How Large is the Owner-Renter Divide in Energy Efficient Technology? Evidence from an OECD Cross-section

Abstract When the agent making an investment decision is different from the one bearing the costs of the decision, the outcome (energy usage, here) is socially sub-optimal, a scenario known in the energy efficient technology case as “split incentive” effect. Using a sample of households (from a survey conducted in 2011) from 11 OECD countries, this paper investigates the magnitude of the “split incentive” effect between home occupants who are owners and those who are renters. A wide variety of energy-related “technologies” are considered: appliances, energy efficient bulbs, insulation, heat thermostat, solar panels, ground source heat pumps and wind turbines. Mean difference in patterns of access to these technologies are consistent with the “split incentives” hypothesis. Regression results suggest that, even after controlling for the sizeable differences in observed characteristics, owners are substantially more likely to have access to energy efficient appliances and to better insulation as well as to heat thermostats. For relatively immobile investments such as wind turbines and ground source heat pumps, we find no differences between owners and renters.


INTRODUCTION
This paper addresses the question of whether differing incentives between home owners and tenants lead to sub-optimal investment decisions on the part of the owners.Beginning with the premise that more energy efficient investments are expensive, we address the following question: do problems related to imperfect information regarding home energy consumption characteristics (such as efficiency of insulation or of appliances already installed) on the part of the renter lead to under-investment in energy efficiency on the part of the home owner (who foresees, or has experienced, the lack of a premium for such characteristics)?This issue, labelled the "split incentive" effect in this context, has long been viewed as being of some importance; it is, in addition, viewed as part-explanation of the wider phenomenon of under-investment in energy efficiency, the so-called "energy efficiency gap" (see Allcott and Greenstone (2012) and references therein).
1.It is important to note that the practical implications from these studies, in terms of either CO 2 emissions or energy savings, are usually relatively small, either since the effects are quite small (Gillingham et al. (2012)) or since the appliances investigated are a small part of the total energy consumption (Davis (2011)).
An early study in the empirical strand is that of Levinson and Niemann (2004), which attempts to quantify the impact of the usage inefficiency for the U.S., using the Department of Energy's RECS (Residential Energy Consumption Survey).This study reports modest usage-related "split incentive" effects, and therefore, low impact on total energy consumption from this effect.Gillingham et al. (2012) is a study which quantifies, using a representative sample from the RASS (Residential Appliance Saturation Survey) for California, both usage-and investment-related agency effects.This study finds only modest evidence for higher heating settings (i.e.usage effect) when tenants do not pay for heating, and a substantial effect on investment, of up to 20% reduction in probability of being insulated.Another important study is Davis (2011), which attempts to quantify the magnitude of the investment-related "split incentive" effect for appliances.This study reports moderate effects for different appliances, of at most 10% reduction in probability of a tenant having access to an energy efficient (energy-star rated) appliance.Finally, Maruejols and Young (2011) assess the significance of both variants of agency problems for the case of apartments (multi-unit dwellings) in Canada, and report only moderate usage effects and a small investment effect. 1 The current study is similar to those of Davis (2011) and Gillingham et al. (2012), in using observed market equilibria to rationalize and quantify the agency effect.However, the analysis here moves beyond these two in two important dimensions: first, geographic and dwelling coverage is broader and second, it addresses virtually all energy consuming technologies, both appliances and heating/cooling.Finally, the somewhat unique nature of the survey allows a disentangling of the issue of a tenant "having access to" a certain energy efficient technology, in that we can identify if the energy efficient technology in question, in tenant-occupied homes, exists as a result of the owners' investment or the tenant's.The main drawback of the current study is the lack of significant sample sizes for energy consumption and the difficulty of dealing with unobserved heterogeneity (an issue whose importance is highlighted in Gillingham et al. (2012) and Allcott and Greenstone (2012)), given that only a single cross-section is available for use.Further, we are unable to identify which appliances are top-rated for energy efficiency and can only provide an overall figure for the presence of at least one top-rated energy efficient appliance.
We find a sizeable investment-related agency effect, quantified in terms of "ownership effects" on the probability of having access to the relevant energy efficiency technology.For appliances (excluding ACs, which are usually not provided by owners), this effect is very large, about 45%, which is at least four times as large as the largest effect in Davis (2011); for energy efficient bulbs it is an even larger 50%, although the practical implications of this are rather unclear, given its extremely portable nature.For technologies such as ground source heat pumps, wind turbines and solar panels, we find either no effect at all or only minimal effects (for solar panels, at 2.5%), which is consistent with the empirical observation that these tend to be very location specific and are likely driven by both increasing returns (at the level of the country) and local regulations/ incentive structures (heat pumps and solar panels).For the heat thermostat, a relatively mobile yet moderately expensive investment, we find lower (yet sizeable) effects, at 9.8%.Finally, for different types of insulation (roof/walls, windows), we find sizeable effects (9.5 and 12.4% respectively), which are well within the ranges for insulation of different types reported in Gillingham et al. (2012).
