA n the sampled countres, whle Portugal has the least mpact. However based on percentage contrbuton of an ncrease n energy effcent R&D nvestment on percentage reducton n energy demand, Portugal s top ranked and USA the least ranked among the countres n the sample. The rest of the paper s organzed as follows. Secton 2 presents the theoretcal background of the study, the econometrc model and data s presented n Secton 3. Secton 4 dscussed the results, whle secton 5 presents the concluson of the study. 2. Theoretcal background The general background on economc analyss on energy effcency ssues usually stem from cost-mnmzaton/ utly or prof maxmzaton behavor of households and frms. 3

5 The general dea on energy effcency s what represents, snce energy s not an end n self but rather the servces provdes for nstance heatng, lghtng, moton among others. It s therefore mportant to conceptualzed energy as an nput nto the producton of energy servces and as a consequence, energy effcency can generally be defne as energy servces per un of energy nput. Based on ths we can use the producton theory framework to derve and understand energy effcency. In the producton theory perspectve, capal and energy are the nputs nto the producton of energy servces (Gllngham et al., 2009). In ths framework, energy effcency s located at the pont of tangency between an soquant (that specfy a gven level of energy servce) and rato of prces between capa and energy. However nvestment n new capal n the quest to mprove energy servce per un of energy nput nvolves comparng the current and future returns from the nvestment to the cost. Ths makes relatve prces (capal and energy) to depend on the cost of capal ncurred n energy servce mprovement, the dscount rate that lnks the current and future cost and returns on energy effcent nvestment, future energy prces, and equpment utlzaton. Achevng sgnfcant mprovement n capal and energy usng equpment performance n terms of energy use depend almost exclusvely on technologcal development, whch s greatly nfluence by R&D actves. There s large lerature on knowledge development and s mpact on output, nfluenced by Grlches (1979) semnar paper that ncorporated a knowledge producton functon nto a standard Cobb-Douglas producton functon. In ths framework, output depends on capal and labor nputs n addon to knowledge capal, precsely R&D capal. The unque feature of knowledge capal (non-excludably and non-rvalry) as noted by Arrow (1962), makes dffcult to estmate precsely own - knowledge capal elastcy. Ths s because of the nherent spllover effects assocated wh knowledge capal and the dffculty assocated n quantfyng spllovers. In the appled lerature, researchers generally mpose lnear separably assumpton on Grlches s type of producton functon to enable estmaton of own knowledge capal elastcy (as well as prvate return to knowledge capal) or a set-up n whch an attempt s made to quantfy spllover effect va ad-hoc weghts. The weghts are constructed based on the relatve nteracton of actors (ndustres, countres etc.). Gven the dffcult n separatng spllover effects from own -knowledge effect, makes more mportant to try to estmate own-r&d capal by methods that are able to elmnate the potental spllover effect n the estmaton process. 4

6 In the energy lerature, most of the studes on energy effcency tend to focus more on eher behavoral neffcences n energy demand or market falures and the approprate polcy desgn to correct these neffcences and market falures. Ths has led to many studes lookng at the so call rebound effect wh less studes lookng at the drect effect of knowledge capal on energy demand, especally the energy savngs potentals of nvestment n knowledge accumulaton that result n energy effcent capal and drect effect of that on energy savngs or potental energy savngs. 3. Emprcal approach and the Data Formally we specfy the basc emprcal model as: e p y hhd R Dcap u & 1,..., N Countres, t 1,..., T Tme perod (1) Where the varables are n natural logs and e s aggregate energy consumpton per capa, p s the energy prce, y s real ncome per capa, hhd s heatng degree days, R & Dcap s energy effcent R&D capal and u s a compose term that ncludes a random error, country fxed effects and possbly unobserved common factors. The above model s a varant to the model used n Ryan and Ploure (2009), Flppn and Hunt (2011), Karmu and Brännlund (2013), n the sense that also ncluded energy effcency polcy varable n the form of energy effcent R&D capal smlar to the model used by Aroonruengsawat et al. (2012). The theoretcal background for the above models s from the utly framework as done n Karmu and Brännlund (2013). Move over, n order to approprately estmate the own R&D capal elastcy, we apply an econometrc modellng approach based on the unobserved common factor framework that wll enable strppng off potental unobserved factors ncludng spllovers on the estmated parameters of nterest. The approach s brefly presented below, n whch for easy exposon we restrct to a model wh one explanatory varable a smplfy verson of equaton (1) s presented as follows: e x u (2) u f (3) 1 t x f g (4) s 2 t t 5

