Free riding in tax credits for home insulation in France: an econometric assessment using panel data.

Transcription

1 Free riding in tax credits for home insulation in France: an econometric assessment using panel data. Abstract This econometric study assesses the efficiency of the tax credit implemented in France in 2005 on the extensive margin of households retrofitting investments, focusing on insulation measures. A logit model with random individual effects is estimated using an unbalanced panel of households surveyed over After pre trend tests, a difference estimator is performed to identify the policy. The tax credit has had no significant effect during the first two years, suggesting a latency period related to uncertainty and inertia in households investment decision, partly due to the complexity of the tax credit scheme. The tax credit has had an increasing significant positive effect from 2007 to 2010, before slightly decreasing in This is in line with evolutions in the tax credit rates, suggesting a correlation with the level of subsidization. Defined as the situation in which the subsidized household would have invested even in the absence of the subsidy, free ridership has progressively decreased and has been lower for opaque surfaces insulation (roof, walls, etc.) measures than for windows. Estimated shares of free riders range between 40% and 70% after Finally, we assess the potential bias caused by time varying unobservable variables and conclude that our estimates are conservative. Keywords: Energy conservation, residential sector, thermal insulation, tax credit, free ridership, difference estimation, panel data, France Acknowledgement: Results and interpretations from the exploitation of ADEME TNS SOFRES «Maitrise de l énergie» survey, are those of the author and do not reflect the opinion of ADEME and TNS SOFRES. Les résultats et interprétations issus de l'exploitation de l'enquête ADEME TNS SOFRES "Maitrise de l'énergie" sont ceux de l'auteur(e) et ne reflètent en aucun cas l'opinion de l'ademe et de la TNS SOFRES. 1

2 1. Introduction and Motivation In the current context of climate change and given the weight of the residential sector in industrialized countries energy end use and CO2 emissions, the promotion of energy efficiency investments in the existing building stock is a major issue in climate policy. This sector is all the more targeted as it is considered as one of the sectors having the best energy savings potential (Levine et al. 2007), although this potential is discussed (Allcott & Greenstone 2012). Consequently, in a lot of countries, incentives like tax credits have been recently implemented to encourage households to invest in energy efficient retrofitting in their dwellings. The first wave of tax credit implementation occurred in some countries in the 1970 s, in the post oil crisis period, and led to a first group of empirical studies, mainly dealing with the US tax credit from 1977 to Among all barriers faced by households to retrofitting their dwellings, reviewed notably by Jakob (Jakob 2007), some market imperfections provide justification for such instruments. Called investment inefficiencies by Allcott and Greenstone, these market failures mainly refer to imperfect information and may cause households not to undertake privately profitable investments in energy efficiency. When they cannot be addressed directly through mechanisms such as information provision, energy efficiency subsidies and standards may be merited (Allcott & Greenstone 2012). Besides, in case of negative externalities related to fossil fuels and political infeasibility to implement first best solutions such as pigouvian taxes, energy efficiency policies might be a reasonable second best substitute for Pigouvian taxes, although at a higher social cost (Allcott & Greenstone 2012). These policies can also address behavioral failures even if they have major shortcomings when consumers are heterogeneous (Gillingham & Palmer 2013). Therefore, even if such instruments are largely implemented, concerns arise about their real efficiency. Ideally in public economics, we should assess such policy with regard to social welfare, in the framework of a cost benefit analysis taking into account all the significant direct and indirect effects of the policy on the economy and society. In practical, empirical papers rather adopt a cost effectiveness perspective due to the limitations of empirical cost benefit analysis, which is as difficult as it is ambitious 2

3 (Ientile & Mairesse 2009). In order to do so, we first have to assess to what extent households respond to the policy. Consensus does not exist in the literature, as regards studies on the first US tax credit scheme for example. Results differ depending on data and methodology. Dubin and Henson (Dubin & Henson 1988) studied it on the 1979 tax year, using fiscal data aggregated by Internal Revenue Service (IRS) district and audit class. They assessed the tax credit s effects on both the probability to declare conservation investment and the amount of expenditures. They found no significant incentive effects (Dubin & Henson 1988), as well as Walsh (Walsh 1989) using micro data from the 1982 Residential Energy Consumption Survey, and Cameron (Cameron 1985) using a nested logit model. On the other hand, Hassett and Metcalf (Hassett & Metcalf 1995) found a positive significant incentive effect of the tax credit on the likelihood of performing energy efficiency improvements, using micro panel data covering 1979/1981 and an alternative definition of the tax credit variable. Furthermore, there is concern about free ridership. To our knowledge, Joskow and Marron were the first in 1992 to discuss free ridership in this context by conducting a meta analysis of surveying evaluations of demand side management (DSM) programs conducted by U.S. utilities. Free riders are program participants for whom the utility incurs administrative costs and provides subsidies but who would have invested in efficiency even in the absence of a utility program. (Joskow & Marron 1992). Free riding 1 can be depicted in several ways (Cohen & al. 2012) but the related literature has a common acceptance of its meaning. We therefore adopt a definition in line with it: free riding behavior occurs when the agents targeted by the policy take the incentives but would have done the investment anyway (Alberini & al. 2013). Free ridership has to be taken into account to avoid overstating the cost effectiveness of a program. Moreover, the social cost of a policy arises if free ridership is high due to opportunity costs (in addition to administrative costs): public revenues could have been allocated to other expenditures and subsidies may cause price distortions and bear anti redistributive effects. Alberini and al. (& 2013) give three possible causes of free riding, factors leading to self selection of people already willing to invest into the program: i) the energy efficiency characteristics of the renovation are not separable from other technical or aesthetic features that would have motivated the renovation anyway, ii) the agents were already convinced 1 Free riding might not be the most appropriate term. We could prefer windfall gains in order to do not stress on households intention to free ride but we will use free riding as this expression is the one commonly used in the related literature. 3