The rest of the paper is structured as follows: section 2 provides an outline and basic summaries of the survey data used for the analysis; section 3 lays out the empirical framework, discusses the main results and provides evidence for robustness of the results to a variety of plausible alternative explanations for the observed effects; section 4 concludes.

Survey and Data
Data for the analysis were drawn from the OECD's project on Greening Household Behaviour, as part of which a periodic survey on Environmental Policy and Individual Behaviour Change (EPIC), covering a number of countries and areas, is carried out.The second survey was carried out in 2011, and included 11 countries: Australia, Canada, Chile, France, Israel, South Korea, Japan, the Netherlands, Spain, Sweden and Switzerland.We provide a brief description of the survey, and refer to OECD (2013, Annex B) for survey details and to Kristro ¨m (2013) for a brief overview of the energy efficiency attributes analyzed here.
Approximately 1,000 individuals in each country were surveyed in 2011 by the survey provider using an internet-based questionnaire, for a total sample size of 12,200 households.The questionnaire consisted of 90 questions, with a median completion time of about 30 minutes; respondents were provided with a small in-kind gift (worth between $5 and $10) to promote participation.The survey was administered in nine languages (Dutch, English, French, German, Hebrew, Japanese, Korean, Spanish and Swedish), and had a drop out rate of 21%.Sample selection followed a strategy of stratification based on income, age-group, region and gender.The survey collected information regarding household behavior in five distinct areas (in addition to household characteristics and environmental attitudes): residential energy use, waste generation and recycling, food consumption, personal transport, and water consumption.The current study uses data from the energy section, for households which pay for electricity based on use. 2 In order to account for non-random sampling, ex-post probability weights were provided.
The key question-and the basis for our analysis-is the following (Q77, (OECD, 2013, Annex A)): "Has your household installed any of the following items over the past ten years in your current primary residence?".The possible responses are one of the following: "Yes, No, already existing, installation not feasible".The energy efficient technologies the respondents were asked about are: "top-rated" energy efficient appliance, 3 energy efficient bulbs, ground source heat pumps, solar panels, roof/wall insulation (labelled "Thermal Insulation" in table 2), heat thermostats, wind turbines and insulated windows.Unfortunately, the question does not allow us to distinguish which of these appliances (and how many, since many households report owning more than one) were top rated energy efficient. 4As a result, our discussion pertaining to appliances will stand for any of these appliances (excluding ACs).
5. The specific question asked is (Q76, OECD (2013, Annex A)): "How often do you perform the following in your daily life?", with choices: (a) "Turn off lights when leaving a room"; (b) "Cut down on heating/air conditioning to limit energy consumption"; (c) "Only run full loads when using washing machines or dishwashers"; (d) "wash clothes using cold water (e.g.30ЊC)"; (e) "Turn off stand-by mode of appliances"; and (f) "Air dry laundry rather than using clothes dryers".The response choices are: "Never, Occasionally, Often, Always, Don't know/Not applicable".
6.The actual WTP question asked of the respondent was (Q 71): "What is the maximum percentage increase on your annual bill you are willing to pay to use only renewable energy?Please assume that your energy consumption remains constant".The respondent then could choose a number, from 0 to 100, using a slider (with a choice of "Don't know" available separately).See also Kristro ¨m (2013) for an analysis of this data.
7. A large part of the sample, about 60%, did not report data on electricity consumption-either kWh or euro.Including these variables in the main specification therefore results in substantial loss of sample size.Nonetheless, we report agency effects using both derived average price and annual expenditure on electricity as a robustness check in section 3.3.
It is important, at this point, to note that we have finer data on hand than in the analyses in Davis (2011) and Gillingham et al. (2012) (as well as the residential consumption related studies in de T'Serclaes and Jollands (2007)), which only have the response to a question asking if the respondent "had the appliance in question" in their residence.The wording of the EPIC survey, on the other hand, does provide enough information to evaluate whether the appliance in question already existed when the home was bought/rented or was acquired subsequently, an especially important issue for tenants who live in a rented home.This is an important point since the split incentives issue refers to homeowners differentially investing in energy efficiency depending upon whether they live in the home.Therefore, the relevant category of rented and owned homes for comparison are those where owners invested in energy efficiency after they bought the home (thus excluding those owners who inherited efficiency) and those where renters gained access to energy efficiency as a result of renting (thus excluding those renters who invested in energy efficiency).In prior studies, tenants who purchased the appliances/technologies they report having access to are also included; including these tenants is likely to lead to a lower estimate of the effect of ownership (see section 3.2 for a discussion).In our case, a substantial number of tenants report investing in energy efficiency, including making potentially large fixed investments in certain cases (see table A.1).
