WP 16-01 
Moral Hazard and the Energy Efficiency Gap: 
Theory and Evidence 
 
Louis-Gaëtan Giraudet 
Ecole des Ponts ParisTech 
Centre International de Recherche sur l’Environnement et le D´eveloppement (CIRED) 
Sébastien Houde 
University of Maryland  
Department of Agricultural and Resource Economics 
Joseph Maher 
University of Maryland  
National Socio-Environmental Synthesis Center (SESYNC) 
 
 
 
 
 
 
 
 
 
 
Copyright © 2016 by Louis-Gaëtan Giraudet, Sébastien Houde, and Joseph Maher 
All rights reserved.  Readers may make verbatim copies of this document for non-commercial purposes by any 
means, provided that this copyright notice appears on all such copies. 
Moral Hazard and the Energy Efficiency Gap:
Theory and Evidence
Louis-Gae¨tan Giraudet,∗ Se´bastien Houde,†Joseph Maher‡
December 10, 2016
Abstract
We investigate how moral hazard problems can cause sub-optimal investment in energy efficiency, a
phenomenon known as the energy efficiency gap. We focus on contexts where both the quality offered by the
energy efficiency provider and the behavior of the energy user are imperfectly observable. We first formalize
under-provision of quality and compare two policy instruments: energy-savings insurance and minimum
quality standards. Both instruments are second-best, for different reasons. Insurance induce over-use of
energy, thereby requiring incomplete coverage in equilibrium. Standards incur enforcement costs. We then
provide empirical evidence of moral hazard in the U.S. home retrofit market. We find that for those measures,
the quality of which is deemed hard to observe, realized energy savings are subject to day-of-the-week effects.
Specifically, energy savings are significantly lower when those measures were installed on a Friday—a day
particularly prone to negative shocks on workers’ productivity—than on any other weekday. The Friday effect
explains 65% of the discrepancy between predicted and realized energy savings, an increasingly documented
manifestation of the energy efficiency gap. We finally parameterize a model of the U.S. market for attic
insulation and find that the deadweight loss from moral hazard is important over a range of specifications.
Minimum quality standards appear more desirable than energy-savings insurance if energy-use externalities
remain unpriced.
JEL: D86,Q41
Keywords: Energy efficiency gap, moral hazard, energy-savings insurance, minimum quality standard, cre-
dence good, day-of-the-week effect.
∗Centre International de Recherche sur l’Environnement et le De´veloppement (CIRED), Ecole des Ponts
ParisTech. Email: giraudet@centre-cired.fr.
†Department of Agricultural and Resource Economics, University of Maryland. Email: shoude@umd.edu.
‡National Socio-Environmental Synthesis Center (SESYNC) University of Maryland. Email:
jmaher@sesync.org.
21. Introduction
Energy efficiency measures are widely advocated as a means of both saving money and cost-
effectively reducing externalities associated with energy use. Yet in practice, they are little
adopted. This phenomenon is commonly referred to as the energy efficiency gap. A vari-
ety of explanations have been investigated to explain the energy efficiency gap (Gillingham,
Newell, and Palmer 2009; Allcott and Greenstone 2012). Jaffe and Stavins (1994), who
first conceptualized the problem, emphasized the difference between market-failure and non-
market failure explanations of the gap. Market failures such as information asymmetries,
positive externalities from innnovation or negative externalities from pollution and security
of energy supply may distort incentives for energy efficiency investment. This motivates
implementation of corrective policies. Non-market failures such as heterogeneity in con-
sumer valuations of energy services or hidden costs (e.g., inconvenience caused by measure
installation) may also prevent widespread adoption of energy efficiency. Yet unlike market
failures, these are normal components of markets. As such, they should be accounted for in
economic assessments but do not per se warrant any particular intervention. More recently,
the dichotomy has been enriched with the concept of behavioral anomalies to account for
apparent under-valuation of energy savings by energy users (Gillingham and Palmer 2014;
Allcott, Mullainathan, and Taubinsky 2014).