4 that the resulting efficiency improvement was worth its cost, or iii) the agents replace existing equipment only when it breaks beyond repair. Recent literature is focusing on free ridership assessment. Using a discrete choice model on German cross section data, Grösche and Vance (Grösche & Vance 2009) designate potential free riders as those whose estimated marginal willingness to pay for a particular retrofit option is higher than the observed investment cost, and find a free rider share approaching 50%. On the same database, Grösche, Schmidt, and Vance (Grösche & al. 2013) use a more flexible discrete choice model to generate predicted choice probabilities for each retrofitting option. They find that, as the size of the subsidy increases, the share of program funds allocated to free riders decreases even as the overall cost of the program increases. However, as incentives increase, households switch to retrofit options which are more expensive but less cost effective in terms of energy savings. Alberini & al. (Alberini & al. 2013) studied the effects of an Italian tax credit implemented in Taking into account geographical heterogeneity, they found that the tax credit has led to an increase in the windows replacement rate of 37 to 40 percent in the colder regions only. They also found that free riding seems to be more important in case of heating system replacement than window replacement. A possible explanation is that heating system replacement only happens when the old equipment breaks beyond repair. In this paper, we assess the French tax credit, called CIDD 2, implemented in 2005 in order to trigger households investment in energy conservation and renewable energy equipment in their dwellings. Mauroux (Mauroux 2012) used tax declarations data to analyze the effect of the 2006 reform increasing the CIDD rate from 25% to 40% rate in case of investments in old buildings after a recent housing transfer on households intention to apply for the subsidy over the period 2006/2008. Her results suggest the presence of an important free riding in case of a marginal increase in the tax credit rate. However, she could not observe non subsidized retrofitting investments and therefore investigate CIDD s effect at the time of its introduction. The originality of the present paper is that it first assesses the impact of CIDD s introduction on energy savings investments on the extensive margin at the time of its introduction in The extensive margin corresponds to the number of retrofitting investments that were actually triggered by the CIDD implementation, in other words the CIDD s effect on the probability to retrofit, whereas the intensive 2 Crédit d Impôt Développement Durable in French, for Sustainable Development Tax Credit. 4

5 margin measures the adjustment in households energy conservation expenditures after the implementation of the CIDD. Second, the paper assesses how CIDD s effect evolves through time over 2005/2011. Third, those estimations enable us to assess the share of free riders. We use an unbalanced panel of individual data coming from the Energy Management (EM) annual survey conducted over 2002/2011 and dedicated to energy related households behaviors and investments in their dwellings. In terms of econometric method, we use a quasi natural experiment relying on time and policy design change. In the absence of control group, we use a difference estimator, as in (Alberini & al. 2013). This method relies on strong assumptions but we take care to avoid ascribing to CIDD effects that are due to exogenous unobserved time varying variables. We conduct test for pre trends over 2002/2004 and control for factors likely to affect energy efficiency renovations. We include individual effects to control for unobserved timeconstant variables and discuss the effect of path dependency in households retrofitting investment decision. Action based estimates are more reliable than surveyed declarations (Malm 1996). As the survey also provides questions on CIDD s effects on behavior, the paper will compare both estimates. This study is focused on insulation measures, which include opaque surfaces insulation (walls, indoor or outdoor insulation, roofs, ceilings, floors) and glazed surfaces insulation (windows, doors, shutters). Information on other retrofitting investments, such as efficient heating systems and equipment producing renewable energy, is too scanty to introduce them in the analysis. Moreover, insulation measures less suffer from the third self selection problem described above (replacement after break). Our results suggest a significant and positive effect of CIDD on the investment decision on the extensive margin but with a period of latency during the first two years. Then, estimated CIDD s average marginal effects progressively increase, especially from 2009, before slightly decreasing at the end of the period. These evolutions are in line with the evolutions in the tax credit rates and related to other mechanisms such as uncertainty and inertia in the investment decision. As regards free riding assessment, the estimated shares of free riders among CIDD beneficiaries are between 40% and 70%, in the range of values present in the literature, suggesting that free riding is an important phenomenon. Free riding is found to be lower for opaque surfaces insulation measures than for glazed surfaces ones and progressively decreasing over the period. Finally, the discussion concludes that our estimates would be understated at 5