The EPIC survey provides a large collection of self-reported variables pertaining to distinct environment-related behaviors/attributes.These variables are suited to controlling for differences in "green preferences" between owners and renters, allowing an exploration of whether differences in stated attitudes (or monetary measures) towards environmental attributes between owners and renters can explain some of the investment-related agency effects.We chose the following three measures as being most relevant to our study: self-reported membership in an environmental organization, willingness to pay (WTP) for consuming only green electricity and an index derived from measures taken to save energy, called the "energy behavior index", whose construction is described next.The EPIC survey asks the respondent how often certain specific energy conservation measures are undertaken in the household.The responses to these questions are then scaled to the interval 0-10, with higher scores indicating more "responsible"/"conserving" behavior. 5The respondent was also asked how much more the household was willing to pay to consume only green electricity, with answers restricted to lie between 0 and 100% of current electric bill. 6 In addition, the survey provides self-reported data on income (in euro), electricity consumption 7 (in euro and in kWh), home details (size of home in m 2 , number of residents, apartment or other type of housing, location details) and socioeconomic characteristics of the respondent (age, gender and employment status), some of which are used in the regressions (see section 3.1 for a Notes: Means are based on the relevant sub-sample, with (sub-) sample sizes shown in the table.For binary variables, the "means" are simply proportion of the sample with category 1. "p-values" refer to the p-value of a two-sided, unpaired ttest for mean difference (assuming unequal variances) between the two sub-groups, Owner and Renter.The summary statistics use probability weights to account for survey sampling (including adjustments to test statistics and the mean itself).This implies, in particular, that the means reported above are not simply the raw means of the sample.
full list of the regressors).As regards the energy consuming appliance data, individuals are also asked how many of refrigerators, freezers and clothes dryers they own.

Summary Statistics
Table 1 presents summary statistics of important variables for the analysis.A few features of the data are worth noting: it is evident that owners as a group are sampled more than tenants, a common issue with many energy-related surveys (see also the remarks in Davis (2011) and Maruejols and Young (2011)) and further, that while "green tastes" are identical for both groups (see next), incomes are substantially different.In more detail, tenants are just as likely as owners to be members of an environmental organisation (at about 8%), are just as "responsible" in their attitude towards energy saving (i.e. have very similar "energy saving behavior" scores, at about 7.5) and have almost identical WTP.Taken together, these clearly imply that there is no discernible differences between owners and renters in terms of "green taste".Owners and tenants however differ in Notes: Definition of mean and computation of test statistics are as reported in table 1.Individuals "having access to" the energy efficient technologies, in the rows under "Our definition", correspond to owners who invested in, and tenants who gained access by renting a unit already equipped with, the relevant technology.Under the "Extant definition", owners who gained access to the relevant technology by purchasing a home already equipped with it and tenants who invested in the relevant technology are also counted as "having access to" that technology.
the size (in m 2 ) of residence, and in annual electricity spending (in euro), conforming largely with differences in income.
Owners as a group are also more likely to report "Married or Living together", compared to renters, and are also substantially older.These help explain the slightly larger household sizes on average, at 3.16, for owner-occupied dwellings in comparison to renter-occupied ones, at 2.6.Both factors indicate that owners, as a group, tend to have a very different social profile, one not reffected in the "Employment" category, since they are also much more likely to be pensioners (not reported).In summary, therefore, a picture emerges of renters as being younger, less well off, less likely to be married and/or having children, spending less on electricity and yet, exhibiting identical "green preferences".We consider next mean differences in access to specific energy efficiency/ generating technologies, the effect of interest.In keeping with our discussion in section 2.1, we report mean differences for two measures of "having access to" energy efficient technologies.In the first, labelled "Our definition" in table 2, only owners who invested and tenants who "inherited" the technology by renting a unit with which it was equipped are counted while in the second, labelled "Extant definition", in addition to those in the first, owners who "inherited" the relevant technology by purchasing a home equipped with it and tenants who invested in it, are also counted.