7 Where x s the explanatory varable, u s a compose term that s composed of country fxed effects ( 1 ), unobserved common factors ( f t ) and a random error term ( ). The explanatory varable ( x ) s also drven by unobserved common factors ( gt and f s t, a subset of f t that also affect x ) and a stochastc error term ( ). Whereas, and are the factor loadng parameters that vary across panel uns. The overlap of common factors n Eq.(3) and (4) creates endogeney problems that render the dentfcaton of very dffcult n the usual panel estmators such as fxed effects, dynamc GMM estmators and varant of them that are not desgn whn the context of unobserved common factor framework as descrbed above. The estmators that are desgn to accommodate ths type of endogeney are the common correlated mean group (CCEMG) and augmented mean group (AMG) estmators proposed by Pesaran (2006) and Eberhardt and Teal (2010), respectvely. The ntuon s that, s assumed latent processes drve both the dependent varable va equaton (3) and the explanatory varable va equaton (4) wh possble dfferent strength va, and. If on average the factor loadng parameters are zero, then the usual panel data estmators such as fxed effect, dynamc panel estmators and varants of them produce consstent and unbased estmator for the parameter f the assumpton that s true. However, f on average the factor loadng parameters are not zero, then the usual panel estmators wll be based and nconsstent as shown n Eberhardt et al. (2013). We can generalze equaton (1) to a multvarate case f we assume that x s a vector of explanatory varables and s a vector of parameters correspondng to the vector of regressors. The expresson as n Eq. (2) to (4) s estmated for each panel un and the average panel coeffcent for s calculated as N, hence gven a long tme perod we can have estmates for each panel un as well as the average over all the panel uns to assess f the parameters vary across countres. In specfc reference to ths study, e denote aggregate per capa energy consumpton, the vector x comprses energy prce, real ncome, heatng degrees days and energy effcent R&D capal. 1 x N 1 The econometrc strategy s to apply a non-heterogeneous panel estmator (fxed effect model) and three varants of the heterogeneous estmators Pesaran and Smh (1995) mean group (MG), Pesaran (2006) CCMG and the Eberhardt and Teal (2010) AMG estmator, where the last two heterogeneous estmators are based on the unobserved 6

8 common factor framework descrbed above. Further, we assess the dfferent estmators on how each f the data generaton process based on dagnostc testng to choose the best model among them, and base our analyss and polcy mplcatons on the chosen estmator, detals on the three heterogeneous estmators s presented n the appendx. Each of the three heterogeneous estmators dffers n ther modellng approach. Whereas the MG-estmator s desgned n a way that allows for heterogeney n parameters across the panel uns, however does not account for possble common unobservable factors. The AMG and CCMG estmators on the other hand account for common unobservable factors n addon to allowng for heterogeney n the parameters. The above framework s therefore the approprate approach that wll help answer the key questons that the paper ntends to answer as enumerated n the ntroducton. These heterogeneous estmators relatve to the homogeneous estmators (e.g., the fxed effect estmator) are able to address the followng. Frst, they relax the constant parameter assumpton n a sense that they allow for varably n the parameters across the panel uns. Second, both the AMG and CCMG estmators also accounts for cross sectonal dependence (effects of common unobservable factors, such as energy/ol crss, global fnancal crss, spllover effects of mprovement n technology, etc.) and hence account for possble spllover effects whch allows for estmatng the own R&D capal elastcy by strppng-off the possble confoundng effects from spllovers akn to the work by Eberhardt et al. (2013). Ths s partcularly mportant for our study, snce our focus s on estmatng the own R&D capal elastcy and n the obscene of an approach that can relably estmate socal returns (spllovers), becomes handy to apply ths approach that strp-off potentally all the effects of spllover n addon to other unobservable factors from the estmates. Lastly, both the AMG and CCMG estmators, relaxes the constant effects of common shocks such as global recesson by allowng these factors to vary across country va factor loadng approach unlke the conventonal technque of usng tme dummes to try and capture unobservable that are assumed to affect all panel uns albe wh the same strength. Data The consumpton, prce and ncome data seres for ths paper were all taken from Adeyem et al. (2010), Karmu and Brännlund (2013) - an annual data set for a panel of OECD 7