6 the end of the period due to limitations intrinsic to the difference estimation. We first review the French context and describe the studied tax credit scheme in section 2. Section 3 describes the data and the variables selected in the model. Econometric methods are explained in section 4. Section 5 presents first descriptive statistics and then econometric results. Those are discussed in section 6, before concluding in section Description of the French context and tax credit scheme 3. Inside a EU legislative framework 4, France is committed to reduce its greenhouse gas emissions by 75% by 2050 compared to 1990 level, and to improve final energy intensity by 2% a year from 2015 onwards (French Climate Plan and Energy Program Act of 13 July 2005 establishing France's energy policy priorities). The residential sector consumed 30% of the total French energy supply in 2011 (in final energy) 5, essentially for heating and hot water purposes. The French policy package called Grenelle de l environnement (voted in 2009) aims at cutting energy consumption in the existing building stock by at least 38% by 2020 compared to their 2008 level and fixed targets of retrofitted dwellings per year from 2013 and of worst energy efficient dwellings in the social housing stock by The French tax credit scheme CIDD started in Tax credit was initially implemented between 2005 and 2009 and then extended until It could possibly run later on. The purchase of energy efficient durables for the main home 7 is eligible for income tax credits, with rates ranging from 15 to 50% of investment cost. Eligible investments include both energy conservation measures and renewable energy systems. As for conservation, it applies to insulation, for both opaque surfaces and windows/shutter, and to 3 This review of the French tax credit scheme is based on information coming from the several Official Tax Bulletin publications (BO n 147 on September 2005, BO n 183 on May 2006, BO n 88 on July 2007, BO n 38 on April 2009, BO n 65 on June 2009, BO n 77 on August 2010, BO n 84 on December 2011) and public reports or publications (Mauroux et al. 2010) (Pelletier 2011) (Mauroux 2012). 4 the Directive 2012/27/EU establishes a common framework of measures for the promotion of energy efficiency and the Directive 2010/31/EU focuses on the energy performance of buildings. 5 Source: Le bilan énergétique de la France en durable.gouv.fr/img/pdf/lps130.pdf 6 Earlier legislation allowing fiscal deductions already existed (since 2001 for thermal insulation material) but did not specifically target energy efficiency renovations and its scope was not comparable with CIDD s one : lower public expenses around ten times lower ( lower level of subsidies and lower recourse rate, especially for insulation measures (given households declarations in the EM survey). 7 Tax credit had only subsidized owner occupiers and tenants but was extended to landlords renting their dwelling in

7 heating system improvements, such as heating regulation systems (mainly including thermostatic valves and programming equipment) and performing systems (low temperature and condensation boilers). Renewable energy production refers to wood heating appliances, photovoltaic panels, solar heaters and domestic wind turbines. Tax credit subvention is capped at for a 1 person dwelling, for a 2 persons dwelling (with surplus per child) for a 5 consecutive years period. Renewable energy production systems (including heat pumps) are eligible for all types of building, whereas insulation measures are only eligible for building older than 2 years. Tax credit rates are specific to each retrofitting types and are based on energy performance criteria. Erreur! Source du renvoi introuvable. gives details for tax credit rate evolutions. Inside each category, rates have evolved through time and can be specific to certain households situations. Evolutions in the tax credit rates, as well as energy performance eligibility criteria, have resulted from a compromise between the aim of targeting the most energy efficient systems, the will of limiting public expenses and the lobbying from the supply side. Heat pumps are a good example: air air heat pumps were only eligible between 2006 and 2008 whereas thermodynamic heat pumps for water heating started to be eligible in As regards specifically opaque surfaces specifically, whereas the tax credit base had only subsidized material cost since 2005, it included labor cost (installation expenditures) in the tax credit base in Finally, an overall tax credit cut of 10% (called rabot in French) occurred in 2011, due to the economic crisis and concerns about public deficit. These recurrent modifications could have increased households uncertainty. The CIDD originality in the French context resides in its target and scope. In the current context of global warming awareness, it was the first specific economic incentive encouraging all French households to conserve energy and produce renewable energy in dwellings. Moreover, CIDD has been widely used in France: between 2005 and 2008 about one primary residence out of sixteen was renovated while benefiting from CIDD, yielding 4.2 million of households (Mauroux & al. 2010). Results from the Energy Management (EM) 9 survey showed that the French households have used the CIDD far more than other contemporary economic incentives. Since 2005, more than half of the households investing each year in retrofitting have 8 Labor cost has always been excluded for other retrofitting types. 9 The Energy Management survey provides the data for this study and will be described in section