It is evident from table 2 that irrespective of which measure is used, the differences in access to technologies between owners and renters is sizeable.For instance, for the definition we work with, these range from a very high 66% for bulbs to a rather low 0.5% for heat pumps, with a -0.3% difference for wind turbines.There is also an apparent pattern in these differences, larger values (at 50 and 66%) are associated with appliances and bulbs, lower but rather sizeable ones 8. Substantial ownership of heat pumps is restricted to the four colder countries in the sample, France, Canada, Sweden and Switzerland.Full country-specific tables are available upon request.9.The sample sizes for wind turbines are rather small, with only 158 households (115 of whom are home owners) in the sample having access to them; see table A.1.
10.It might be puzzling at a first glance that while more owners than renters have access to wind turbines (see table A.1), the mean difference is -0.3%.The reason is that the means presented in table 1 are weighted by sampling probability (see section 2.1).
(between 14% and 18%) for insulation-related technologies (insulation, windows and thermostats) and very low and negligible (between -0.3 and 0.6%) for difficult-to-port technologies with large investment and low access levels (heat pumps 8 and wind turbines 9,10 ).Solar panels, which are geographically more restricted, exhibit slightly larger differences, at 4%.While the ranking of technologies based on mean difference is different for the second definition used, the overall mean differences, while lower, are still rather sizeable.
To summarize, there are clearly significant differences in mean outcomes between owners and renters across many dimensions, providing a basis for, and impetus to, further investigation.However, it is important to note that these observed differences in access to specific technologies do not control ("condition") for differing attributes-such as education, income levels or agebetween owners and renters.Given the unconditional nature of such comparisons, it is not clear if observed mean differences are a reffection of the underlying owner-renter dichotomy, which is the effect of interest, or are being driven by common, unobserved factors (or combinations of observed factors).In order to address this issue, we turn next to a regression framework.

Empirical Framework
Recall that the main objective of the paper is to quantify the investment-related agency effect i.e. the size of the principal-agent problem in the rental market.Similar to prior analyses (Davis (2011); Gillingham et al. (2012); Maruejols and Young ( 2011)), we use a binary choice framework to quantify the relevant effect for each of the eight energy efficient technologies/appliances.In more detail, we estimate below the following generic model: where F(.) is a generic conditional distribution, usually the normal or the logistic, Y i is an indicator for "having access to" the relevant energy efficient technology/appliance, X i is a set of conditioning variables described next, and I{owner} is an indicator of ownership.The effect of interest is the average marginal effect (AME) of the impact of ownership i.e.
corresponding to the difference in the probability of having access to the energy efficient technology in question between owners and renters.The estimation takes into account the non-random nature of sampling (i.e.uses sampling probabilities) and the resulting likelihood function is known as the "pseudo-likelihood".Standard error for the AME is, in all cases here, computed as a linearized version-instead of the more common delta-method version-accounting for sampling variability 11.It is important to note that, with many indicator variables-as in the case here-reporting the grand average of marginal effects ensures that the (marginal) effects correspond to an "average household", not to those of any particular category (for instance, one where the respondent is male, employed, and owner).
12. It is difficult to disentangle whether space heating or cooling is the primary goal for all countries; for certain countries e.g.Sweden, Switzerland and Canada, heating is most likely the major goal.For certain other countries e.g.Spain, France and Chile, it is likely to be a combination of both heating and cooling.It is important to note that there are far fewer options for space cooling other than using electricity, strengthening the case for controlling for electric heating/cooling as a driver for technology choice.
13. Households are asked in the survey (Q67, OECD (2013, Annex B)) to list their major source of space heating/ cooling, and one of the choices is electricity (the others being gas and other fossil fuels, wood, district heating, ground source heat pump).
of the conditioning variables.In addition, the standard errors are robust to arbitrary correlation within countries (i.e. are clustered at the country-level).
We emphasize that the main benefit of estimation of agency effects for a wide cross-section of countries, using a common data set and methodology, is the resulting comparability of estimates.To our knowledge, this is the first study to undertake such an analysis.We capture country-specific unobserved common factors by using a country-fixed-effect in the estimation, written implicitly as a part of the X i in eq. ( 1).We note that, unlike in the case of the linear model, this is not as restrictive an approach as appears at a first glance, in that despite not explicitly allowing for country-specific coefficients (e.g.α i ), the marginal effects-being a function of the country-fixed-effects-vary by country.We do not allow explicitly for country-specific coefficients in the interests of both interpretation and parsimonious estimation (the country-specific sample sizes vary and, for many countries, are rather small) and report (in tables 3-5) the marginal effect averaged over all countries and covariate values. 11 A positive AME implies that owners are more likely to own the device in question, relative to tenants; in other words, a positive AME implies an agency effect.The magnitude of the AME on the ownership variable provides a quantitative measure of the agency effect for that particular energy consuming technology.Sample sizes vary between technologies due to the varying number of non-responses (see table A.1 for details).For the main specifications, we use probit (and logit) models, and in these models, the marginal effect has the same sign as the coefficient in the regression.