9 countres coverng the perod 1960 to 2006; however we lmed the coverage for ths paper to the perod 1980 to The reason for ths lmaton s that, the data on our nterest varable (energy effcent R&D expendures) starts from 1974 but to be able to calculate R&D capal from the expendure seres, we used the frst 6 years to buld the R&D capal that takes effect from 1980 (how ths s done s explan below). The varables comprse of aggregate energy consumpton (E) for each country expressed n thousand tons of ol equvalent (ktoe), GDP (Y) n bllons of 2,000 US$ usng PPP for the entre perod and the real energy prce ndex (P) at 2,000 US dollars. Both Y and E are converted nto per capa terms by dvdng each country's Y and E by ther respectve populatons. We used heatng degree days to proxy for the effects of clmate on energy demand and the data were retreved from Eurostat - Statstcal Offce of the European Unon se for all the countres except that of the USA, whch was taken from Natonal Oceanc and Atmospherc Admnstraton ( NOAA). Data on energy effcent R&D expendures are retreved from the Internatonal Energy Agency (IEA) energy technology RD&D statstcs for OECD countres. These expendures are used to calculate the R&D capal, constructed by applyng the perpetual nventory model, n whch R&D capal s derved as: R & Dcap (1 ) R & Dcap R & D (5) 1 Where R&D denotes real R&D expendures, s the deprecaton rate whch we follow prevous lerature (Hall, 2007 and Eberhardt et al., 2013) and assume a 12% rate. The nal capal stock s calculated as: R& D 0 R & Dcap 1 g (6) where R & Dcap 1 s the nal capal, g s the country specfc growth rate, whch we used the frst 6 years of the observed R&D expendures to compute. The calculated R&D capal calculated va equaton (5) to (6) revealed varatons across the sampled countres based on the boxplot presented n fgure1. The country wh the least medan value s Portugal (16.78 express n logarhms) and the U.S has the hghest medan value (21.66). The countres wh the hghest R&D capal n the sample nclude; Sweden, UK, Netherland, Italy and the US, however these values are not n per capa terms. From the boxplot, one can dentfy few countres wh potental outlers for the R&D capal seres, 8

10 Log R&D Capal n a whole, these few outlers are due to the usual perodc ncreases n energy effcent R&D expendures n these countres and not due to errors. Further, gven the estmaton methodology, these outlers wll not sgnfcantly nfluence the parameter estmates and hence lkely not to bas the estmates. The boxplot n fgure A2 n the appendx ndcate some varably for each of the other varables n our data set across the sampled countres. Both the heatng degree day s seres and energy consumpton per capa seres have the hghest varaton relatve to real prce and ncome seres. The Countres n the study are: Austra, Belgum, Denmark, France, Greece, Ireland, Italy, Netherlands, Norway, Portugal, Span, Sweden, Swzerland, the UK and the USA. Dfferences n R&D Capal across the 13 OECD Countres Country-ID Fgure 1: Boxplot showng the varably of the medan value for R&D capal across 13 OECD countres. (Note: the Country-ID, 1=Austra, 2=Belgum, 3= Denmark, 4=France, 5=Italy,6=Netherland,7=Norway,8=Portugal, 9=Span, 10=Sweden,11=Swzerland, 12=UK, 13=USA) 4. Results Table 1 shows the results from cross-sectonal dependence test based on the raw seres. The results ndcate each of the seres n our data could not pass the null hypothess of cross-sectonal ndependence, mplyng that each of the data seres are correlated across panel uns and therefore the econometrc strategy should ncorporate ths nto the estmaton process, n order to reduce the potental problem of producng bas estmates. 9

11 Ths frst step n our estmaton strategy therefore means that both the AMG and CCMG estmators are lkely to be the best estmators for ths study as they are purposely desgn to handle data wh cross sectonal dependence/common factors. We also tested for the tme seres propertes of the data, specfcally un root test, whch show evdence of I(1) 3 as presented n Table A1 n the appendx. However to be sure the estmators based on the common factor frame work are approprate relatve to say fxed effect model n ftng the data generatng process, we also estmated a fxed effect model and used varous dagnostc tests of the resduals to assess whch of the models f the data generatng process (DGP). Specfcally we estmated four dfferent estmators fxed effect, MG, AMG and CCMG. Table 1: Cross sectonal dependence test Varable CD-test P-value Correlaton e p y R&Dcap hhd Notes: Under the null hypothess of Cross-secton ndependence CD ~ N (0, 1), R&Dcap and hhd denotes R&D capal and heatng degree days, respectvely. The results for each of the four estmators ndcated above s presented n Table 2 and reveal that the estmates from the homogeneous fxed effect model are bgger n magnude relatve to the heterogeneous estmators (except the estmates for heatng degree days) and ths s partcularly evdent for our nterest varable R&D capal. Another observaton from the results s that, the estmated coeffcent for R&D capal s approxmately the same across the three heterogeneous models. Further, each of the estmators ndcates a negatve own R&D capal elastcy, mplyng a negatve response of aggregate energy demand to energy effcent R&D capal. The estmated value s statstcally sgnfcant at the 5% for all except the CCMG model. Addonally, all the estmates from the four estmators have the expected sgns from theory. 3 Two dfferent panel un root tests are mplemented, specfcally Pesaran (2007) CIPS test that allows for heterogeney as well as cross-sectonal dependency and Maddala and Wu Panel Un Root test that does not allow for cross-sectonal dependence n the testng procedure. Detals on the CIPS un root test specfcaton are n Pesaran (2007). We also tested for contegraton and the results are reported n Table A1 n the appendx. 10