Several ex post evaluations of energy efficiency programs also find that realized energy
savings underperform those predicted by engineering models (Metcalf and Hassett 1999;
Fowlie, Greenstone, and Wolfram 2015; Zivin and Novan 2016; Maher 2016; Davis, Fuchs,
and Gertler 2014; Houde and Aldy 2014). This measurement gap between predicted and
realized savings can arise from a variety reasons and is subject to an intense debate.1 In this
paper, we provide a market-failure explanation for the discrepancy between predicted and
realized savings: moral hazard in the provision of quality. Combining theoretical, empirical
and numerical approaches, we argue that the problem is empirically relevant, economically
1Fowlie, Greenstone, and Wolfram (2015) investigate the role of the rebound effect, which appears to not
play a major role in their context. Davis, Fuchs, and Gertler (2014) discuss the role of inframarginal program
participants and the fact that some participants might replace technologies that are not being used in the first
place, leading to higher energy consumption after investment. Houde and Aldy (2014) present evidence that
consumers might be upgrading toward larger and higher quality models when energy efficiency technologies
are subsidized. Other than the reason discussed here, the gap between predicted and realized savings can
be attributed to a variety of technical reasons (de Wilde 2014), in particular the so-called prebound-effect,
according to which energy use before investment tends to be overestimated in engineering predictions of
energy savings (Sunikka-Blank and Galvin 2012).
3important and partially remediable by policy interventions such as minimum quality stan-
dards and energy-savings insurance.
Our motivation comes from the credence-good nature of energy efficiency, a little-studied
aspect of this technology. Just like taxi rides or auto repairs, many energy efficiency measures
are subject to verifiability and liability issues which make their performance never completely
revealed to the buyer (Sorrell 2004; Dulleck and Kerschbamer 2006). This is especially the
case in buildings, where energy use depends on unobservable factors such as weather fore-
casts, occupants’ behaviors and the quality of energy efficiency equipment. These properties
are conducive to a set of information asymmetries, of which two have received most attention.
One is adverse selection in housing decisions. The intuition is that the energy efficiency of a
dwelling is hard to observe and therefore will not be capitalized into sale prices or lease con-
tracts. The intuition is proving correct in rental housing, as rented dwellings are found to be
less energy-efficient than owner-occupied ones (Scott 1997; Davis 2012; Gillingham, Harding,
and Rapson 2012; Myers 2013; Krishnamurthy and Kristro¨m 2015). The effect is less clear
in home sales. More energy-efficient homes, as measured by their energy performance certifi-
cate, are found to sell with a premium, but no counterfactual situation without certificates is
available for comparison (Brounen and Kok 2011; Murphy 2014; Fuerst, McAllister, Nanda,
and Wyatt 2015). Another much-studied information asymmetry associated with energy
efficiency is moral hazard in energy demand. It is well established that building occupants
use more energy when they face zero marginal cost for energy usage, for instance because
they signed up a utility-included rent contract (Levinson and Niemann 2004; Maruejols and
Young 2011; Gillingham, Harding, and Rapson 2012).
The information asymmetry considered here is related but involves different parties. We
are interested in under-provision of quality in the installation of energy efficiency measures.
We specifically examine home energy retrofits, where a contractor may take advantage of
the lack of expertise of the homeowner to perform insulation or duct sealing poorly. This
can be interpreted as supply-side moral hazard. As we shall see, full analysis of the problem
and solutions thereto requires one to also consider the demand-side moral hazard discussed
above.
Our contribution is three-fold. First, we formalize how supply-side moral hazard can cause
an energy efficiency gap and examine two little-discussed policy remedies: energy-savings
insurance, a private solution, and minimum quality standards, a public one. The analy-
sis builds on a double-moral-hazard framework borrowed from the literature on warranties
(Cooper and Ross 1985). We articulate the mechanism by which asymmetric information
4induces the contractor to cut quality in equilibrium. This deters adoption of energy efficiency
measures. Both policy solutions are found to be second-best: energy-savings insurance re-
duce marginal energy expenditures, which induces demand-side moral hazard and therefore
requires incomplete coverage in equilibrium; minimum quality standards incur enforcement
costs.