The regression results for our main specification are reported in table 3. The controls, X i in eq. ( 1), used in each regression include: age and gender of the respondent, number of individuals in the household, size (in m 2 ) of the home, an indicator for whether the home is in an apartment building, defined here as part of a multi-unit complex with more than 12 units and, finally, cubic splines for income, and country-fixed effects.We also include for all technologies-except for appliances (see below)-an indicator for whether the primary source of space heating/cooling is electricity.We do not use, in this set of regressions, either the quantity of electricity used or the electricity price (see also footnote 7).
We turn briefly to discussing the rationale for inclusion and the direction of effect for various controls.It is likely, a priori, that certain technologies (e.g.heat pumps) are more difficult to install in apartments, especially for a single apartment.In addition, apartment-dwellers, owners and tenants both, are likely to face other restrictions on installation which either makes it infeasible to use certain technologies (e.g.solar panels) or lead to increases in cost of installation.In addition, in countries where electric space heating is used, it is likely more expensive than alternatives (such as district heating, in Sweden) and we account for these factors by including an indicator for households which use electricity as a primary source of space heating/cooling. 12,13In most countries, 14.To illustrate a practical case, consider the case of energy efficient bulbs, from table A.1: the precise definition of the agency problem yields 5,946 owners who have access, compared to 360 renters.The conventional definition, however, would add the 3,045 renters who invested in bulbs to the 360 renters who did not; since the number of owners who inherited such bulbs is far smaller (at 545) than the number of renters who invested, one anticipates, a priori, that a smaller agency effect should be detected in the latter case than the former.This situation is reversed in the case of the ground source heat pumps.
15.Given that result from the probit and logit specifications are almost indistinguishable for all cases, we focus on the results of the probit specification.We also note that the close agreement between probit and logit is a well known empirical fact.Attempts were made to estimate specifications other than these, such as the skewed logit or heteroscedastic probit, but were unsuccessful (likelihoods did not converge) due to the inclusion of fixed effects and sampling weights.Heteroscedasticity, in particular, is an important issue in this setting (since ignoring heteroscedasticity can lead to not just inefficient but inconsistent parameter estimates) but one which we are unable to deal with here.
the dynamics of space cooling and heating are rather different: for instance, it is likely that owners do not provide air-conditioners, an important space cooling equipment in many countries.Since these are not directly subject to the split incentives issue, we exclude them from the remit of our appliance access decision and consider only households which do not report owning an air conditioner, for the appliance access decision (row I in table 3).
A final issue is related to a priori expectations regarding the magnitude of effects for different technologies.To illustrate the considerations involved, we compare two types of energy efficient technologies, bulbs and solar panels.The former is quite portable but liable to damage and frequent failure (as well as to theft) and has a short return period; the latter is difficult-to-port and has a much longer return period.A priori, therefore, one may anticipate that landlords are less likely to risk investing in energy efficient bulbs; however, this is counteracted by the substantially smaller magnitude of investment.On balance, it is unclear which of these effects dominate, without greater knowledge of a few specifics, such as average (expected) tenure length, building codes etc.This is also seen in the empirical results in Davis (2011), wherein lighting is seen to have a smaller magnitude of the agency effect than many other categories with higher fixed costs and return periods (e.g.dishwashers).