8 then used the CIDD, this proportion having reached nearly 70% since 2007, whereas less than 7 % of them used other subsidies in the meantime 10. Moreover, according to the same survey, CIDD is widely known, 56.9% of households surveyed only once were aware of it in 2005, steadily increasing to 85.2% in 2009, and has been considered by households as the most decisive incentive since 2006 (see Figure 1). Due to its success, CIDD has led to large public expenses: 985 million in 2005, 1.9 billion in 2006, 2.2 billion in 2007, 2.8 billion in 2008, 2.6 billion in 2009, 1.96 billion in 2010 and 1.1 billion in After CIDD s implementation, other economic instruments were created. A zero rate loan started in 2009 but has not been used as much as CIDD. A total of EcoPTZ were hence emitted in 2009 (resp in 2010 and in 2011), while by contrast, more than 1 million households have benefited from CIDD every year since Moreover, statistics based on EM survey show that EcoPTZ has never been perceived as a key incentive, contrary to CIDD (see Figure 1). CIDD and EcoPTZ could have been cumulated for the same investment except in We will discuss how EcoPTZ could have interacted with CIDD in section 5.1. White certificates for residential energy savings measures have also existed in France since 2006 but this scheme is nearly unknown among consumers 13. [Insert Table 1] [Insert Figure 1] 3. Data 3.1. Dataset description The data used in this paper come from the annual Energy Management survey (EM survey) supervised by the French Agency for Environment and Energy Management (ADEME) and conducted by the 10 Available subsidies gather regional or local subsidies, national subsidies delivered by the ANAH (National Agency of habitat), for which modalities can vary with income, retrofitting performance, etc. 11 Data from the Public Finances general Directorate (DGFiP). Public expenses (in million ) for opaque and glazed insulation were estimated at 518 in 2005, 741 in 2006, 948 in 2007, 811 in 2008, 902 in 2009, 417 in 2010 and 363 in 2011 mainly for windows (CGDD 2012). 12 This has been mainly due to the reluctance of the bank sector to propose such loans (Pelletier 2011). 13 Households are usually not informed by the supply side, at least directly, of the subsidization coming from the white certificates. Energy suppliers may transform them in their own subsidizing programs and/or installers, who can receive some part of the white certificates subsidy, is not obliged to inform the household. 8

9 French market research institute TNS Sofres. It provides detailed information on the retrofitting decision process, the retrofit options, the households and dwellings characteristics, and on the subsidies they received. We used data collected from 2002 to This survey is based on a renewed panel of households, the number of annual observations going from 6148 households in 2005, to 8498 in The turnover is due to attrition due to households will to leave the panel and recruitment made by TNS Sofres to ensure the representation of all the socio economic profiles. Over 2002/2011, households were interrogated with an average turn over rate between two successive years was of 49% and 40.5% of the sample was observed only once. In the analysis, we will introduce individual effects but we will keep households observed only once in order to avoid dropping too many observations and to limit the attrition bias in the data 15. Every year, households are asked about their residential energy consumption and the investments they have or not made, in order to improve the energy efficiency of their dwelling. A first questionnaire provides socio economic variables, housing information (type of building, heating energy source, building date, etc.), and information about dweller's situation (occupation status, move in date). Those who have invested in retrofitting during the last year (7 12% each year) answer a second questionnaire to provide information on retrofitting types, investment costs, some payment modalities, the economic or non economic incentives investors have benefited from (including tax credit), as well as other qualitative information such as their motivation, personal context, satisfaction, etc. In this second questionnaire, each investment is described by 1 to 4 items taken from a retrofitting options list. Retrofitting options include insulation (external insulation of wall, internal insulation of wall, roof, attic, ceiling, windows, shutters), heating system improvement (thermostatic valves, heat cost allocators, ambient thermostat, programming equipment), new heating system (radiator, boiler, wood stove, heatpump, solar heater) or heating system replacement (with information on fuel switching) Variables selection In order to estimate CIDD s effect on the investment decision on the extensive margin, the 14 Survey are conducted over 2000/2012 but we restricted the sample over 2002/2011 due to the time required to clean the data and the fact that some explanatory variables such as the Individual preferences are unavailable for 2000 and Compared to the national scale, the couple of elderly and inactive homeowners are over represented in the EM survey, all the more if we consider people who are present in the survey several years. 9

10 endogenous variable is the retrofitting investment decision, equal to one if the respondent has invested in retrofitting during the past year. We have restricted our analysis to the retrofitting measures that were: i) homogeneous enough regarding energy performance and CIDD s eligibility, and ii) frequently found each year and over the period. This leads us to retain only the opaque (roofs, indoor walls, outdoor walls, ceilings, floors) and glazed (windows, doors, shutters) surfaces insulation types. Indeed, data are inappropriate to deal with other retrofitting measures mainly concerning systems (heating, ventilation, renewable energy production, etc.). As for the category heating systems installation or replacement, it is too aggregated to be studied with respect to CIDD s impact. We do not observe the energy performance of each system, on which is based CIDD eligibility 16. As opposed to heating systems, we assume that available information is sufficient to consider insulation measures without leading to major measurement bias since the non eligibility of the material based on technical criteria is a minor reason for no subsidization, the major reason being the auto production 17. Even if hiring a declared professional retrofitter is a necessary condition to benefit from the tax credit, we chose to keep considering undeclared works (auto produced works or realized by an undeclared professional) in the sample of retrofitting observations. Indeed, we assume that households decision to make the retrofit done by a declared worker, an undeclared one or by themselves is endogenous to the CIDD. Equipment producing renewable energy is also excluded from the analysis since the EM survey did not explicitly mention these systems in the retrofitting options listing before CIDD s introduction 18. As for heating regulation and ventilation systems, the sample size data cannot provide robust statistics given their low retrofitting rates and the limited sample size. The explanatory variables were selected based on the abundant literature on households investment modeling in residential energy, which provides guidance on the main drivers and barriers to consider (Jakob 2007). The basics of those models consists in calculating the return on retrofitting investment by comparing its initial cost with its future economic savings in a cost benefice analysis, in which technological, socio economic and contextual constraints can interact. 16 Moreover, in case of shift in the heating energy source, the old energy heating source is not reported. 17 Given households declarations reported in the EM survey, households who retrofitted to insulate and did not benefit from CIDD declared that is was due to auto production at 75% and due to non eligibility at 26% (among others secondary reasons ) whereas those who retrofitted to replace their boiler and did not benefit from CIDD declared that is was due to autoproduction at 18% and due to non eligibility at 60%. 18 Except for wood stoves which have been observed since 2000 but the annual installation rates are too low to provide robust statistics given our sample size. 10