Second, we provide the first empirical evidence of supply-side moral hazard in home energy
retrofits. Using a dataset of more than 4,000 retrofits sponsored by Gainesville Regional
Utilities (GRU) in Florida, we exploit variation in the type of measures (classified as easy or
hard to observe) and the day of the week on which they were installed. We find that realized
energy savings underperform predicted ones, and in particular that the former can vanish
for hard-to-observe measures if those were installed on a Friday. The result is robust to a
number of robustness checks, including testing for contractors selecting specific measures on
Fridays. It suggests that perhaps due to fatigue, retrofit workers are likely to shrink if the
informational context gives them the opportunity to do so. Controlling for the Friday effect,
the discrepancy between predicted and realized energy savings shrinks by 65%. Incidentally,
the exercise contributes to the empirical literature on day-of-the-week effects on workers’
productivity (Campolieti and Hyatt 2006; Bryson and Forth 2007) and more generally to
the empirical analysis of moral hazard and credence goods (Abbring, Heckman, Chiappori,
and Pinquet 2003; Dulleck, Kerschbamer, and Sutter 2011; Schneider 2012; Balafoutas, Beck,
Kerschbamer, and Sutter 2013).
Third, we integrate the two approaches and conduct a numerical assessment of the size
of and policy solutions to supply-side moral hazard in the context of attic insulation. The
model developed in the theoretical exercise is specified and parameterized with data derived
from the GRU program and the U.S. Residential Energy Consumption Survey (RECS). We
find that the deadweight loss from moral hazard is commensurate with that from energy-use
externalities over a number of parameter specifications, in some cases even largely exceed-
ing it. Energy-savings contracts with insurance coverage of about 20% can mitigate losses,
while minimum performance standards deliver benefits which hardly compensate enforce-
ment costs. The merit order is however reversed if energy-use externalities are left unpriced.
Then, the demand-side moral hazard induced by insurance magnifies losses and standards
appear as the best policy option.
Our analysis is a first step pointing to supply-side moral hazard in home energy retrofits
as an important problem, both empirically and economically. Policy instruments already
5existing in the marketplace could be adjusted to deliver their full potential. Energy per-
formance contracts, which are common in commercial buildings, could be promoted in the
residential sector. Certification of professional installers, which is so far voluntary, could
evolve toward a mandatory regime. Hybrid instruments combining standards and insurance
might also produce substantial benefits.
The outline of the paper is as follows. Section 2 introduces the theoretical model, derives
key predictions and examines policy solutions. Section 3 presents the empirical approach
and the results. Section 4 provides numerical simulations of the U.S. home insulation market
to compare policy impacts. Section 5 concludes.
2. A model of energy efficiency investment with double moral haz-
ard
Our model builds upon the double-moral-hazard model of Cooper and Ross (1985). Invest-
ments in energy retrofits for the residential sector, which involve hidden actions from both
the homeowner and the contractor, are considered as a canonical example. Other situations
that give rise to one-sided moral hazard, for instance energy retrofits in the commercial and
industrial sectors, can be viewed as special cases of this general model. They are occasion-
ally discussed in the text. The exposition here focuses on the key elements and predictions
of the model. Formal assumptions, additional propositions, complete proofs and graphical
illustrations are provided in an online appendix.
2.1. Setup
Consider an ower-occupant who uses energy to heat her home. This energy service s, mea-
sured in indoor temperature, provides her with increasing comfort V (s), multiplied by a
taste parameter θ > 0. For instance, a person who prefers a high indoor temperature would
be characterized with a high value of θ.
