energy demand and income elasticity: a cross-country analysis · energy demand and income...
TRANSCRIPT
Department of Economics, Umeå University, S-901 87, Umeå, Sweden
www.cere.se
CERE Working Paper, 2013:5
Energy demand and income elasticity: a cross-country analysis
Chandra Kiran B Krishnamurthya, Bengt Kriströmb
aCenter for Environmental and Resource Economics and
Umeå School of Business and Economics, Umeå University, Sweden
bCenter for Environmental and Resource Economics and
SLU Umeå, Sweden
The Centre for Environmental and Resource Economics (CERE) is an inter-disciplinary and inter-university research centre at the Umeå Campus: Umeå University and the Swedish University of Agricultural Sciences. The main objectives with the Centre are to tie together research groups at the different departments and universities; provide seminars and workshops within the field of environmental & resource economics and management; and constitute a platform for a creative and strong research environment within the field.
Energy Demand and Income Elasticity: Across-country analysis
Chandra Kiran B Krishnamurthy∗
Center for Environmental and Resource Economics andUmeå School of Business and Economics,
Umeå University 90187 Sweden
Bengt Kriström†
Center for Environmental and Resource Economics andSLU Umeå
August 13, 2013
Abstract
We provide consistent, cross-country estimates of price and incomeelasticity for 11 OECD countries, using data from a survey conductedby the OECD in 2011. Using data for annual consumption of electric-ity and sample-derived average electricity price, we provide country-specific price elasticity estimates and average income elasticity esti-mates, using the double-log form of the demand function.
For most countries in our sample, we find strong price responsive-ness, with elasticities varying between −0.27 for Korea and −1.4 forAustralia, with most countries’ elasticity being above −0.5. Using theunique nature of the dataset (in having many attitudinal indicators),we find evidence for non-price related factors to significantly affectenergy demand. In particular, we find households’ self-reported en-ergy saving behavior to reduce energy demand between 2 and 4%. In
∗[email protected]†[email protected]
1
contrast, we find very weak income response, with income elasticitiesvarying from 0.07 to 0.14 and no evidence for heterogeneity in incomeresponse across the countries in our sample. Our results regardingprice elasticity are in contrast with many existing studies which findlow-to-moderate price responsiveness, and adds to a few recent stud-ies indicating more policy space for demand reduction than previouslythought.
Keywords: Residential Electricity Demand, Price ElasticityJEL Classification: Q4, Q41, C5, D12
1 Introduction
The issue of climate change mitigation, extremely important in particularfor the EU and selected OECD countries, has led to a resurgence in studiesregarding household electricity demand. Such studies have focused on threedistinct issues; understanding price responsiveness, appliance choice and theimpact of policy on energy demand, including the issue of “rebound” effects.
Most studies on the first and third issue work with two types of data,household level or aggregate (local, regional or state/province/country-level).As indicated in, for instance Fell et al. (2011), aggregate data are able todistinguish larger-scale patterns, such as regional variation in key param-eters, which household data typically do not allow one to infer, primarilydue to the lack of panel household data sets. On the other hand, parame-ters estimated from such aggregate studies are difficult to interpret or applyto policy, in addition to suffering from well known aggregation and otherbiases.
Household data sets however are being increasingly used in recent studiesin the US (Fell et al. (2011); Alberini et al. (2011)) to provide detailedestimates of household-level parameters at the regional level. In general,the substantial heterogeneity which exists in consumption behaviour acrossspace is acknowledged but rarely addressed.
The current study attempts to address, for the case of 11 OECD coun-tries, the issue of estimation of price and income elasticities of electricity
2
demand, using a sizeable cross-section of survey data. This is, so far as weare aware of, the first systematic attempt at assessing important demand pa-rameters from household data in a cross-country setting. In particular, thisstudy addresses the following issue: estimation of income and price elastic-ities, exploring both the commonly used double-log and a more reasonablealterantive, the trans-log functional form. In addition, the issue of possibleparameter heterogeneity is addressed by estimating country-specific coeffi-cients on income and price, using the popular double-log functional form.Finally, the study also provides some indications of the potential empiri-cal significance regarding non-price-related factors (“attitudinal” factors),preference heterogeneity and the issue of split incentives.
It is important to mention at the outset that the price elasticities areestimated here based on a small sample, with average prices derived fromthe survey data; as a result, these elasticity estimates need further valida-tion. Unlike in recent papers which have used either utility-level aggregatedata (Fell et al. (2011), Alberini et al. (2011)) or instruments (Alberini &Filippini (2011)), our data set does not provide information regarding ei-ther utility-level or even region-level price (or other utility data), and wecan use neither of the above strategies. We therefore derive the household-specific average price from (possibly mismeasured) survey-reported annualquantity estimates. We note that consumers in our survey are in generalbilled on a per-unit basis, and indicate that they are aware of the marginalprice (about 90% of the sample, see Kriström (2013)); nonetheless, there islittle evidence regarding the basis (marginal or average price) upon whichconsumption decisions are made.
In using average price data (at the individual level) we assume, as in someof the recent literature (Borenstein (2009); Fell et al. (2011); Ito (2012)),that consumers make decision based on average, not marginal prices. Inour case, in common with Alberini et al. (2011), we resort to using averageprice due to missing data on marginal price. We are unable to obtain dataregarding marginal price (and fixed components of price) both due to thesize of the task (of collecting utility-region level prices for the many differentcountries), as in (Alberini et al., 2011), as well as due to lack of data on the
3
utility serving the individual households. We stress that we do not take anyposition regarding which measure of price, marginal or average, is actuallyused by individuals in our sample to make their decision.
However, we emphasise that our identification of the income elasticity(as well as other relevant parameters) does not depend on which price is usedby consumers to make these decisions in our main specification, the double-log functional form. Since the double log functional form, in addition to itswell documented drawbacks (see for instance Plourde & Ryan (1985)), hasan added disadvantage that it cannot be consistent with any utility func-tion when price (or income) elasticities are close to unity1, we also provideelasticity estimates from an alternative and encompassing functional form,the translog, which does not suffer from such drawbacks.
It is pertinent to point out that in a large number of countries in thesample, electricity pricing does not follow the non-linear increasing blockpricing so prevalent in the US; as a result, the debate about marginal-versus-average prices induced by the complexity of rate schedules is less important 2.In particular, therefore, the issue of endogeneity between price and quantityof electricity is not relevant. Nonetheless, as a robustness check, we excludethe data for a few countries which follow some form of block-pricing and showthat our estimates are qualitatively unaffected, with (almost) unaltered priceelasticites and only slighly increased income elasticity.
We briefly summarize our main results here. We find a very high, in cer-tain cases unitary, price elasticity for most countries in the sample (rangingfrom −0.25 for Korea to −1.4 for Australia), with most countries (except
1To see this, note that the indirect utility function for an individual with the double-log demand function can be written (Hausman (1981, equation (21))) as: v(p, y) =
−eΓX p+ γ
1 + γ+ y1−β
1− β , using notation from eq. (2). It is evident that, for γ → −1, this
utility function is infinite, ∀(p, y). We note that Fell et al. (2011) ignore this issue in theirestimation using the double-log form, an issue particularly relevant for them due to theirfinding of a unitary price elasticity.