12 In dscrmnatng between the estmators, we reled on the models dagnostcs, especally f the models resduals pass the cross sectonal ndependence test (Pesaran, 2006 CD-test) and are statonary I(0). The dagnostc test results as reported n Table 2, favor both the AMG and CCMG models. Ther respectve resduals are statonary mplyng non-spurous regresson and also pass the CD-test at the 5% level. Whle the MG model s also nonspurous, however fals the CD-test. The FE model s resduals follow an I (1) process and also fals the CD-test. The dagnostc tests therefore provde strong support for the heterogeneous models relatve to the fxed effect model. We can conclude that, the estmates from the fxed effect model are bas upward due to falng to correctly model common factor effects. More mportantly, gven the nature of R&D wh nherent spllovers, s mpossble to precsely estmate prvate R&D elastcy from non-common factor models such as a fxed effect model. Both the AMG and the CCMG models have approxmately the same level of performance based on the two dagnostc tests, however we decded to based our dscusson on the estmates from the AMG model on the grounds that produces a sgnfcant ncome varable that s also whn range of values usually found n the lerature, whereas n the case of the CCMG the ncome estmate s statstcally not dffer rent from zero at the 5% level of sgnfcance, whch ntuvely does not make sense, snce ncome s a key varable that nfluence energy demand and even on the evoluton of energy use from pre-ndustral error to modern economc systems of producton and consumpton (Stern and Kander, 2012). The AMG estmates shows that prce elastcy s rather low but consstent wh fndngs from prevous studes such as Adeyem and Broadstock, 2009; Karmu and Brännlund, 2013; Welsch, Ths mples less response to prces across the sample countres whch can be due to many thngs ncludng hgh ncome levels whch make prce ncreases less panful for the average person whn these countres. It could also be the case that, overtme, the average stock of applances and equpment that reles on the use of energy servces has ncreased and couple wh the wealth effect, consumers and producers are becomng less and less responsve to prce ncreases and also because of less avalable substutes for energy whch tend to make energy more of a necessy n the modern economy. 11

13 Table 2: Regresson Results FE MG AMG CCMG p ** *** *** ** (0.098) (0.035) (0.034) (0.073) y ** *** *** (0.413) (0.095) (0.106) (0.170) R&Dcap *** ** ** (0.025) (0.020) (0.016) (0.032) hhd *** ** *** (0.023) (0.035) (0.044) (0.033) Trend yes yes yes yes Constant (12.391) (0.769) (0.698) (1.077) Dagnostcs CD-test [0.015] [0.022] [0.108] [0.067] Integraton I(1) I(0) I(0) I(0) N Note: Standard errors n parentheses (robust standard errors), * p < 0.10, ** p < 0.05, *** p < 0.01, CD-test s the cross-sectonal dependence test base on Pesaran (2004) test for the Null of cross-sectonal ndependent resduals, N s the number of observatons, hhd s the heatng degree days and R&Dcap denote R&D capal. Numbers n square bracket are the p-values for CD-test. The estmated ncome elastcy have the expected sgn and sgnfcant but low n magnude compare to some of estmates from prevous studes (Fllpn and Hunt, 2011; Atknson and Mannng, 1995) but consstent wh the estmates from Adeyem and Broadstock, 2009 and Welsch, The mplcaton s that, people on average demand more energy as ther ncome level rses. The estmated hhd elastcy s posve and sgnfcant, whch shows that outsde temperatures below whch heatng s requred tend to ncrease energy demand. Further, the estmated R&D capal elastcy s and sgnfcant at the 5% level. Next we calculate the effect of the share of R&D capal on energy demand by usng the 1 average estmated elastcy ( N N 1 4, ) multple by the R&D capal. Ths can be nterpreted as the predcted effect of R&D capal on energy demand based on the AMG model, whch wll gve us the share predcted mpact of R&D capal for each of the countres n our sample. The average mpact of R&D capal across the countres n our sample as reported n fgure A1 n the appendx range from approxmately 15% to 20% (cumulated over 1980 to 2006). The US s the country that has the largest energy savngs (predcted) from R&D nvestments, whlst Portugal has the least savngs. Ths shows that polces targetng energy effcency nvestments are havng the ntended effects (reduced 12