The homeowner faces energy expenditure pE0(s), where E0(·) is the energy use and p the
price of energy. She sets her energy service s0θ so as to maximize the discounted sum of net
utility:
(1) U0(θ, s) ≡
(
θV (s)− pE0(s)
)
Γ,
6where Γ is a discount factor.2
The homeowner can undertake a retrofit investment supplied by a contractor. Each party
can take hidden actions that influence ex post energy use, E(s, q). The homeowner’s ex
post energy service s is unobserved to the contractor. In turn, the quality q with which
the contractor completes the retrofit is unobserved to the homeowner, who as a non-expert
cannot verify insulation installation or duct sealing, for instance. Energy use, which is
reported on the homeowner’s utility bill, is common knowledge to both parties, who are also
aware that it increases with s and decreases with q. The framework is deterministic and
linearity of utility with respect to energy expenditures reflects risk-neutrality.3
Upon investing, the homeowner maximizes utility U(θ, s, q), net of upfront cost T > 0
and including an idiosyncratic value, : 4
(2) U(θ, s, q) ≡ (θV (s)− pE(s, q)) Γ− T + 
The contractor maximizes profit formed by the revenue from the sale T minus the cost
C(·) of providing quality q. We assume zero profit, so that: T = C(q). The assumption
is meant to reflect the competitive nature of the industry.5 As we explain in the Online
Appendix, the equilibrium analysis is nevertheless robust to alternative market structures.
The energy efficiency contract can be modeled as a two-stage game, of which the home-
owner is the principal and the contractor is the agent. In the first stage, the homeowner of
type θ invests if her net present value NPV (θ) is positive, given her energy service s∗θ and
the quality q∗θ she expects to be offered in equilibrium:
2We assume time invariance of energy price, energy technology and consumer value. The energy service
vector is therefore constant over time, which we simplify as a scalar s. The discount factor implicitly factors
in a discount rate r and an investment lifetime l such that Γ ≡ Γ(r, l) ≡∑lt=1(1 + r)−t = (1− (1 + r)−l) /r.
3We ignore uncertainties coming from the weather variations determining heating needs, from measure-
ment errors propagated in the complex engineering models used to predict energy savings, or from the
volatility of energy prices. This simplification is equivalent to assuming that the effects of s and q on energy
use both satisfy first-order stochastic dominance.
4The term  captures aesthetic and acoustic benefits associated with the measure, as well as the inconve-
nience costs incurred forinstallation.
5The home energy retrofit industry is highly fragmented. For instance, firms operating in the heating,
ventilation and air conditioning (HVAC) industry in California are typically small, offer low wages, face low
barriers to entry and an annual turnover as high as 25% (Zabin, Lester, and Halpern-Finnerty 2011).
7(3) NPV (θ) ≡ U(θ, s∗θ, q∗θ)− U0(θ, s0θ) ≥ 0
In the second stage, both agents determine their optimal action given their belief about
the other party’s action. The game is solved using backward induction.
2.2. Supply-side moral hazard
The social optimum is modeled as a cooperative game with perfect information. We show
in the Online Appendix that this game generates strictly increasing reaction functions: a
contractor will offer more quality to a homeowner he perceives as demanding more energy
service; the homeowner will demand more energy service is she expects to be offered more
quality. The intersection of the two reaction functions defines a perfect-information equilib-
rium that determines the socially optimal level of quality.
If actions s and q are not perfectly observable, the parties each maximize their private sur-
plus, given their beliefs about the other party’s action. This does not affect the homeowner’s
first-order conditions and therefore leaves her reaction function unchanged. In contrast, the
contractor does not to internalize how the quality he provides benefits the homeowner. His
reaction function is now flat: whatever behavior he expects from the homeowner, he sets
quality at the level which minimizes production cost. The intersection of the two reaction
functions defines the asymmetric-information equilibrium.
Equilibrium actions under perfect information (PI) and asymmetric information (AI)
can be unambiguously compared:
Proposition 1. Under asymmetric information, an energy efficiency contract is subject to
supply-side moral hazard. The contractor offers less quality to any homeowner of type θ than
under perfect information (qAIθ ≤ qPIθ ). The homeowner responds by using less energy service
(s0θ < sAIθ ≤ sPIθ ). The two inputs together make investment less profitable (NPV AI(θ) ≤
NPV PI(θ)).