2Indeed, excluding taxes and other fixed charges, which can form a substantial partof the electrcitiy bill, the average and marginal prices coincide, in our setting. Morerealistically, average and marginal prices can still be substantially different, with varyingdifferences over time (months of year and year) since “fixed charges”, such as energyescalation surcharges, on the bill can also be variable over time.
4
Korea and the Netherlands) having a high price elasticity. In particular,Netherlands and Korea are the only countries with an elasticity below −0.5,about −0.25 and − 0.26 respectively. We also find low, between 0.05 and0.12, and rarely significant, income elasticity, and further, in our main spec-ification, income elasticity does not vary across countries. In addition, wefind only modest differences between home owners and renters in terms ofelectricity demand, after controlling for other observables including totalnumber of appliances.
Illustrating the potential significance of non-price-related variables, wefind that an increase in the energy behaviour index reduces energy consump-tion by 2 to 4%, while membership in an environmental organization has nosignificant impact on demand. We also find the usual relationships to holdwith typical control variables: households with: greater size, larger homes,more appliances and using electric-based heating (and/or cooling) use moreelectricity. Finally, all our results are robust to a range of factors, includingfunctional form, with very similar results obtained using the more generaltrans-log demand function.
While the income elasticity estimated here is well within the range ofestimates for other settings in the literature, we find the price elasticites tobe very large in comparison to most of the recent literature. Only Fell et al.(2011) and Alberini et al. (2011), among recent studies, report elasticitiesas high as we do.
The plan of the paper is as follows. We lay out the econometric ap-proach in section 2, discuss the data and summary statistics in section 3and present our results in the context of existing literature in section 4.Section 5 provides a discussion and concluding remarks.
2 Econometric Approach
Denote by Ei,j , Qi,j & Pi,j the levels of (annual) expenditure (euro), quan-tity consumed (KwH) and (marginal) price of electrcitiy (euro) faced byconsumer j = 1, . . . , Ji in country i = 1, . . . , I. Data on Ei,j is available,while that on one of Qi,j , Pi,j is not. Posit the usual relationship between
5
Qi,j and Pi,j
ln(Qi,j) = β ln Ii,j + γ lnPi,j + ΓXi,j + εi,j (1)
where X is a vector of control variables, including appliance holdings,individual characteristics and other relevant non-economic factors. Writ-ing out explicitly the equation for expenditure, Eij = QijPij , in log andsubstituting eq. (1), we have
ln (Ei,j) = β ln Ii,j + (1 + γ)︸ ︷︷ ︸γ̃
ln(Pi,j) + ΓXi,j + εi,j (2)
Thus, eq. (2) is easily estimable, with data on P and E. In a large number(seven) of the countries sampled here, endogeneity of price is not an issue,given the constant marginal price (with substantial variation at the regionand home-type level). Presuming data on quantity consumed (KwH) areavailable (as in our case), it is possible to compute P̄i,j := Qi,j
Ei,j. Provided a
sample of data on Ei,j and P̄i,j is available, it is straightforward to estimatethe parameters of interest, γ and β, the price and income elasticity3,4.
We note here that we have not indexed any of the coefficients with i, butthat is purely for notational convenience; in the setting above, equations (1)-(2) may well be thought of as being country-specific, with country-specificcoefficients. This speaks to our issue of addressing (observed) heterogeneity;estimated elasticities vary substantially over countries, even within a similarmethodology (provide citation). In addition, for the larger countries, thereis likely to be substantial heterogeneity within a country, and finally, acrosscountries, heterogeneity in preferences across individuals.
Our strategy for addressing this issue is the following: we do not wish3Note that the coefficient on lnPi,j in the equation for expenditure, denoted γ̃, is not
the elasticity of interest, γ; evidently, γ = γ̃ − 1, and its standard error can be computedfrom that of γ̃.
4Using data on E and Q to estimate P , as we do, it is then evident that both eq. (1) andeq. (2) yield identical estimates. We write out both equations for purposes of comparabilityto the framework in Fell et al. (2011). In subsequent developments, we will always estimatethe quantity equation, eq. (1).
6
to impose parameter homogeniety in our estimation of eq. (2) (eq. (1))above. Instead, we take the approach of testing for parameter homogeneity.The strategy is to test for similarity of coefficients between the unrestrictedmodel, with many coefficients being country-specific, and a restricted modelwith only country-fixed-effects, a traditional approach to addressing hetero-geneity5.
In more detail, we consider testing between the following two models:
ln (Ei,j) = αi + βi ln Ii,j + γ̃i lnPi,j + ΓXi,j + εi,j (Unrestricted)
and
ln (Ei,j) = αi + β ln Ii,j + γ̃ lnPi,j + ΓXi,j + εi,j (Restricted)
It is then straightforward to estimate the two models and to test thefollowing hypothesis:
H0 : γ̃i = γ̃ & βi = β ∀i (3)
This approach, while intuitive, also illustrates the limitation of this frame-work. There is no reason to restrict differences in elasticities to be onlydependent on the country, any other observed characteristic (such as posses-sion of an energy efficient appliance, of being a home owner, or belonging toan environmental organization etc) can, in addition, be envisaged as givingrise to differences. At the extreme, such an approach would lead to an esti-mation of a very large number of parameters, one for each country-categorycombination. In order to obtain interpretable and usable results, we restrictsuch testing to only country-specific coefficients, although interactions be-tween other key variables (possession of an energy efficient appliance and ofbelonging to an environmental organization were the two indicators postu-lated as being relevant, one at a time) and countries were not found to be
5A more efficient alternative, the SUR framework, exists; however, in our case, theestimation sample is so unbalanced that implementing the SUR presented great difficultiesin estimation (due to the difficultues involved in computing the covariance matrices).
7
significant.It is to be noted that the coefficients in equations estimated above (in
eq. (Unrestricted)) are not equivalent to country-by-country estimation, fortwo reasons: there are many cross-equation restrictions (most coefficients arerestricted to be equal across countries) and the error variance is identicalacross countries. An important advantage of estimating demand parametersfor many different countries in the same framework and using comparabledata is the consistency in estimation and resulting comparability of param-eters across countries.
Yet another issue worth highlighting is the dependence of interpretationon functional form: for a general functional form (even, for instance, alinear form), elasticities are a function of the level of a variable, as well aspossibly of all other variables in the model. As a result, we present estimatesof elasticity using an alternative and encompassing functional form, thetranslog, more as a check of robustness of results rather than presentinga preferred functional form6.
We provide results using the trans-log functional form:
ln(Qi,j) = β0 ln Ii,j + γo lnPi,j + ΓXi,j + β1 (ln Ii,j)2 +
γ1 (lnPi,j)2 + δ (lnPi,j ln Ii,j) + εi,j(4)
Even in the absence of country-specific coefficients, the elasticities ofprice and income each depend on all other parameters and variables and, inparticular, on each other. The price and income elasticities, denoted ηP , ηIrespectively, are: ηPij = γ0 + 2γ1 lnPi,j + δ ln Ii,j and ηIij = β0 + 2β1 ln Ii,j +δ lnPi,j . Thus, each observation has its own elasticity.
In this functional form, on the other hand, there is no consistent way ofspecifying country-specific coefficients7. Therefore, for this specification, we
6The linear functional form was also estimated, and found to have a significantly worsefit than either the double-log or the tranlog. As a result, we do not provide the results ofestimation with this functional form. In general, the elasticities were of the correct sign,and of similar magnitude to the other two; however, they are very imprecisely estimatedand are rarely significant. These results are available from the authors upon request.