14 energy demand). However the effects are relatvely modest gven that the values are the averages over the perod , mplyng an annual reducton of approxmately 1%. Ths smaller contrbuton to energy savngs from energy effcent nvestment means that we need to combne energy effcency polces wh other polces such as taxes to have a reasonable reducton n energy demand and consequently on CO 2 reducton. Polcy mplcatons of R&D nvestment In assessng the effect of a gven change n energy effcent R&D nvestment on energy demand, we follow the approach mplemented n Davs and Kllan (2011) by expressng the percentage reducton n energy demand from an x mllon US$ ncrease n energy effcency R&D nvestment evaluated at a chosen base of R&D capal (taken as the volume-weghted mean of R&D capal at 2006), ths s formulated as 1 where N N 1 ˆ 4, N N 1 ˆ 4,( )*100, 1 R & Dcap 13 x s the estmated average R&D capal elastcy, R&Dcap s the base for the evaluaton of the change whch s the volume-weghted mean n year 2006.To demonstrate ths polcy effect of an ncrease n energy effcent R&D nvestment, we consder a 100 mllon US$ ncreased n R&D nvestment, however the choce of sze of the nvestment s rrelevant. The key dea s to show the lkely percentage change n energy demand from a gven ncrease n energy effcent R&D nvestment n order to assess the lkely effcacy of energy effcent polcy/polces related to energy effcent capal. Table 3 present these effects for each of the countres n the data set and show vared dfferences n energy demand response to the polcy. Portugal stands out wh the hghest (34.8%) reducton n energy demand from a 100 mllon US$ ncreased n R&D nvestment, whle the USA has the least reducton (0.08%). Ths complement the results on predcted energy savngs from energy effcent R&D nvestment n the sense that the country wh the least savngs from 1960 to 2006 tends to have the hghest reducton n energy demand from a polcy that wll ncrease energy effcent R&D nvestment relatve to a country/countres that has/have the hghest savngs. Ths s n lne wh the economc prncple of dmnshng returns (n ths case to R&D nvestment). The reducton n energy demand s mnmal for countres such as Netherland, Italy, France and the UK. Ths result ndcates sgnfcant heterogeney n the mpact of an ncrease n R&D nvestment n energy effcency polcy.

15 Table 3: The effects of 100 mllon US$ ncrease n R&D nvestment n energy effcency on energy demand. Country Austra Belgum Denmark France Italy Netherland Norway %Energy Reducton Country Portgal Span Sweden Swzerland UK USA %Energy Reducton Table 4: Carbon doxde emsson reducton from 100 mllon US$ ncrease n energy effcent R&D nvestment. Country Austra Belgum Denmark France Italy Netherland Norway %CO 2 Reducton Country Portgal Span Sweden Swzerland UK USA %CO 2 Reducton Further we extended the above polcy analyss to carbon reducton from the effect of energy effcent R&D nvestment and ths s done va usng the energy reducton from the polcy as presented n Table 3 and the converson rate ( from OECD countres n 2010) of energy to CO 2 emsson. The carbon doxde reducton from ths s reported n Table 4 and show CO 2 reducton that vary from 0.03% for the USA to 13.3% for Portugal. Robustness Analyss As a robustness check on our key results reported n Table 2, we also relax the statc structure for each of the three heterogeneous estmators by applyng a dynamc model. The result for ths analyss s reported n Table A2 n the appendx. Generally, there s no sgnfcant dfference n the sze of the long-run estmates from both the dynamc and statc model based on the AMG and MG estmators, especally on our varable of nterest (R&D capal). The CCEMG estmates tend to vary greatly between statc and dynamc verson of the model for the varable of nterest, however none of dynamc long-run estmates for R&D capal s statstcally sgnfcant at the 5% level. 14