Results
As indicated earlier, our data set allows us to explicitly identify whether tenant access to a given technology is due to their investment or to owner investment.Prior studies (Davis (2011); Gillingham et al. (2012)) are unable to make this distinction, as a result of which, under their measure (described in table 2), a larger share of tenants (and owners) have access to energy efficiency (every row in the table 2 under "Our definition" is smaller than that under "Extant definition").Note however, that the direction of the impact of this definition upon the agency effect depends upon the following factors: the number of home owners who "inherit" energy efficiency and the number of renters who invest in them.If the latter category is larger (smaller) in size than the former, then it is unlikely that the inclusion of these cases in the "having access to" category will lead to an increase (decrease) in the magnitude of the split incentive effect. 14 We turn now to assessing the magnitude of the agency effects in different energy efficient technologies, presented in table 3. 15 Note that all subsequent references to magnitudes of mean measures of access, including mean differences, are to those reported in table 2. Consider first the case of energy efficient appliances: the positive significant coefficient of 0.45 indicates that owners are 45% more likely than renters to have access to top-rated energy efficient appliances.To understand the magnitude of the effect, observe that the raw differences in means are about 50% (table Notes: Average Marginal Effects (t-statistics in parentheses) for the indicator variable on home ownership from probit and logit regressions.Standard error computations take into account the nature of sampling, account for sampling variation in the conditioning variables and are robust to arbitrary correlations within country.The dependent variable in each regression is an indicator variable for "having access to" the energy efficient technology in question; see p. 71 for the full list of regressors.The definition of "having access to" differs between the columns under "Our definition" and "Extant definition" (definitions in table 2).All regressions include country-fixed effects (not reported).For all regressions above, the included covariates (excluding the country-fixed-effects) were jointly significant, based on a Wald test (not reported), indicating that all regression models are well specified.Significance levels: ***pϽ0.01,**pϽ0.05 and * pϽ0.1 2, row I).Thus, controlling for income and other relevant covariates does not substantially alter the magnitude of the ownership effect.The effect is much smaller (about a quarter), at 12%, with the prior definition of "having access to".These differences are slightly larger than the mean differences (row IX, table 2), which are only 10%, the reason being the significantly greater number of renters investing in a top-rated energy efficient appliance in comparison to owners inheriting the same category of appliance (792 versus 357, from table A.1).In summary, we report a sizeable agency effect on energy efficient appliances, which is much larger than the largest effect for appliances reported in Davis (2011, table 2)-the study closest to ours-at 9.5% for dishwashers.
Turning now to the case of energy efficient light bulbs, our a priori anticipation is of a large effect, if the insecurity already alluded to dominates.In accordance with this intuition, we observe a sizeable effect, at about 50%.This is also evident from an inspection of the mean differences between owners and renters, which are substantial-and slightly larger than the estimated effect-at 66%.We also observe that mean difference when renter investment is included (under the "Extant definition"), at about 10%, is a mere sixth of the case when renter investment is excluded; correspondingly, the estimated effect, at 5.4%, is about half of the mean difference and slightly over 10% of that obtained with the more precise definition of "having access to".Indeed, this estimate is very close to the estimate for lighting in Davis (2011), at 4.9% (mean differences are also close, at 39%).Maruejols and Young (2011) also report an agency effect in energy efficient lighting, although they do not quantify the magnitude.To summarize, there is a large agency effect in the case of energy efficient bulbs, although it is not clear how practically relevant the effect is.
We turn next to ground source heat pumps and wind turbines, which are relatively rare (in terms of access) in the population, at 2.8% and 2.2%, respectively, for owners and renters (heat 16.These issues are somewhat more relevant for the colder countries in the sample and the raw data display a clear pattern confirming this.Australia, Israel, Japan and Spain have relatively lower penetration of roof and wall insulation, while Chile, Israel, Japan and Australia have a lower penetration of heat thermostats (tables available upon request).Nonetheless, even in these countries, the penetration of these technologies is non-trivial and generally above 5%, warranting their inclusion in the analysis.17.The country-specific distribution of these technologies is similar to those for roof/wall insulation, with homes in Australia, Chile, Israel and Japan having relatively smaller proportion of households with access to this technology.As before, these proportions (at above 5%) are reasonable enough to warrant inclusion of all countries in the regressions.pumps) and 1.4 and 1.7% (wind turbines); heat pumps are also relatively country-specific (see footnote 8).The mean differences between owners and renters (rows III and VII, table 2) are rather small (at 0.6% and -0.3% respectively) and insignificant for wind turbines.This fact, in combination with the rather low number of households who have access to it, provides a basis for an a priori expectation of moderate or no effect.These expectations are confirmed by the regression results, where for both these technologies, the effects are less than 1% and insignificant (rows III and VII, table 3).
Turning next to solar panels, the agency effects are estimated at about 3% using the more precise definition and roughly half as large, at 1.6%, with the commonly used definition.The magnitude of the agency effect is about three quarters the mean difference (at 4% and 2.5%, rows IV and XII, table 2) for both definitions.It is important to highlight that cost savings with a relatively expensive and difficult-to-port technology such as solar panels are accrued only after a substantial time period.In this context, it is not clear if these smaller-than-anticipated differences are a result of unusually high renter access or to the overall low penetration of this technology in the population, in combination with country-specific factors such as subsidies (it is interesting that this is the technology for which the highest proportion of participants reported having received a government grant/subsidy-figures not reported).