7If country-specific coefficients are postulated on price and income, then all transfor-mations of price and income, including the squared and interaction terms, must also have
8
do not allow for country-specific coefficients on any variable; nonetheless,since the price (income) elasticities are, in addition to their own levels, afunction of income (price), there is substantial variation in country-specificprice (income) elasticities. This variation is essentially a reflection of thecountry-specific variation in price (income).
We report, for eq. (4), country-specific (and overall) mean elasticities,computed at the country-specific mean of price and grand mean of income(with standard errors computed using the delta-method) are reported. Moreprecisely, ηPi = γ̂0 + 2γ̂1 lnPi + δ̂ ln I and ηI = β̂0 + 2β̂1 ln I + δ̂ lnP are re-ported8. Finally, since for four countries (Korea, Australia, Japan and Israel)in the sample, certain regions (or all of Korea, which is served by a singleutility), increasing-block-pricing is practiced; we exclude these countries andre-estimate eq. (1) and eq. (4). We emphasize that this is in the nature of arobustness check.
3 Data Description and Summary Statistics
3.1 The Survey
Data for the analysis was drawn from the OECD’s project on GreeningHousehold Behaviour, as part of which a periodic survey on Environmen-tal Policy and Individual Behaviour Change (EPIC), covering a number ofcountries and areas, is carried out. The second survey was conducted in2011 and included 11 countries: Australia, Canada, Chile, France, Israel,Korea, Japan, the Netherlands, Spain, Sweden and Switzerland. We pro-vide a very brief description of the survey and refer to OECD (2013, AnnexB) for details.
About 1000 individuals in each country were surveyed using an internet-based questionnaire, for a total sample size of 12,200 households. The
country-specific coefficients, for sake of consistency. Given the explosion of parameterswhich results (an additional 30 parameters, compared to equation (2)), we do not pursuethis route here.
8We do not report country-specific income elasticities since they turn to be almostidentical, when evaluated at the country-specific means.
9
questionnaire collected information regarding household behaviours in fivedistinct areas (apart from household characteristics and environmental at-titudes): residential energy use, waste generation and recycling, food con-sumption, personal transport, and water consumption. The present analysisuses data from the energy section.
Sample selection followed a strategy of stratification based on income,age-group, region and gender. In order to account for sampling-related is-sues, ex-post probability weights were provided, which may be used to ren-der estimation results using this sample comparable to those using randomsamples from country-level population distributions.
We note that individuals were requested to provide data on their elec-tricity bill (annual) and quantity consumed in Kwh (annual). Very few in-dividuals provided billing data, and of those, a few provided quantity data,allowing computation of the average price. As a result, the final sample size,about 1100, is a fraction of the usable responses of approximately 11, 000households.
3.2 Variable Descriptions and Summary Statistics
We provide a brief discussion of selected summary statistics for the regressionsample, as well as discuss prior expectations (based on existing literature)regarding the signs of many of the control variables, appearing in eq. (1)9.
We begin with summarizing the reported electricity usage and billingdata. The mean expenditure varies widely across countries, from a low of170 Euro in Korea to a high of 2000 in Sweden, which is consistent withthe variance in usage, Korea being the lowest and Sweden, the highest. The
9We carry out no particular data cleaning exercises such as dropping outliers etc forour analysis below, despite the presence of a few outliers. Our primary concerns centeredaround (i) wildly overstated prices (ii) substantial changes for households which reportlow spending and (iii) very different behaviour of high income households from those withvery low income. To account for these issues, we carried out analyses winsorizing our dataat the 5% threshold, for price, income and quantity. On repeating the analyses with thesedata for the double-log functional form, we found no qualitative difference in results andminor quantitative differences. The only noteworthy difference was the wider variationof income elasticity and some evidence to the effect that income elasticity should also becomputed by country.
10
derived average price mirrors the variation above; noteworthy however isthe relatively low Swedish electricity prices, at 19 (euro) cents, compared toboth the Netherlands and Spain (of the EU members). Overall, it is evidentthat colder countries (Canada, Sweden and Switzerland) typically consumemore electricity, while a variable average price and possible existence ofsubstitutes render cross-country comparisons tricky. All of these figuresare consistent with country-reported data (see Kriström (2013) for a moredetailed analysis of the EPIC 2011 data on energy).
Other key control variables in our regression include: household char-acteristics such as household size, home size, age of respondent, years ofpost-secondary education, urban location, behavioral attributes such as en-ergy behaviour index and membership in an environmental organization,appliance characteristics such as possession of at least one top-rated energyefficient device (labelled “energy efficient device”), appliance stock and anindicator for electric space cooling/heating . We summarize them, in table 1and table 2, by country in order to highlight the cross-country variation andexhibit country-specific characteristics which motivate us, in our regressions,to allow for (observed) heterogeneity in demand parameters across countries.
Average income, at 39800 Euro, is consistent with measures derivedfrom national income statistics but certain countries, Korea and Switzerlandin particular, appear to have far higher income than respective country-averages. In addition to issues with exchange rate dynamics (a concernparticularly for Switzerland over this period), this highlights a strong urbanand high income bias with Korea, with 71% of the sample residing in ur-ban areas. Chile has the lowest income in the sample countries (average of14800), but also suffers from a strong urban bias, with 75% of the samplelocated in urban areas, in addition to having higher-than-average years ofeducation, at 4.1.
There are smaller variations in household size and years of educationacross countries, on average, with Chile, Netherlands and Japan havinghouseholds with more years of education of the respondent (above 3.5) andChile, Israel and Korea having larger family sizes (above 3.5). Chile andIsrael have the youngest responders, on average (average age below 42 years)
11
while Sweden and Switzerland have the oldest (average age about 51). Wenote that the sample average age of 48 is likely lower than in other surveysdue to the exclusion of respondents older than 70 (65 for Korea) from thesample, as well as possibly due to an internet bias.
Home size (size of primary residence) varies widely across countries, with(as anticipated) Korea and Japan having the smallest (at 97 sqr. mtr.) andAustralia by far the largest, at 174 sqr. mtr.
Home ownership does not vary substantially, with only Japan and Switzer-land having significantly lower rates than the average, of 76%. Similarly,apart from Chile, Korea and Israel, there is not a substantial urban bias,with a mean urban residence of 36%. We categorize homes into two typesfocused on in the literature, flats or “multi-dwellings” and (semi-)detachedhouses, and find that there is substantial variation across countries alongthis dimension. Australia has the smallest proportion of households in flats(5%) and Spain the largest (73%), with most countries in between.
Finally, membership in environmental organizations, postulated to be akey preference attribute impacting energy demand, varies across countries,from almost no participation in Japan (5%) to more than a quarter of thesample in Switzerland (28%). Overall, we find that many of the key variablessuch as price, consumption, income, membership and size of home, varywidely across countries.
We turn next to summarizing our prior expectations, based on exist-ing literature, regarding the effect of control variables on energy demand.Income, members in household, and size of residence (in sq. mtrs) are typi-cally thought of as positively related to electricity consumption, with pricenegatively related. Recent literature (provide citations) postulates electricspace heating (cooling) and age to be positively related, the latter operatingpossibly through the channel of energy efficiency and habits, while urbanresidence, residence in apartments (“multi-dwelling”) and home ownershipare posited to negatively influence electricity demand.