11 To describe the household s situation in their dwelling, we used the status of occupation (rental or ownership), which is a key variable to characterize the important barrier linked to the split incentives between renters and owners. In the econometrics model, we will restrict the analysis to homeowners. Indeed, the question in the EM survey about retrofitting investment is ambiguous is case of tenants since it is not clear what a tenant of a dwelling in which retrofitting has been undertaken by the owners has to answer, which can lead to potential measurement bias. Moreover, even if tenants were potentially concerned by the program, preliminary statistics show that CIDD might have only impacted home owners. The move in date was also included to represent dwelling situation, as a recent change in occupancy can be a proper timing to retrofit (Gans 2012). The socio demographic variables influencing the investment decision in the model were the Annual income of the dwelling, the Socio professional category, the Family size and the Age of the head of the household. The Annual income of the dwelling determines the households financial possibilities and their opportunity cost of time 19. This variable can also reflect the households discount rate, included in each profitability calculation. Indeed, several studies showed that the discount rate, in other words the preference for the present, decreases with income (Train 1985). Besides, this variable can reflect the impacts of the overall economic variations on individual situations, notably the potential consequences of the crisis occurring at the end of the period. Given life cycle theory, the Age of the head of the household or the Family size may reflect the financial and situational constraints of the dwelling. Distinguishing between Entrepreneurs (in a wide meaning including famers and retailers), Managers (including Liberals professions), Employees and Inactives, the Socio professional category captures aspects linked to education and the opportunity cost of time. In order to capture evolutions in individual preferences about the environment and the economic context, possibly linked to either macroeconomics or social evolutions, we also included data on households main concerns. In the EM survey, households are hence asked every year to rank by order of importance their concerns about environmental (e.g. pollution, climate change, renewable energy ) and economical contexts (unemployment). Environmental concern and Economic concern were included as explanatory dummy variables, which equaled one if the household had claimed that pollution, 19 Since the information collection and the implementation phases of a retrofitting project are time consuming. 11

12 and, respectively, unemployment, were one of their main concerns. The Building completion date, the Building type, the Dwelling size (surface area in squared meter) and the Heating energy sources were the home characteristics variables included in the model to describe the energy performance and the importance of energy consumption, conditioning the profitability of the investment. The Building type variable, which differentiates between individual houses and collective flats, also characterizes the potential barriers raised by a collective decision process. The Heating degree days (HDD) and the Category of city were used to represent the climatic and spatial characteristics of the dwelling. The regional HDD variable, taken from external data source 20, influences the energy performance of a retrofitting investment, as the energy needs vary according to the outside temperature. The Category of city, allows for the differentiation between urban and rural regions, and captures aspects such as storage space availability or supply side structure of the residential energy efficiency market. Finally, the probability to invest in retrofitting is impacted by former retrofitting investments. We deal with dynamic aspects in the retrofitting investment decision including the Former retrofitting variable. From 2004, households have been asked about their former retrofitting which allows us to know if households did not undertake such investment because insulation measures have already been done. This variable is also used in robustness checks as it compels us to drop 2002 and However, since the implementation of dynamics panel data model is impossible due to such short and unbalanced panel, this last estimation allows us to identify any potential bias neglecting this time dependency in former estimations. 20 From statistics made by the French Ministry of Ecology. durable.gouv.fr/energieclimat/r/statistiques regionales.html?tx_ttnews. Heating degree day (HDD) is a measurement based on the gap between outside temperatures and an inside temperature of comfort. The heating requirements for a given structure at a specific location are considered to be directly proportional to the number of HDD. 12