16 6. Concluson The man objectve of ths study s to examne the mpact of energy effcent R&D nvestment on aggregate energy demand for OECD countres, the possble dfferences n the mpact across the countres n the sample and how ths translate to CO 2 emsson reducton across the countres. Our analyss mplemented both homogeneous and heterogeneous panel estmators wh dfferent estmaton assumpton. The estmators that are based on the unobserved common factor framework allow for both heterogeney n the parameters and unobserved common factors. Ths approach (unobserved common factor models) enables a proper estmaton of the own-r&d capal elastcy. The reason for focusng on the prvate/own elastcy rather than the socal or both s because the current econometrc methods do not provde accurate way/s to precsely estmate the spllover effect, the component that accounts for the dfference between prvate/own and socal elastcy, hence the choce of the unobserved common factor framework to help strp-off all the unobserved common factors ncludng spllovers n order to at least estmate the prvate elastcy more precsely. Our key result ndcate a negatve own R&D capal elastcy whch s however small n our preferred model relatve to estmates based on the fxed effect model, whch we argue that the dfference between the fxed effect estmates and AMG estmates s due to the nably of the fxed effect model to strp-off spllover effects as well as not allowng for possble heterogeney n the parameter estmates. Ths concluson s based on the dagnostc test that ndcates non-statonary n the resduals from the fxed effect model and the nably to deal wh cross-sectonal dependency. The poor dagnostcs from the fxed effect model, specfcally the nably to correct for cross-secton dependence provde evdence to support the argument that the addve separably assumpton usually mposed to enable estmatng prvate elastcy and return to R&D capal s lkely napproprate as ndcated n Eberhardt et al.(2013). Ths result s n lne wh those from Eberhardt et al. (2013) albe wh dfferent applcaton (producton n Eberhardt et al. and energy demand n our study), tme frame and countres n the sample. Further, our result ndcate that a polcy of ncreasng energy effcent R&D nvestment result n reducton n aggregate energy demand that vares sgnfcantly across the sampled 15

17 countres from a relatvely low change for the USA to a hgh change for Portugal, whch we argue s due to a relatvely hgh nvestment n energy effcent R&D capal n the USA relatve to Portugal, such that an addonal ncrease n such nvestments tend to contrbute less and less to the total mpact n the USA than does n Portugal. Ths argument s supported by the hgher cumulated ( ) predcted mpact of energy effcent R&D capal on energy demand of approxmately -20% for the USA relatve to -15% for Portugal. Our analyss shed lght on the mpact of energy effcent R&D capal on energy demand whch can be mportant for polces focusng on energy effcency measures n reducng energy demand, where the focus of the polcy s on energy effcent R&D nvestment. It also hghlght the mportance of spllover effects and other unobserved common factors n nfluencng the estmates f we only rely on the separably assumpton for dentfcaton of prvate/own R&D capal elastcy and return as usually done n the lerature on knowledge producton. It also shows that whle energy effcency measures are mportant, we need other measures to complement effcency measures to acheve szeable reducton n energy demand and the assocated CO 2 reducton. Irrespectve of the evdence of a negatve effect of energy effcent R&D capal, s mportant to stress that ths s only the drect effect. It s possble that the ndrect effect of energy effcent R&D capal could be negatve as a result of the rebound effect va lower energy prces. As a consequence, the total effect from energy effcent R&D nvestment on energy demand could as well be posve dependng on whch of the effects domnates (the drect or the ndrect effect). Ths study only focused on the drect effect. Further s mportant to recognze that, promotng only effcency measures whout mplementng other measures such as taxes and regulaton mght not lead to a sgnfcant reducton n CO 2 emssons f the goal s to reduce CO 2 emssons. 16

18 References Adeyem,O.I., and Broadstock, D.C. (2009). Underlyng consumer preferences and ther contrbuton to energy demand. OPEC Energy Rev. 33 (3 ): Adeyem, O. I., Broadstock, D. C., Chns, M., Hunt, L. C., and Judge, G. (2010). Asymmetrc Prce Responses and the Underlyng Energy Demand Trend: Are they Substutes or Complements? Evdence from Modellng OECD Aggregate Energy Demand. Energy Economcs 32(5): Aroonruengsawat, A., Auffhammer, M., and Sanstad, A.H. (2012). The mpact of state level buldng codes on resdental electrcy consumpton. Energy Journal, 33(1): Arrow, K. (1962). Economc Welfare and the Allocaton of Resources for Inventons, n R. Nelson (ed.), The Rate and Drecton of Inventve Actvy, Prnceton, NJ: Prnceton Unversy Press. Bon, H., and Wllam, P., (1991). Better Buldng Codes through Energy Effcency. Washngton, D.C.: Allance to Save Energy Costa, D., and Matthew, K. (2011). Electrcy Consumpton and Durable Housng: Understandng Cohort Effects. NBER Workng Paper Davs, L.W., and Kllan, L. (2011). Estmatng the effect of a gasolne tax on carbon emssons. J. Appl. Econometrcs 26(7): Eberhardt, M., Teal, F. (2010). Productvy analyss n global manufacturng producton. Economcs Seres Workng Papers, 515. Unversy of Oxford, Department of Economcs. Eberhardt, M., Helmers, C., Strauss, H. (2013). Do spllovers matter when estmatng prvate returns to R&D? Rev. Econ. Stat. 95(2): Geller, H., Harrngton, P., Levne, M.D., Rosenfeld, A.H., and Tanshma, S. (2006). Polces for Increasng Energy Effcency: Thrty Years of Experence n OECD Countres. Energy Polcy 34(5): Gllngham, K., Newell, R., and Palmer, K. (2006). Energy Effcency Polces; A Retrospectve Examnaton. Annual Revew of Envronment and Resources (31): Grlches, Z. (1979). Issues n Assessng the Contrbuton of Research and Development to Productvy Growth. Bell Journal of Economcs 10:

19 Hall, B.H., Jacques, M., and Perre, M. (2009). Measurng the Returns to R&D, NBER workng paper Jacobsen, G. and M.J. Kotchen. (2013). Are Buldng Codes Effectve at Savng Energy? Evdence from Resdental Bllng Data n Florda. The Revew of Economcs and Statstcs 95(1): Karmu, A. and Brännlund, R. (2013). Functonal forms and aggregate energy demand elastces: A nonparametrc approach for 17 OECD Countres. Energy Economcs 36: Metcalf, G.E. and Hassett, K.A. (1999). Measurng the Energy Savngs from Home Improvement Investments: Evdence from Monthly Bllng Data. The Revew of Economcs and Statstcs 81(3): Meyers S, McMahon, J.E, McNel, M., and Lu, X. (2003). Impacts of U.S. federal energy effcency standards for resdental applances. Energy 28: Flppn, M. and Hunt, L. C. (2011). Energy demand and energy effcency n the OECD countres: a stochastc demand fronter approach. The Energy Journal, 32(2): Pesaran, M. H., and Smh, R.P. (1995). Estmatng Long-Run Relatonshps from Dynamc Heterogeneous Panels. Journal of Econometrcs 68: Pesaran, M. H. (2006). Estmaton and Inference n Large Heterogeneous Panels wh a Multfactor Error Structure Econometrca 74(4): Pesaran, M. H. (2007). A smple panel un root test n the presence of cross secton dependence. Journal of Appled Econometrcs 22: Ryan, D.L.and Ploure, A. (2009). Emprcal modelng of energy demand. In: Evans, J., Hunt, L.C.(Eds.), Internatonal Handbook on the Economcs of Energy. Edward Elgar, Cheltenham, U.K. NOAA, Stern D. I. and Kander, A. (2012). The role of energy n the ndustral revoluton and modern economc growth. Energy Journal 33(3): Webber, C.A., Brown, R.E., and Koomey, J.G. (2000). Savngs estmates for the ENERGY STAR voluntary labelng program. Energy Polcy 28: Welsch, H. (1989). The relably of aggregate demand functons: an applcaton of statstcal specfcaton error tests. Energy Economcs 11:

20 Appendx Abref account of the heterogeneous estmators: Suppose we have two set of varables, y the dependent varable and an explanatory varable, x, whch are related va the expresson: y x u u f 1 t x f g s 2 t t A1 Where u s a compose term that comprses both unobserved common factors ( f t,a country fxed effects, 1 ) and a random error term,. Further, s unobserved common factors ( f, g ). t t x s also nfluence by The Pesaran (1995) mean group estmator (MG) allows for heterogeney n the but assumes cross-secton ndependence, as a consequence, the MG estmator s mplmented by estmatng N country regresson as follows: y x t ˆ MG N N 1 1 ˆ A2 Where t s a lnear trend to capture unobserved factors that are tme-nvarant and by constructon, the MG estmator allows for parameters to vary or dffer across panel uns (countres n ths study). The Pesaran (2006) common correlated mean group (CCEMG) estmator, unlke the MG estmator, does not assume cross-secton ndependence. It however allows for heterogeney n the parameters as does n the MG estmator. The CCEMG accounts for cross-secton dependence va cross-secton average of both the dependent ( y ) and explanatory varable ( x ) as addonal covarates as follows: y x y x 1 2 ˆ CCEMG N N 1 1 ˆ A3 19