These factors essentially reflect the spatial and incentive structures, sincetypically, urban residences have higher density and lower area while homeownership alters incentives for investment in energy efficient appliances, the
12
so-called “split incentives” issue (provide citations). It is important to em-phasize that some of the these indicators act through similar channels, andit is sometimes difficult to interpret these variables in isolation10.
There has been a recent emphasis on (provide citations) the impor-tance of non-price-related “attitudinal” characteristics determining energydemand (see for instance Martinsson et al. (2011) andWilson & Dowlatabadi(2007)); all of these essentially involve some form of inertia (“habits”) or“lock-in” and posit different behavior depending on attitude towards theenvironment (see also Maréchal (2010)). To capture these effects, if theyexist, the OECD survey provides at least two indicators11: membership inan environmental organization, and an energy behavior index. The former isan indicator for unobserved differences in behavior related to environment,and by analogy to energy, while the latter is a more direct, self-reportedmeasure of certain measures taken to save energy. This index is definedto lie between 1 and 10, with higher values indicating greater tendency tosave energy. Energy behaviour index has comparatively lower variabilitythan many other variables, with only Spain, Chile and Korea having highscores, above 8; somewhat surprisingly, Sweden has the lowest score by far,at 5.5, the only country below 6. Both variables are anticipated to lead toreductions in total electricity usage, conditional on other factors.
4 Results
We begin with discussing the results from our preferred specification, thedouble-log, wherein country-specific coefficients are allowed on price and all
10For instance, higher income households tend to have larger homes and to be moreurban; thus, many of these variables can sometimes represent very similar effects. However,dropping certain variables which had identical channels of effect did not appreciably alterresults, and therefore, we do not refer to this issue in subsequent developments.
11There are many other indicators which can be used to indicate different attitudes toenvironmental issue, both general and specific, such as importance of environmental issues,importance of climate change etc. None of these indicators, however, are directly relatedto the issue of energy and further, if an unambiguous indicator of differences betweenindividual preferences are to be used, a revealed measure, such as membership or energybehaviour index, is likely more preferable to ambiguous measures such as importance.
13
other coefficients are restricted to be identical across countries12. Results arereported below in table 3 and table 4. For models with probability weights,the covariance matrix estimated is the standard Huber-White sandwich ma-trix while for the model without weights, we estimate the more robust boot-strap standard errors, with independent resampling within countries. Wenote that both approaches account for heteroscedasticity across countries.
In all tables below, the “Model Significance Test” reported is a (joint)Wald test of all coefficients except the country-specific fixed-effects; the teststatistic is either χ2 or F−distributed for, respectively, unweighted andweighted models, and in all cases we report, are significant. The “ModelComparison Test statistic” reported is a (joint) Wald test of equality of allcoefficients, i.e. the test indicated in eq. (3); as for the Model SignificanceTest, the test statistic is either χ2 or F−distributed and, in all cases here,is significant.
The test results for individual coefficients indicated that, except for priceand home ownership, all other coefficients, in particular on income, wereidentical across countries. For brevity, we report results from a model inwhich only the price coefficient was allowed to vary over countries, notingthat allowing home ownership to vary across countries does not substantivelyalter the results on price and other coefficients.
The column reading “Exogenous-only”, in all tables below, indicatesregressions restricted to only countries in which increasing-block-pricing isnot applied for any region/utility.
We turn next to discussing the results and begin with by noting that, fordummy variables in double- (or semi-) logarithmic equations, the coefficient
12We note that regional variation in electricity pricing, as well as other characteris-tics, within countries is quite substantial; the more robust approach to dealing with thesewould be to use country and country-region specific fixed effects, in the pooled models,and country-region-specific fixed effects in models with country-specific coefficients. Inthe results we report, we do not take such an approach due to the substantial increasein the number of parameters estimated, relative to the sample size; nonetheless, in unre-ported results, we estimate versions of models with these fixed-effects and find virtuallyno qualitative difference in our results. The only change, as anticipated, is slightly largerstandard errors, due to the substantial number of additional parameters (approximately90) estimated.
14
on dummy variables can be interpreted as a proportional effect, approxi-mately13. We find, in most specifications (from table 3, table 5 and table 6)that the coefficients on these control variables accord with both intuitionand prior results, whenever significant. Household size and home size (notshown) are positvely associated with energy demand, and so are the numberof appliances.
Urban residence is negatively associated with demand, with a coefficientwhich varies in magnitude depending upon the specification. In unweightedspecifications, using the full sample, the effect is quite strong, about 9−11%and significant; restricted to exogenous-only countries, the effect is a moremodest 7% and not significant. Overall, it is clear that the impact of urbanresidence varies across countries and that larger sample sizes are possiblynecessary before any definitive assessments can be made14.
Age is positively associated and is significant while education is neversignificant. Home ownership is never significant, for all samples and spec-ifications, and we speculate that it is due to the inclusion of the channelthrough which it operates, presence of energy efficient device(s). It is curi-ous that the dummy for presence of at least one top-rated energy efficientappliance, while of the correct sign in most specifications, is never significant.This result is similar to that in Alberini et al. (2011, Table 9).
A surprisingly strong result is that regarding the difference betweenapartments and (semi-) detached housing; across all specifications and alldata configurations, residence in apartments has a very strong negative ef-fect on electricity demand, varying between 22% and 38%, and always sig-nificant. Finally, in common with Alberini et al. (2011), we find that having
13Let D be a dummy variable for urbanisation, and the effect of urbanisation on elec-tricity demand as a proportion be u = Q1 −Qo
Q1where 1 and 0 indicate respectively urban
and non-urban data points, and c be the estimated coefficient on D in eq. (1), for instance.Then, as indicated in Halvorsen & Palmquist (1980), c 6= u, in general; to be more precise,u = exp(c)−1. In our case, the difference between c and u (computed from the regressionresults), for all dummy variables included in the regression, were found to be very small,and we therefore discuss and interpret c (the estimated coefficient) as if it were u (theproportional effect of interest).
14A conclusion which is hinted at in the weighted regressions in which results are alwayssignificant and substantially higher in magnitude.
15
electric heating (and/or cooling) makes a substantial impact on electricitydemand; across specifications, it varies between 22 and 41%, and is alwayssignificant.
We turn now to understanding the impact of attitudinal determinantsof energy demand, membership in an environmental organization and en-ergy behaviour index. From the results in table 5 we see that while energybehaviour index is indeed negatively significantly associated with energydemand, membership in an environmental organization is not, in any speci-fication. The magnitude of the impact, between 2 and 4%, depending uponthe specification, indicates a substantial impact of increases in the index.
Turning now to the key effects of interest, price- and income-elasticity,we note that income elasticites are rather low, and vary widely across spec-ifications, between 0.08 and 0.14. Typically, when weighting is resortedto, income turns insignificant, but is significant in all specifications with un-weighted data. The magnitude, and variability based on specification, of theincome coefficient are both consistent with, for instance, Fell et al. (2011)and Reiss & White (2005) as well as with much of the existing literature.