13 4. Econometric strategy The difference estimation. Let ˆ be the difference estimator: ˆ CIDDit 1 CIDDit 0 Yit Yit (1) with Yit the dependent variable, e.g. the retrofitting investment decision, CIDD it Y it 1 and CIDD it 0 Y it the empirical means of Yit after and before CIDD s implementation, respectively. Let CIDD it be a dummy variable equal to one after CIDD s implementation and zero before. ˆ captures the average effect of CIDD s introduction on the dependent variable. It is identified by the marginal effect of CIDD it on Y it and is unbiased if all unobserved explanatory variables are constant over time (Crépon & Jacquemet 2010). All unobserved explanatory variables having evolved over time would be captured by CIDD it. Therefore, in order to avoid ascribing to CIDD effects that are due to exogenous unobserved time varying variables, we first include all the relevant available variables likely to evolve over time. Second, the eventual presence of trends in potential time varying unobservable variables is checked over the period before CIDD s implementation. Third, we will discuss about potential shocks in unobservable variables over the period after CIDD s implementation in order to see in which extent identification could have been biased Econometrics specification. We aimed at determining the amount by which the probability to invest was increased when households benefited from a partial refund of their expenditures. To estimate CIDD s effect on energy saving investments on the extensive margin, we used a Random effect (RE) dichotomous logit model: Xit ui e EI ( it Xit, ui ) PI ( it 1 Xit, ui ) FX ( it ui ) Xit u, i 1 e (2) with PI ( 1 X, u) the probability to invest in retrofitting for the household i at time t and F a logistic it it i cumulative distribution function, X it ( x1 it,..., xkit ) the exogenous observed covariates presented in section 3.2, the vector of coefficients to be estimated and ui the unobserved individual effect. We model 13

14 ui as random individual effect, individual error terms it 2 2 e /3 2 ui Xi ~ Normal(0, u), assuming that i c and x i are independent. The e are assumed to follow a standard logistic distribution with mean 0 and variance giving the latent intra class correlation 2 u logit 2 2 u /3 (Rodriguez & Elo 2003). The model is estimated by maximum likelihood. We first checked the absence of temporal trend by fitting the following model on dataset collected before CIDD s implementation: Tt X ' it ui e EI ( it Tt, Xit, ui ) PI ( it 1 Tt, Xit, ui ) (3) 1 Tt X ' it ui e Then, we fitted the full model over the entire period: CIDDt X ' it ui e E( Iit CIDDt, X it, ui ) P( Iit 1 CIDDt, X it, ui ) (4) 1 CIDDt X ' it ui e, with CIDD t the dummy equal to one after CIDD s implementation and zero before with its coefficient. Contrary to linear model, before/after estimate is not directly derived from. Indeed, marginal effects of a particular xik on PI ( it 1 Xit, ui ) write: PI ( it 1 Xit, ui ) k (1 P( Iit 1 Xit, ui )) P( Iit 1 Xit, ui ). (5) x kit In order to estimate, we computed the average of all individual marginal effects. From those estimates, we derive the free riding share. Recall that free riding is defined as the situation in which the subsidized household would have undertaken the energy saving investment even in the absence of the subsidy. The free riding rate is defined as: ˆ FRS 1, with R the retrofitting rate among the * CIDD R occupying homeowners and CIDD the share of retrofitters applying for CIDD, assuming that the measurement error caused by identifying households benefiting from CIDD with those applying for it can be ignored 21. As for robustness checks, we also estimate models linear probability models. We include individual 21 More accurate information is unavailable. 14

15 effects as both random 22 and fixed. We do not estimate logit model with fixed individual effects because of the incidental parameter problem (Wooldridge 2002). Conditional fixed effects logit model is not used either since it is estimated only on individuals having variation in the outcome, which makes drop too many observations. Finally, based on the same RE logit model as (4), a last model is estimated in order to control for path dependency in the retrofitting investment decision. 5. Results 5.1. Descriptive statistics In order to contextualize the results, Figure 2 shows retrofitting rates in the main dwellings among all homeowners for all the main retrofitting categories. Retrofitting types are here accounted separately, even if the same households have undertaken several measures belonging to different categories. As for the installation or replacement of standard heating systems with non renewable energy, we do not observe any trend. As we have said in section 3.2, this category is too aggregated to study CIDD s impact since we do not observe the energy performance of each system, on which is based CIDD eligibility. However, these statistics suggests that CIDD would only potentially impact the decision to invest at the intensive margin (leading to choose more energy efficient equipment for example), since the retrofitting rate seems unresponsive to CIDD s introduction. Unfortunately, this cannot be properly measured as data do not provide precise energy performance characteristics of each system. No trend appears for heating regulation/ventilation systems either. Post 2005 retrofitting rates seem even lower. We cannot say anything for equipment producing renewable energy as the series start in An exception would be wood stove installation or replacement (observed since 2000), which retrofitting rate seems to slightly increase after But statistical tests cannot be performed due to such low rates and our limited sample size. Finally, focusing on insulation measures, we do not see any particular trend before After 2005, we may see a slight increase until 2008, especially for glazed surfaces insulation, then a peak from 2009 onwards for opaque surfaces 22 In the random effect linear probability model, the individual error terms e it (resp. the individual effect u i ) are assumed to follow a normal distribution with mean 0 and variance (resp. (Rodriguez & Elo 2003). 2 e 2 u ) giving the latent intra class correlation 2 u 2 2 u e 15