21 where the cross-secton averages ( y, x ) are to account for the unobserved common factors ( f s t, ft, g t ). The augmented mean group estmator (AMG) s a varant of the CCEMG estmator that accounts for both heterogeneous parameters and cross-secton dependency. The crosssecton dependency s accounted for by the ncluson of a common dynamc effect n the country specfc regressons (Eberhardt and Teal, 2010). The estmaton s done n two stages, where the frst stage s to obtan the common dynamc varable whch s ncluded n the second stage as an addonal covarate to account for the common dynamc effect. The two stages can be express as: stage(1) y x c D A4 k k k 2 cˆ k ˆ t T stage(2) y x ˆ t t ˆ AMG N N 1 1 ˆ A5 where the frst stage s based on frst dfference (FD)-OLS regresson wh T 1 frst dfference year dummes denoted by Dk to get the estmated common dynamc varable ( ˆt ). The estmated common dynamc varable s ncluded n the second stage regresson to capture the effects of potental unobserved common factors. 20

22 Table A1: Un root and contegratng test for each of the seres Un root test Varable lags CIPS p-value MW p-value e e e e y y y y p p p p hhd hhd hhd hhd R&Dcap R&Dcap R&Dcap R&Dcap The null hypothess s that of a un root, the Pesaran (2007) CIPS test allows cross-sectonal dependency n the testng procedure, whle Maddala and Wu (MW) Panel Un Root test does not allow for cross-sectonal dependence. The frst dfference (change) s denoted by and the numbers n parenthess are the P-values for the un root test statstc. Contegraton test (pedron resdual base contegraton) Test Stats Panel Group t ADF All test statstcs are dstrbuted N (0, 1), under a null of no contegraton, and dverge to negatve nfny. There are two groups of the test statstcs panel and group. The group statstcs average over the ndvdual country test statstcs, whle the panel statstcs pool the test statstcs over the tme dmenson. The and t are nonparametrc test statstcs, whle ADF and are parametrc statstcs. The estmatons for the seven test statstcs (,, t, ADF) are done wh the ncluson of common tme dummes to correct for potental smple cross-sectonal dependence. 21

23 Predcted Impact of Energy Effcency R&D Investment on Energy Demand Country-ID Fgure A1: Predcted mpact (cumulated over ) of Eneregy Effcency R&D nvestment on energy demand for 13 OECD countres. Note: the Country-ID, 1=Austra, 2=Belgum, 3= Denmark, 4=France, 5=Italy,6=Netherland, 7=Norway,8=Portugal, 9=Span, 10=Sweden,11=Swzerland, 12=UK, 13=USA 22

24 p 4.8 lhhd e y Varably of the medan value of e across countres Varably of medan value of y across countres Country-ID Country-ID Varably of medan value of p across countres Varably n the medan value of lhhd across countres Country-ID Country-ID Fgure A2: Boxplot showng the varably of the medan value for energy consumpton(e), ncome(y), prce (p) and heatng degree days (hhd) across 13 OECD countres. Note: the Country-ID, 1=Austra, 2=Belgum, 3= Denmark, 4=France, 5=Italy,6=Netherland, 7=Norway,8=Portugal, 9=Span, 10=Sweden,11=Swzerland, 12=UK, 13=USA 23

25 Table A2: Heterogeneous estmates (dynamc) and respectve long-run elastcy MG AMG CCEMG e lag * ** * (0.095) (0.071) (0.096) p ** *** ** (0.033) (0.030) (0.059) p lag *** *** (0.026) (0.023) (0.077) y *** *** ** (0.125) (0.138) (0.253) y lag (0.128) (0.088) (0.197) lnrndstock (0.023) (0.015) (0.044) lhhd *** ** ** (0.017) (0.048) (0.040) Trend yes yes yes Const *** (0.093) (0.431) (1.269) Dagnostcs CD-test [0.224] [0.232] [0.045] Integraton I(0) I(0) I(0) N Note: Standard errors n parentheses (robust standard errors), * p < 0.1, ** p < 0.05, *** p < 0.01, CD-test s the cross-sectonal dependence test base on Pesaran (2004) test for the Null of cross-sectonal ndependent resduals and N s the number of observatons, hhd s the heatng degree days and R&Dcap denote R&D capal. Numbers n square bracket are the p-values for CD-test. Long-run eslastcy estmates MG AMG CCMG p *** *** ** (0.055) (0.049) (0.119) y ** *** ** (0.226) (0.206) (0.397) R&Dcap (0.028) (0.018) (0.052) hhd *** *** ** (0.040) (0.060) (0.049) 24