Our key result (column 4 in table 4), however, is a unitary price elas-ticity for four countries, Australia, Canada, Chile and Spain and very high,but less-than-unitary elasticities for France and Israel. Only in the caseof Korea and the Netherlands are elasticities below 0.5. Price elasticitesas low as this are well beyond the typical short-run elasticites reported inthe literature, which are much lower, about 0.4 or below. Only Fell et al.(2011) and Alberini et al. (2011), using average prices, document very highprice elasticities, unitary in the former and between −0.6 and − 0.87 in thelatter. Borenstein (2009) and Ito (2012), using average prices, find muchsmaller elasticites, less than 0.3, using a sample restricted to only (parts of)California. However, some of the older literature, beginning with Halvorsen(1975) and summarized in Bohi & Zimmerman (1984, Table 1, pp 116-118),find a substantially higher price elasticity, closer to the ones obtained in thisstudy.
We provide evidence to the effect that our results are robust to en-dogeneity of price and functional form chosen, and are not an artifact of
16
our accommodating for heterogeneity (in estimating country-specific coeffi-cients). When four countries with suspected endogeneity issues are excludedfrom the regression sample (column 8 in table 4), the three remaining coun-tries (Canada, Chile and Spain) nonetheless have almost identical (unitary)elasticities as before15.
An alternative way to see this is using the pooled model, where all co-efficients (except the intercept) are asssumed identical across countries, andcountry-specific fixed effects are used to allow intercepts to vary. In thisspecification (table 5, columns 2 and 4), we note two important points:first, that the pooled model yields unitary income elasticity (−0.99), withthe full sample, and second, that excluding the countries with suspected en-dogeneity from the regressions increases the elasticity slightly, to −1.18. Inother words, the very high estimates of elasticity are not an artifact either ofendogeneity in a part of the sample or a result of allowing for country-specificelasticities.
Finally, given the substantial variation in elasticity estimates across func-tional forms and in general, the somewhat unrealistic nature of the double-log form (as noted in Plourde & Ryan (1985) and Ryan & Plourde (2008),among others), we estimate a more general and encompassing trans-log func-tional form16. Results (column 2, table 4) clearly indicate very high andunitary price elasticities; indeed, the number of countries with unitary priceelasticities now rise by 2 (Korea and Sweden) and in general, the trans-logprovides slightly higher elasticities than the double-log form. These resultsare unaltered when countries with suspected price endogeneity are excluded(column 6)17.
15We recall here that our estimation is not equivalent to country-by-country estima-tion. Nonetheless, estimates of coefficients which are unrestricted typically are not alteredsignificantly by moderate changes to the sample, such as we undertake here. It is nothowever, always the case, as may be observed by a comparison of the coefficient, underthe trans-log specification, for the Netherlands.
16The trans-log has been found to provide more reasonable price elasticity estimates thanthe double-log for other fuels, for instance, gasoline (Wadud et al. (2010); Liu (2011)).
17We note that, for all model specifications for the translog, we are able to reject thefollowing null, via a Wald test: β1 = γ1 = δ = 0, which is essentially a test of the double-logversus the trans-log functional form. We do not emphasize the model selection aspect here,since both models provide results which are qualitatively consistent and quantitatively
17
We turn next to providing a brief comparison of our results to those avail-able in the existing literature. We note that for the Netherlands, France andChile, we have been unable to obtain references to any studies and so can-not provide explicit comparisons. An explicit comparison is also complicatedby the very different methodologies (time-series versus panel versus cross-section); nonetheless, we provide comparisons with short-run estimates ofelasticity wherever possible.
The only study for Australia using individual-level data (at the half-hourly frequency), Fan & Hyndman (2011), found relatively low price elas-ticities, between −0.36 and −0.43. However, it is rather difficult to compareour estimates to that in this study due to the unrepresentative sample inthat study (drawn from only one town in Southern Australia). For Canada,our estimate, of −1.1, is far larger than the short-run coefficients in the cur-rent literature, which vary between −0.3 and −0.67 (Bernard et al. (2011,Table 4)) but are quite close to the long-run elasticity of about −1.3.
For Israel, our estimate of −0.85 is higher than the range(−0.21 to −0.59) reported in Beenstock et al. (1999), while for Japan, our estimates of−0.5 to −0.8 are much lower (in absolute values) to that in Nakajima (2010),of −1.3, but are comparable to those in the prior literature (Nakajima (2010,Table 1)), which lie between −0.37 & − 0.47.
For South Korea, estimated price elasticity varies widely between thetrans-log and the double-log, at −1.04 and −0.27 respectively, possibly as aresult of data collection or sample selection issues (already referred to above).Neither of our estimates accord with the estimated (long-run) elasticity of−0.5 , in the most recent study, Saad (2009), but the double-log estimateis almost identical to the that obtained obtained for Seoul (−0.24) in Yooet al. (2007). For Spain, our estimates, for all specifications, are quite close,between −0.9 and −1; however, existing literature in Spain using householdlevel data report widely varying elasticities, between −0.87 (LABANDEIRAet al. (2006)) and −0.25 (Labandeira et al. (2012)). Our estimates are closerto the former, estimated for survey data, as is ours.
For Sweden, our estimates, which vary between −0.68 and − 1, are very
comparable.
18
similar to those reported in the literature using household data. Dams-gaard (2003) provides varied estimates, based on type of heating, between−0.37 and −1.35, while Andersson (1997), considering only households withelectric heating, finds a rather high elasticity of −1.37. Finally, for Switzer-land, our estimates vary between −0.49 and − 0.63, which are remarkablysimilar to the short-run elasticities estimated using house-hold data for alarge sample, in Filippini (2011), which vary between −0.65 and − 0.85.
5 Conclusions
Consistent estimates of price and income elasticity are an important part ofthe understanding the policy space for climate change mitigation in manyOECD countries. There are few existing studies estimating these impor-tant parameters across countries using a consistent household dataset andmethodology. This study is a first attempt at addressing the afore-mentionedshortage.
The primary objective of our study is an estimation of price- and income-elasticity across 11 OECD countries using survey data, while accommodatingobserved heterogeneity through country-specific coefficients. In addition,this study is among a few in the literature on electricity demand to havegone beyond the almost universal double-log functional form, in order toovercome its well known deficiencies, using in addition the encompassingtrans-log demand function.
Unlike in some of the existing literature, we find very strong price re-sponsiveness, ranging from a low of −0.25 (for the Netherlands) to a highof −1.4 (for Australia). In general, except for the Netherlands and Korea,we find price responsiveness to be substantial. In contrast, our estimatesof income elasticity, varying between 0.07 and 0.14 (and always significantin the unweighted specifications), are quite low but consistent with those inthe existing literature.
In addition, due to the unique nature of the dataset, we find that non-price related attitudinal measures are also of some importance in explainingenergy demand, a theme increasingly being explored in the literature. In
19
particular, self-reported energy saving measures undertaken by householdsreduce energy demand between 2 and 4%, depending upon the specification.
Our findings suggest that households, even in the short-run, appear toexhibit substantial response to average price changes. However, before thesecan be taken directly to policy (as in Fell et al. (2011)), two important issueshave to be addressed. First and foremost, what prices do consumers actuallyrespond to? Evidence that consumers respond to average prices is rathersparse and rely on very low samples in specific regions. Second, obtainingestimates similar to those we do in a panel context (with its attendantbenefits), would likely provide a stronger evidence base upon which policiesmay be designed.