We next consider issues which are specifically related to home insulation: roof/ wall insulation and use of a heat thermostat. 16Roof/wall insulation is evidently more expensive than a thermostat; nevertheless, from the discussion in section 3, we note that it is difficult to make a specific prediction regarding the direction of the effect.From the regression results (rows V and VI, table 3) we find, surprisingly, identical agency effects for both insulation-related technologies, of around 10%, which is about two-thirds the mean difference (of 14% for insulation and 15% for thermostat, rows V and VI, table 2).We observe a curious reversal for thermostats: despite slightly larger mean difference (at 16%, row XIV, table 2) for the conventional definition, the agency effect is only half as large as that using our preferred definition, at 5%.The estimate for insulation, using the conventional definition, is doubled, to 19.8% (row V, table 3), virtually identical to the mean difference (about 19.1%, row XIII, table 2).
Overall, for both roof/wall insulation and thermostats, we find sizeable agency effects whose magnitudes are very similar to those found in Gillingham et al. (2012).They report (table 11) that owners are 20% more likely than renters to live in homes with attic/ceiling insulation and 6% more likely to live in homes with walls insulated.We cannot distinguish between wall and roof insulation, and so it is not clear which of their estimates are directly comparable; nonetheless, our estimates lie roughly between their ceiling and wall insulation estimates.Strictly speaking, their estimate is to be compared to our estimate using the "Extant definition" for insulation; at 19%, our estimate is almost identical to the 20% they report for attic/ceiling insulation.Finally, for window insulation, 17 we find a sizeable agency effect, at 12% (row VIII, table 3), which is two-thirds the mean difference, at 18% (row VIII,table 2).Curiously, the magnitude of this effect is virtually identical for both definitions of "having access to", while the unconditional mean difference is only slightly smaller, at 16% (row XVI,table 2), with the latter definition.

Robustness Checks
From the foregoing discussion (as well as that in section 1), it should be evident that while agency problems are a very likely cause for the effects identified above, two alternative explanations are equally plausible: first, that owners and renters differ in unobserved (to the econometrician) dimensions and second, that they differ in observed characteristics not already included.Addressing the first issue, of unobserved heterogeneity, is more challenging in our case; unlike in Gillingham et al. (2012), we do not have access to more than one observation per household and cannot use a fixed-or random-effect framework.However, unlike in the cases in Gillingham et al. (2012) and Maruejols and Young (2011), who had access to a rather narrow geographic region (California and apartments in Canada, respectively), we have a much broader sample, covering many countries.It is arguably less plausible that, across the range of institutional frameworks in our multi-country context, there are unobserved factors which systematically drive these differences.In addition, these institutional differences (which do no vary across households within a country) are accounted for by the use of country-fixed-effects.
Prominent in the second category is the hypothesis of "green tastes" i.e. owners, who are wealthier in our sample, have a preference for "green" goods.A related issue is the lack of accounting (in the results in section 3.2) for intensity of use, the hypothesis being that owners, with greater usage, benefit more from greener goods which in turn tends to reduce the cost of usage.In order to address these concerns, we provide (in table 4) a check of the hypothesis that the agency effects estimated in section 3.2 are driven primarily by differences in the following observable characteristics: green tastes (proxied by the three measures of environmental "tastes" referred to in section 2.1: membership in an environmental organization, WTP and the energy behaviour index), different data categorizations (in particular differences across living in an apartment and having electric space heating/cooling) and usage (proxied by either average price or annual electric expenditure).As already noted in section 2.2, owners and renters are very similar in environmental "tastes" and so, a priori, we do not anticipate that "green tastes" will provide an explanation for the estimated ownership effect.Recall also that due to substantial non-response, we do not include average price (unlike in Davis (2011)) or electric expenditure variables in the main specification.
Note that simply including an indicator for some of these characteristics referred to above in the specification in eq. ( 1) does not address the issue of robustness.Rather, the indicator for ownership must be interacted with, say, the indicator for membership (alternatively, the regression may be estimated separately for members and non-members); thus, the effect of interest-the coefficient on the ownership variable-varies between members and non-members.For the case of continuous controls (usage, WTP and energy behavior), however, simply adding the variable itself to the regression in eq. ( 1) is sufficient; the hypothesis here is that sizeable variation in the added control variable between owners and renters drives the agency effect estimates in table 3.If the effect of interest is unchanged with the addition of, say, WTP, that particular hypothesis is rejected.