16 insulation, from 2009 to 2010 for glazed surfaces insulation 23. This suggests that differentiated effect occurred during two sub periods of CIDD s implementation: 2005/2008 and 2009/2011. [Insert Figure 2] Figure 3 and Figure 4 detail retrofitting rates for opaque and glazed surfaces insulation taking into account interactions between different retrofitting types. Indeed, different retrofitting types can be complementarily part of the same global retrofitting project, all the more considering insulation measures. Figure 3 focuses on cases in which households only insulate one type of surfaces, without any other retrofitting. We find the same evolution as in Figure 2, especially for window and roof insulation measures. Roof insulation was the most common opaque surface insulation measure over the period, and also the retrofitting work that has increased the most after the CIDD implementation, especially after Other opaque surfaces insulation measures series do not show any visible trend, except a slight increasing trend for indoor wall insulation during the CIDD period. Figure 4 considers combined insulation measures. Comparing to Figure 3, Figure 4 shows that retrofitting rates for combined insulation measures were consistently inferior to unitary measures. Their apparent increasing trends after 2005 with a peak in 2009 (except for the combination of window and multi opaque surfaces insulation) are less clear, probably because of the low number of observations. [Insert Figure 3] [Insert Figure 4] In order to provide robust statistical results, we have to choose the appropriate set of measures gathering a sufficient number of observations. Erreur! Source du renvoi introuvable. gives the results of a Pearson's chi squared test statistic testing the equality between the investment rates during the pre ( ) and post ( ) CIDD periods for opaque surfaces insulation and window insulation measures. The equality of the detailed opaque surfaces (except for roofs) investment rates before and after the CIDD implementation could not be tested, due to their small sample size and low power value. We consequently pooled all opaque surface insulation measures in one category to ensure the test good statistical power. The 23 We use window as an extensive meaning for all glazed surfaces. 16

17 equality of the detailed opaque surfaces investment rates (null hypothesis) is then rejected for both opaque surfaces insulation and windows insulation. [Insert Table 2] Finally, it is important to note that the increase of the retrofitting rate after 2005, and especially from 2009 onwards for opaque surfaces, occurs even considering unitary measures exclusively. Indeed, it means that the shock in 2009 cannot be attributed to the introduction of the EcoPTZ since unitary measures are not eligible to this program. Table 3 gives households intentions to benefit from CIDD and EcoPTZ over 2005/2012. We first see that the average intention to benefit from CIDD has been largely more important than the intention to benefit from EcoPTZ. Moreover, as seen in section 3.1, CIDD and EcoPTZ could not have been cumulated in In most cases, EcoPTZ recourse rate dropped in 2011 compared to 2009/2010 level (except for unitary glazed surfaces insulation), which suggests that EcoPTZ alone has had a minor effect on retrofitting decisions. CIDD recourse rate also slightly dropped in 2011 compared to 2009/2011 but this decline has gone on in 2012 and may be due to the fact that households have already benefited from the subsidy 24. [Insert Table 3] Table 4 presents the summary statistics of the households socio economic variables and dwellings characteristics computed on different subsamples to better assess the CIDD effects on energy efficiency investments. These are: i) the full sample dataset, ii) the pre CIDD subsample ( ), iii) the post CIDD subsample ( ), iv) the CIDD aware dwellers subsample, v) the dwellers having invested in opaque or glazed surfaces insulation after 2005 subsample, and vi) the dwellers who invested in opaque surface insulation and intended to benefit from CIDD subsample. Lower income households are under represented among investors and CIDD beneficiaries, as the lowest income bracket represents 33.7% of the post CIDD sub sample, 24.3% of the households having invested in insulation and 21.4% of the CIDD beneficiaries. Older households are over represented among CIDD beneficiaries as they represent 63% of this subsample whereas they represent 53.7% of the post CIDD sample and 55.1% of the households having 24 Households declarations in the EM survey confirm that assumptions (statistics are not reported in the paper). 17

18 invested. In the same way, owners dominates the CIDD beneficiaries as they are 96.8% of this subsample whereas they represent 64% of the subsample and 90.3% the households having invested. Households living in rural areas invested more in energy efficiency than urbans but benefited less from CIDD: they are 24.9%, 34.1% and 29.7% of the post CIDD datasets, the investors and the CIDD beneficiaries, respectively. This might be due to the fact that rurals have more capacities to retrofit by auto production than urbans, auto production being the main cause of CIDD ineligibility. Environmental concern was higher between 2001 and 2004 than during the period (54.8% of the household top priority was environment over 2001/2004, 51.5% over 2005/2011). Conversely, the economic concern increased during the whole period (unemployment was the top priority of 58.2 and 63.3% of the households during the and periods, respectively). Annual national surveys in fact indicates that the environmental concern of French households started to decline after 2008, while employment was becoming the main concern, probably as a side effect of the economic crisis. [Insert Table 4] 5.2 Econometrics We estimate the CIDD effect on energy saving investments on the extensive margin, i.e. on the probability to invest in retrofitting. Recall that we use a difference estimator and that, in order to avoid ascribing to CIDD effects that are due to exogenous unobserved time varying variables, we first perform pre trend tests over the period before CIDD s implementation. Results are presented in Table 5, showing the average marginal effects of both logit models with and without RE for the two retrofitting subsamples (opaque & glazed surfaces and opaque surfaces only). The temporal trend had no significant effect on the probability to invest in logit models fitted on the pre CIDD sub sample, which provides justification for the estimation in difference. The estimated average marginal effects of most of the covariates have the same level of significance and the same magnitude as those from the full models described below. For a few variables (the Socio professional category and the Income), estimates lose their significance compared to those from the full model due to the lack of observations. 18