In addition, while electricity demand is possibly more price sensitivethan hitherto assumed, welfare implications of such changes will have to beexplored, in order for standard energy policy tools to be both equitable andefficient. Finally, more relevant non-price measures will need to be developedand their significance explored before these measures may be translated intopolicy, given the limited impact that such measures have so far had in reality.
In our research here, we have only addressed some of these issues. Afeasible and possibly immediate extension of our work would be to coupletwo rounds of the OECD surveys (the next round begins in 2014) to accom-modate (the well-documented) variability over time of price- and income-responsiveness.
References
Alberini, A. & Filippini, M. (2011). Response of residential electricity de-mand to price: The effect of measurement error. Energy economics, 33(5),889–895.
Alberini, A., Gans, W., & Velez-Lopez, D. (2011). Residential consumptionof gas and electricity in the us: The role of prices and income. EnergyEconomics, 33(5), 870–881.
Andersson, B. (1997). Essays on the Swedish Electricity Market, chapter
20
Electricity Demand–A Study of the Swedish Residential Sector. PhDThesis, Stockholm School of Economics.
Beenstock, M., Goldin, E., & Nabot, D. (1999). The demand for electricityin israel. Energy Economics, 21(2), 168–183.
Bernard, J.-T., Bolduc, D., & Yameogo, N.-D. (2011). A pseudo-panel datamodel of household electricity demand. Resource and Energy Economics,33(1), 315–325.
Bohi, D. R. & Zimmerman, M. B. (1984). An update on econometric studiesof energy demand behavior. Annual Review of Energy, 9(1), 105–154.
Borenstein, S. (2009). To what electricity price do consumers respond?residential demand elasticity under increasing-block pricing. PreliminaryDraft April, 30.
Damsgaard, N. (2003). Regulation and Deregulation of Electricity Markets,chapter Residential electricity demand: Effects of behavior, attitudes andinterest. PhD Thesis, Stockholm School of Economics.
Fan, S. & Hyndman, R. J. (2011). The price elasticity of electricity demandin south australia. Energy Policy, 39(6), 3709–3719.
Fell, H., Li, S., & Paul, A. (2011). A new approach to estimate residentialelectricity demand using household expenditure data. Resources for theFuture Discussion Paper, (pp. 10–57).
Filippini, M. (2011). Short-and long-run time-of-use price elasticities in swissresidential electricity demand. Energy policy, 39(10), 5811–5817.
Halvorsen, R. (1975). Residential demand for electric energy. The Reviewof Economics and Statistics, 57(1), 12–18.
Halvorsen, R. & Palmquist, R. (1980). The interpretation of dummy vari-ables in semilogarithmic equations. American economic review, 70(3),474–75.
21
Hausman, J. A. (1981). Exact consumer’s surplus and deadweight loss. TheAmerican Economic Review, 71(4), 662–676.
Ito, K. (2012). Do Consumers Respond to Marginal or Average Price? Evi-dence from Nonlinear Electricity Pricing. Technical report, National Bu-reau of Economic Research.
Kriström, B. (2013). Greening Household Behaviour: Overview from the2011 Survey, chapter Household behaviour and energy use, (pp. 77–112).OECD Publishing.
Labandeira, X., Labeaga, J. M., & López-Otero, X. (2012). Estimation ofelasticity price of electricity with incomplete information. Energy Eco-nomics, 34(3), 627–633.
LABANDEIRA, X., LABEAGA, J. M., & RODRIGUEZ, M. (2006). Aresidential energy demand system for spain. The Energy journal, 27(2),87–111.
Liu, W. (2011). Modelling Gasoline Demand in the United States: A Flex-ible Semiparametric Approach. Technical report, Technical Report. De-partment of Economics, State University of New York at Binghamton.
Maréchal, K. (2010). Not irrational but habitual: The importance of “be-havioural lock-in” in energy consumption. Ecological Economics, 69(5),1104–1114.
Martinsson, J., Lundqvist, L. J., & Sundström, A. (2011). Energy saving inswedish households. the (relative) importance of environmental attitudes.Energy Policy, 39(9), 5182–5191.
Nakajima, T. (2010). The residential demand for electricity in japan: anexamination using empirical panel analysis techniques. Journal of AsianEconomics, 21(4), 412–420.
OECD (2013). Greening Household Behaviour: Overview from the 2011Survey. Technical report.
22
Plourde, A. & Ryan, D. (1985). On the use of double-log forms in energydemand analysis. The Energy Journal, 6(4), 105–113.
Reiss, P. C. &White, M. W. (2005). Household electricity demand, revisited.The Review of Economic Studies, 72(3), 853–883.
Ryan, D. L. & Plourde, A. (2008). Energy Demand Models and Modelling.Technical report.
Saad, S. (2009). Electricity demand for south korean residential sector.Energy Policy, 37(12), 5469–5474.
Wadud, Z., Graham, D. J., & Noland, R. B. (2010). Gasoline demand withheterogeneity in household responses. Energy Journal, 31(1), 47.
Wilson, C. & Dowlatabadi, H. (2007). Models of decision making and resi-dential energy use. Annu. Rev. Environ. Resour., 32, 169–203.
Yoo, S.-H., Lee, J. S., & Kwak, S.-J. (2007). Estimation of residentialelectricity demand function in seoul by correction for sample selectionbias. Energy Policy, 35(11), 5702–5707.
23
Australia
Can
ada
Chile
Fran
ceIsrael
Japa
nKorea
Nethe
rland
sSp
ain
Swed
enSw
itzerland
Total
Income(euro)
4832
3.7
4528
3.5
1447
2.2
3937
5.7
2708
9.6
4418
2.6
2960
6.5
4056
2.7
3322
3.8
4660
2.8
6746
4.6
3975
6.1
(260
50.2)
(252
81.1)
(947
5.5)
(172
61.2)
(155
28.2)
(269
35.1)
(123
39.0)
(156
22.7)
(184
60.4)
(196
39.4)
(289
71.8)
(239
62.7)
Averag
eelect.