We first note that sample sizes varied widely for each (sub-) category (see table A.2). We also do not provide an interpretation for each possible specification for each energy technology under consideration: we merely note certain interesting points pertinent to our discussion and provide comments regarding the overall patterns.We begin our discussion of the results of robustness checks (from table 4) with the effect of inclusion of the (average) price variable as a proxy for usage, in a specification otherwise identical to the baseline.We note a striking result (row X, table  3. Rows corresponding to "Apartments", "Electric heating"and "Envt.Membership" report results from probit regressions estimated separately for "category" = 1 and "category" = 0, where "category" refers to the relevant row header.Rows correspondingto "Price", "Expenditure", "WTP" and "Energy behavior index" report results from probit regressions with respective variable included in the regression, in addition to the regressors used in the baseline case in table 3. Standard error and test statistics computations as well as other details are as reported in table 3. differences in unobserved preferences-such as "green tastes" (which we are directly able to account for)-, to specific types of homes (many investments are either infeasible or more expensive) and to electrically heated/cooled homes (which possibly face higher costs and so are likely more motivated).It is important to note that for appliances and bulbs in particular, the estimated agency effects are substantially larger than estimated before, by a factor of four to six.We believe that the main reason for the strength of the effect estimated here is an ability to distinguish between tenant investment and owner-provision.Given the sizeable investment by tenants in these devices in the countries in our sample, ignoring this distinction leads to substantial under-estimation of the ownership effect even in our case.
While these results are promising and indicate scope for policy, there are two types of data-related drawbacks which call for some caution in interpreting our results.First, the crosssection nature of the data do not allow for accommodating individual heterogeneity and product replacement cycles and second, while the data set is relatively rich in certain dimensions (compared to even the RECS, say), it is not rich enough to allow more precise delineation of specific issues/ effect (e.g.appliance-specific effects such as in Davis (2011) or non-electric heating costs).Addressing unobserved individual heterogeneity is considered important in the energy efficiency literature, from both estimation (e.g.Gillingham et al. (2012)) and policy (Allcott and Greenstone (2012)) perspectives.Collection of panel data sets which are comparable across countries would help address, in particular, the issue of unobserved individual heterogeneity.
The investment-related split incentives issue is important since it represents a possible loss in social welfare, due to the externalities imposed in electricity generation.In addition, the residential sector is a key part of the solution to the climate change issue, in particular for the kyoto protocol signatories in the sample (Japan, EU countries of Italy, Spain, Sweden, France as well as for Australia).
The direct policy implications of studies which quantify the magnitude of these agency effects are unclear, however, since policy makers are essentially dealing with two market failures (in the terminology of Allcott and Greenstone (2012)); the first due to the externality from energy consumption (e.g.climate change) and second, that related to information asymmetries and failures (exacerbated by policies, such as rent control, targeting other aspects).A common administrative, second-best, solution consists of stringent building codes (for insulation), minimum efficiency standards (for appliances) and information enhancement campaigns in general.These measure are already in place, to varying degrees, in most countries in our sample, and further increases in stringency are unlikely to increase social returns and may even reduce welfare (Allcott and Greenstone (2012)).Additional specific measures to address this market failure involve country-specific costs (administrative and other costs) and in many cases, these measures may not pass a cost-benefit test (OECD (2010, p.18)).
We remark that for the countries in the sample, the market failure identified here implies welfare losses of varying magnitude, depending both upon factors (to be) identified in more detailed country-specific studies and country-specific welfare functions.In light of these considerations, our estimates of aggregate market failure are presented as a useful starting point for further discussion and analysis for developing country-specific measures.

Table 4 : Robustness Checks for Agency Effects
Reported above are Average Marginal Effects (t-statistics in parentheses) for the indicator variable on home ownership from probit regressions.The dependent variable in all regressions is an indicator for "having access to" the technology indicated in the column header."Baseline" results are from Column II of table

Table 5 : Additional Robustness Checks for Agency Effects on Appliances and Bulbs
Average Marginal Effects (t-statistics in parentheses) for the indicator variable on home ownership from probit regressions reported in the table.The dependent variable in all regressions is an indicator for "having access to" the technology indicated in the column header."Baseline" results are from Column II of table 3. Standard error and test statistics computations as well as other details are as reported in table 3.

Table A .1: Ownership and Investment Patterns in Energy Efficient Technology
Tabulation of sample sizes, for each energy efficient technology analyzed, by home ownership status (owner and tenant) and purchase type (purchased and preexisting).For energy efficient appliances, two categorizations are provided: the first one including air conditioners and the second, excluding it.Sample sizes by owners and renters for all energy efficient technologies (except for the appliances excluding ACs) are: Owner 7,695 and tenant 4,506, for a total of 12,201.For appliances excluding ACs: Owner 3,274 and tenant 2,100, for a total of 5,374.The regression sample sizes reported in table 3 vary due to sample-size differences in included covariates.