19 19

20 Table 6 and Table 7 show logit estimates for the full model over for respectively opaque & glazed insulation and opaque insulation only. The tables provide estimates for logit models with (column 2) and without (column 1) random individual effects (RE) plus those of a second RE logit model (column 3) in which retrofitting observations satisfying EcoPTZ eligibility criteria were dropped from the sample in order to isolate CIDD s effects from EcoPTZ s ones. The results reveal significant determinants on the investment decision. As regards households variables, to be wealthy, to live with a family (family size larger than one), to occupy another job than entrepreneurs, retailers of farmers positively impacts the retrofit decision. The age of the household s head variable was dropped due to correlation between covariates. Households having recently moved in also invest significantly more in energy efficiency. Environmental preferences have a significant positive effect on the investment decision whereas economic concerns have a negative effect though significant only when restricting the sample to opaque surfaces insulation. As regards home s variables, households living in old buildings and/or in individual houses are hence more prone to invest than those living in more recent and/or collective flats. Compared to gas, electricity (resp. district heating) as the main heating energy source reduces the probability to invest for opaque & glazed surfaces (resp. for opaque surfaces only), which is probably due to correlation between dwelling types and energy sources. Larger surface area, correlated with energy consumption, increases the probability to invest. Regarding geographic patterns, households living outside Paris, in small cities or in rural areas, invest more in energy efficiency, although with a low level of significance. The regional average heating degree days (HDD) variable, also correlated with energy consumption, positively impacts the retrofit decision but is significant at the 5% confidence level when focusing on opaque surfaces insulation. In most cases, the introduction of RE in the logit models lowers the magnitude of the average marginal effects, leading to loss in significance in rare cases (for the family size variable for example). The intra class latent correlation is lower in the model including all insulation measures than the one focusing on opaque surfaces, indicating a higher correlation between households propensity to retrofit in different years in the sample restricting the retrofitting observations to opaque surfaces insulation. As regards CIDD s effect, we make CIDD it interact with annual dummies in order to allow for temporal heterogeneity in CIDD s effect. Looking at both glazed & opaque surfaces insulation, we find that, 20

21 after a two year latency period ( ) with no significant effect, the CIDD starts to have a significant positive (at 1% level) effect on opaque and glazed surfaces insulation investments in This positive effect increased after 2009 (doubling from 1.6 percentage points in 2008 to 3.2 in 2009 in the RE logit model), with a slight decrease afterwards. Looking at opaque surfaces insulation only, we have the same temporal pattern in the results, with same level of significance (doubling from 0.7 percentage points in 2008 to 1.2 in 2009 in the RE logit model). The introduction of RE reduces the magnitude of the coefficients. In the second RE logit model (column 3), in which potential eligible retrofitting to EcoPTZ were excluded over the entire period, we observe an increase of similar magnitude in 2009 i.e. a doubling from 1.3 (resp. 0.4) percentage points in 2008 to 2.5 (resp. 0.8) in 2009 for opaque & glazed surfaces (resp. opaque surfaces). This confirms your assumptions that this increase in 2009 is mainly driven by the CIDD and not due to the introduction of EcoPTZ. Annual free riding share estimates are presented in Table 8, in addition with annual statistics on R (the retrofitting rate among the occupying homeowners), (the share of retrofitters applying for CIDD) and confidence intervals computed with the delta method. It does not provide free riding share estimates for 2005 and 2006 as CIDD s effect was not significant, suggesting a complete free riding for the first two years. Annual free ridership rates have globally decreased over the period: from 76% in 2007 to 56% in 2009 for all insulation measures, from 62% in 2007 to 42% in 2011 when focusing on opaque surfaces insulation. Estimates indicate that free riding is more important in case of glazed surface insulations than opaque surfaces insulation. We remark that households intentions to benefit from CIDD also decline at the end of the period, probably linked to the exhaustion of the potential of CIDD beneficiaries over time. Therefore, whereas average marginal effects decrease at the end of the period, free riding share does not increase for opaque surfaces insulation. CIDD Finally, Table 9 and Table 10 provide complementary estimation results for robustness checks for opaque & glazed surfaces and opaque surfaces, respectively. The first three columns of each table show estimates for linear probability models. Linear probability (LP) model without individual effect is shown in column 1 whereas LP model with fixed effect FE (resp. random effects RE) is detailed in column 2 (resp. 21