price(euro/
kwh)
0.20
80.13
70.25
70.14
80.15
10.21
80.05
710.27
70.28
50.16
40.16
40.19
8(0.169
)(0.124
)(0.267
)(0.113
)(0.097
7)(0.169
)(0.045
2)(0.201
)(0.283
)(0.120
)(0.074
1)(0.183
)
Ann
ualE
lectric
ityBill
(Euro)
807.2
914.2
490.6
840.6
884.2
972.0
167.6
944.4
720.9
1924
.975
9.1
962.9
(583
.3)
(597
.3)
(489
.0)
(528
.5)
(607
.7)
(660
.5)
(149
.6)
(654
.0)
(450
.5)
(136
1.2)
(696
.3)
(877
.0)
Electricity
Con
sumption:
Kwh
5221
.010
601.5
3802
.872
29.8
7317
.453
02.9
3937
.538
85.2
3895
.613
666.6
5557
.567
94.6
(438
6.4)
(121
86.7)
(842
9.0)
(536
8.2)
(578
7.6)
(453
2.6)
(465
2.3)
(250
1.8)
(370
5.0)
(109
51.2)
(582
0.0)
(783
3.3)
Mem
bers
inho
useh
old
2.69
42.38
63.99
32.57
83.39
42.69
13.47
22.71
23.02
52.45
92.68
92.87
7(1.195
)(1.198
)(1.520
)(1.227
)(1.567
)(1.342
)(1.170
)(1.132
)(1.074
)(1.116
)(1.319
)(1.353
)
Size
ofPr
imaryResiden
ce17
4.8
132.5
94.93
109.3
102.7
97.30
96.22
133.7
112.5
120.7
127.7
116.9
(117
.2)
(63.03
)(46.56
)(46.51
)(45.74
)(48.67
)(33.47
)(60.85
)(53.17
)(45.21
)(56.60
)(60.19
)
Age
ofrespon
dent
49.40
52.08
43.76
51.01
42.38
49.43
40.47
48.87
47.25
50.90
50.92
48.38
(13.65
)(11.53
)(12.48
)(11.10
)(12.94
)(12.93
)(11.71
)(11.45
)(12.48
)(12.32
)(12.77
)(12.69
)
Yearsof
Post-Secon
dary
Educ
ation
33.05
74.00
72.55
13.32
73.84
03.05
74.23
33.49
22.40
42.90
63.28
1(2.559
)(2.108
)(2.462
)(2.218
)(2.522
)(2.820
)(1.737
)(2.387
)(2.487
)(2.256
)(2.299
)(2.476
)
Energy
Beh
aviour
Inde
x7.91
57.46
58.52
77.89
57.50
37.17
78.01
07.02
48.16
65.55
47.16
67.32
9(1.522
)(1.634
)(1.264
)(1.556
)(1.742
)(1.643
)(1.671
)(1.680
)(1.425
)(1.628
)(1.639
)(1.811
)
Num
App
l15
.11
16.18
13.40
13.40
15.71
16.99
11.89
14.76
14.56
15.06
13.17
14.73
(5.533
)(6.241
)(5.000
)(5.924
)(6.897
)(7.337
)(3.667
)(5.348
)(6.500
)(5.809
)(5.074
)(6.069
)
Table1:
SummaryStatist
ics(M
ean(s.d.))
ofRegressionsampleforCon
tinuo
usVa
riables.
24
Australia
Can
ada
Chile
Fran
ceIsrael
Japa
nKorea
Nethe
rland
sSp
ain
Swed
enSw
itzerland
Total
Hom
eOwne
rship(1=
Owne
r)0.77
60.87
50.66
70.78
20.71
20.63
40.71
70.81
60.86
40.85
30.61
30.75
8(0.419
)(0.333
)(0.473
)(0.414
)(0.455
)(0.483
)(0.455
)(0.389
)(0.344
)(0.355
)(0.489
)(0.428
)
Hom
eTyp
e(1=
Multi-
dwellin
g)0.05
880.13
60.19
00.24
50.70
20.36
00.67
90.17
80.73
70.23
90.53
80.34
0(0.237
)(0.345
)(0.393
)(0.431
)(0.460
)(0.481
)(0.471
)(0.384
)(0.442
)(0.427
)(0.501
)(0.474
)
Urban
Area(1=Urban
)0.20
00.36
40.75
20.19
00.65
40.34
90.71
70.20
90.50
00.18
30.15
10.36
0(0.402
)(0.484
)(0.433
)(0.394
)(0.478
)(0.478
)(0.455
)(0.408
)(0.502
)(0.388
)(0.360
)(0.480
)
Gen
der(1=
Male)
0.50
60.56
80.58
20.55
10.51
90.58
30.66
00.62
60.60
20.63
80.60
40.58
9(0.503
)(0.498
)(0.495
)(0.499
)(0.502
)(0.495
)(0.478
)(0.485
)(0.492
)(0.482
)(0.491
)(0.492
)
Mem
berof
Envt.Organ
isatio
n0.14
10.14
80.24
20.11
60.17
30.04
570.17
00.16
60.12
70.15
60.27
40.15
5(0.350
)(0.357
)(0.430
)(0.321
)(0.380
)(0.209
)(0.379
)(0.373
)(0.335
)(0.364
)(0.448
)(0.362
)
Energy
EfficientApp
lianc
es0.71
80.69
30.46
40.78
20.69
20.54
90.64
20.73
60.81
40.75
20.76
40.68
9(0.453
)(0.464
)(0.500
)(0.414
)(0.464
)(0.499
)(0.484
)(0.442
)(0.391
)(0.433
)(0.427
)(0.463
)
SpaceHeat/Coo
l(Elect.)
0.77
60.56
80.51
00.49
71
0.85
70.35
80.36
80.66
10.41
70.13
20.55
5(0.419
)(0.498
)(0.502
)(0.502
)(0)
(0.351
)(0.484
)(0.484
)(0.475
)(0.494
)(0.340
)(0.497
)
Table2:
SummaryStatist
ics(M
ean
(s.d.))
ofRegression
SampleforCategorical
Varia
bles.
“Mean”
valueis
interpretedas
theprop
ortio
nof
“1’s”
inthesample.
25
FullSa
mple:
OLS
Exo
geno
uson
lySa
mple:
OLS
weigh
ted
unweigh
ted
weigh
ted
unweigh
ted
Incomeelasticity
0.06
690.09
41**
0.15
1**
0.13
1**
(0.052
8)(0.039
2)(0.061
4)(0.053
1)ho
useho
ldsize
(num
berof
mem
bers)
0.09
49**
*0.10
7***
0.06
83**
*0.09
30**
*(0.028
3)(0.018
0)(0.024
1)(0.020
3)Hom
eOwne
rship(1=Yes)
-0.010
30.05
790.12
10.03
26(0.095
9)(0.058
8)(0.084
9)(0.069
1)Hom
eTyp
e(1=
Multi-dwellin
g)-0.284
***
-0.352
***
-0.220
**-0.386
***
(0.104
)(0.059
5)(0.086
9)(0.076
9)Urban
Area(1=Urban
)-0.233
***
-0.092
1*-0.126
*-0.075
4(0.084
6)(0.051
1)(0.071
7)(0.059
6)Age
ofrespon
dent
0.00
853*
**0.00
923*
**0.00
926*
**0.00
897*
**(0.002
64)
(0.001
81)
(0.002
41)
(0.001
90)
Gen
der(1=
Male)
0.05
95-0.016
30.05
350.00
581
(0.065
2)(0.043
8)(0.056
9)(0.045
9)Years
ofPost-second
aryEdu
cation
0.01
42-0.000
962
-0.005
18-0.012
0(0.016
3)(0.009
55)
(0.012
3)(0.011
1)Mem
berof
Env
t.Organ
ization(1=Yes)
-0.123
0.03
170.05
870.04
98(0.170
)(0.055
9)(0.057
7)(0.057
9)Ene
rgyBeh
aviour
Inde
x-0.028
5-0.039
0***
-0.024
7-0.034
5***
(0.024
9)(0.011
7)(0.016
8)(0.013
1)Ownan
Ene
rgyEfficientApp
lianc
e(1=Yes)
-0.003
25-0.036
10.03
24-0.003
90(0.064
0)(0.042
5)(0.068
2)(0.051
0)Num
berof
App
lianc
es0.01
78**
0.01
50**
*0.00
994*
0.01
53**
*(0.007
40)
(0.004
18)
(0.005
69)
(0.004
72)
SpaceHeat/Coo
lElect.(1=
Yes)
0.22
1***
0.32
6***
0.39
4***
0.39
2***
(0.061
9)(0.044
6)(0.054
9)(0.046
9)Observation
s14
1014
1099
399
3R-squ
ared
0.47
40.48
30.39
50.48
8Mod
elSign
ificanc
eTest
16.94
623.3
14.97
509.2
Mod
elCom
parisonTestStatistic
Rob
uststan
dard
errors
inpa
renthe
ses
***p<
0.01
,**
p<0.05
,*p<
0.1
Table3:
Results
from
doub
lelogSp
ecificatio
nin
eq.(
1).
26
Fullda
taExo
geno
uson
lyTrans-log
Dou
bleLo
gTrans-log
Dou
bleLo
gCou
ntry
Elasticity
S.E
Elasticity
S.E
Elasticity
S.E
Elasticity
S.E
Overall
-0.880
60.44
52-0.887
60.70
65Australia
-1.134
00.94
60-1,433
80,33
08Can
ada
-1.022
90.09
11-1,108
90,18
82-1.019
10.08
90-1,107
90,18
73Chile
-1.023
23.24
49-1,096
60,09
23-0.999
00.36
53-1,092
10,09
68Fran
ce-0.969
30.93
26-0,851
80,14
19-0.853
22.22
10-0,813
10,14
73Israel
-0.851
40.08
39-0,887
60,11
78Ja
pan
-0.806
30.07
25-0,500
20,11
71Korea
-1.039
84.06
94-0,262
30,08
4Nethe
rlan
ds-0.370
81.03
50-0,245
50,08
59-0.789
61.31
91-0,241
90,08
63Sp
ain
-0.996
10.15
18-1,044
10,13
26-0.896
13.03
56-1,030
40,12
91Sw
eden
-1.039
00.09
22-0,689
30,17
37-1.032
80.08
71-0,678
80,17
01Sw
itzerlan
d-0.632
10.23
35-0,536
30,19
17-0.508
50.34
13-0,480
10,20
61R-squ
ared
0.45
0.45
48N
1410
993
Table4:
Cou
ntry-spe
cific
priceelastic
ities,for
thedo
uble-lo
g(in
eq.(1))an
dtran
s-log(in
eq.(4))specificatio
ns,
with
outprob
ability
weigh
ts.Fo
rthetran
slogmod
el,t
hecoun
try-specificpriceelastic
ities
areob
tained
byesti-
matingtheelastic
ityat
thecoun
try-specificmean,w
hile
theOverallcoeffi
cientisob
tained
byestim
atingat
the
gran
dmean.
Forde
tails
rega
ding
thedo
uble-lo
gmod
el,s
eetable3.
27
FullSa
mple:
Poo
led
Exo
geno
uson
lySa
mple:
Poo
led
weigh
ted
unweigh
ted
weigh
ted
unweigh
ted
Price
elasticity
-0.990
2*-1.101
7***
-1.246
8**
-1.179
6**
(0.054
2)(0.036
8)(0.051
6)(0.051
7)Incomeelasticity
0.07
150.10
4**
0.14
5**
0.14
4***
(0.073
4)(0.042
2)(0.060
2)(0.054
0)ho
useho
ldsize
(num
berof
mem
bers)
0.10
2***
0.10
6***
0.06
90**
*0.09
79**
*(0.030
2)(0.019
6)(0.023
4)(0.022
5)Hom
eOwne
rship(1=Yes)
-0.007
380.05
140.06
930.00
114
(0.110
)(0.065
7)(0.076
6)(0.069
1)Hom
eTyp
e(1=
Multi-dwellin
g)-0.227
**-0.324
***
-0.226
***
-0.378
***
(0.114
)(0.066
3)(0.086
6)(0.074
9)Urban
Area(1=Urban
)-0.279
***
-0.113
**-0.118
-0.069
1(0.092
2)(0.056
6)(0.071
9)(0.061
8)Age
ofrespon
dent
0.00
851*
**0.00
969*
**0.00
959*
**0.00
918*
**(0.002
79)
(0.001
99)
(0.002
48)
(0.002
14)
Gen
der(1=
Male)
0.07
21-0.013
30.05
050.00
267
(0.070
5)(0.044
5)(0.057
7)(0.046
7)Years
ofPost-second
aryEdu
cation
0.02
430.00
650
-0.000
560
-0.007
75(0.019
1)(0.009
92)
(0.011
6)(0.010
4)Mem
berof
Env
t.Organ
ization(1=Yes)
-0.084
00.00
446
-0.000
561
0.00
653
(0.176
)(0.057
6)(0.062
5)(0.060
0)Ene
rgyBeh
aviour
Inde
x-0.034
3-0.039
2***
-0.029
5*-0.036
9***
(0.025
9)(0.012
6)(0.017
0)(0.013
5)Ownan
Ene
rgyEfficientApp
lianc
e(1=Yes)
-0.114
-0.055
70.03
53-0.009
38(0.083
2)(0.044
5)(0.063
5)(0.056
4)Num
berof
App
lianc
es0.02
29**
0.01
72**
*0.01
31**
0.01
80**
*(0.009
07)
(0.004
17)
(0.006
07)
(0.004
85)
SpaceHeat/Coo
lElect.(1=
Yes)
0.29
2***
0.36
3***
0.41
0***
0.40
2***
(0.067
6)(0.048
8)(0.056
2)(0.050
3)Observation
s14
1014
1099
399
3R-squ
ared
0.41
90.43
70.36
10.44
5Mod
elSign
ificanc
eTest
17.45
411.0
17.02
354.1
Rob
uststan
dard
errors
inpa
renthe
ses
***p<
0.01
,**
p<0.05
,*p<
0.1
Table5:
Results
from
the“p
ooled”
mod
el.
28
Full Data Exogenous Onlyln(Income) 1.563** 1.383**
(0.704) (0.657)ln(Price) -1.795*** -2.026***
(0.566) (0.749)ln(Income)*ln(Price) 0.0581 0.100
(0.0561) (0.0717)ln(Income)2 -0.0667* -0.0520
(0.0343) (0.0333)ln(Price)2 -0.120*** -0.0639
(0.0210) (0.0406)household size 0.116*** 0.104***
(0.0201) (0.0224)Home Size (sq mtr) 0.000461 0.00143***
(0.000570) (0.000468)Home Ownership (1=Yes) 0.0741 0.0463
(0.0620) (0.0711)Home Type(1=Multi-dwelling) -0.337*** -0.378***
(0.0660) (0.0798)Urban Area (1=Urban) -0.0906* -0.0657
(0.0516) (0.0598)Age of HH head 0.00891*** 0.00839***
(0.00183) (0.00213)Gender(1=Male) -0.0140 0.00677
(0.0449) (0.0509)Years of Post-secondary Education 0.00112 -0.0109
(0.00955) (0.0109)Member of Envt. Organization (1=Yes) 0.0165 0.0383
(0.0539) (0.0560)Energy Behaviour Index -0.0394*** -0.0361***
(0.0123) (0.0135)Own an Energy Efficient Appliance (1=Yes) -0.0661 -0.0321
(0.0435) (0.0575)Number of Appliances 0.0139*** 0.0141***
(0.00404) (0.00469)Space Heating Elect.(1=Yes) 0.344*** 0.400***
(0.0491) (0.0502)Observations 1 410 993
R-squared 0,533 0.622
Table 6: Results from the trans-log Specification in eq. (4).29