The resulting quality gap, which corresponds to under-treatment in the credence-good
terminology, would occur even if the homeowner’s behavior did not respond to energy ex-
penditures.6
6This is for instance the case with a tenant subscribing to a utility-included rent contract, or an employee
in a commercial building. Then, the occupant’s reaction function would be flat. The contractor’s reaction
8Comparison of equilibrium outcomes is ambiguous when it comes to energy use. If con-
sumer behavior were held constant at some s¯, as is generally assumed in ex ante engineering
simulations, then moral hazard would still produce a wedge between predicted and realized
energy savings: E(s¯, qAIθ ) > E(s¯, qPIθ ). Yet according to Proposition 1, undoing moral hazard
would increase both q and s. Recall that by assumption, the two inputs have an opposite ef-
fect on E(·, ·). Therefore, the decrease in energy use due to increased quality would be partly
offset by the increase in energy service. This phenomenon is known as the rebound effect.
At some point, it can backfire, that is, be such that energy use is higher after investment.
In the presence of negative energy-use externalities, this can have important consequences,
as we will see later.
We now extend the above result to the whole market. Consider a continuum of home-
owners of mass 1, all living in a similar dwelling and only differing with respect to their
preference for energy service θ. We show in the Online Appendix that the higher the value
of θ, the higher the demand for energy service, hence the higher the quality offered under
perfect information. This shifts the homeowner’s reaction function upward. In contrast,
the quality offered under asymmetric information remains at minimum. As a result, the
moral-hazard effect is increasing in θ. We also show that the net present value of investment
is increasing in θ. This means that there exists a unique marginal investor of cutoff type
θ∗0 such that NPV (θ∗0) = 0. Combining this with Proposition 1 leads us to the following
proposition:
Proposition 2. Asymmetric information creates an energy efficiency gap at the market
level. Both social welfare and the number of investing consumers are lower than under
perfect information: WAI ≤ W PI and NAI ≤ NPI . .7
Anticipation of the quality gap discourages homeowners with low valuations of heating
comfort to invest. As a result, investment is suboptimal on both the intensive and extensive
margins. Again, without further specification of the technology E(·, ·), we cannot conclude
about how aggregate energy use differs in the two equilibria.
function under perfect information would also be flat but set at the maximum quality level, which internalizes
the effect of quality on energy expenditures.
7N∗ ≡ 1 − F (θ∗0) is the equilibrium number of participants, with F (·) the cumulative distribution func-
tion of θ. W ∗ is aggregate welfare, calculated under zero-profit condition as the sum of utility before
investment for non-participants and utility after investment for participants: W ∗ ≡ ∫ θ∗00 U0(θ, s0θ)dF (θ) +∫ +∞
θ∗0
U(θ, s∗θ, q∗θ)dF (θ).
92.3. Policy solutions
The textbook remedy to moral hazard is a risk-sharing contract. In the context of home
energy retrofits, such a contract can take the form of energy-savings insurance. Alternatively,
a regulator may want to address the problem with a verifiable quality standard. A third
option, sometimes found in practice, is to combine the two. In this section, we compare the
two solutions in their purest form in order to identify their relative strengths and weaknesses.
We find that none can achieve the first-best outcome.
Energy-savings contracts and double moral hazard. Energy-savings contracts or in-
surance, more commonly referred to as energy performance contracts, have been offered for
nearly twenty years in the commercial sector (Mills 2003), but less frequently in the residen-
tial sector.8 Such contracts typically have the contractor pay the homeowner any shortfall in
energy savings below a pre-agreed baseline. In our simple framework with no risk-aversion,
insurance can be modeled as a contract specifying a share k of energy expenditures borne
by the contractor in exchange for an actuarially-fair insurance premium I:
(4) I = k pE(s, q)Γ
Such an insurance contract creates an incentive problem which superimposes to the one
it is meant to address in the first place. The contract can be modeled as a three-stage game,
of which the contractor now is the principal and the homeowner is the agent. In the third
stage, the parties cooperatively determine optimal insurance coverage k; in the second stage,
they privately set their own action, given their belief about insurance coverage k and the
other party’s action; in the first stage, they decide whether or not to participate.
The insurance induces the contractor to offer some quality, otherwise he would have
to make excessive payments to the homeowner. In other words, the risk-sharing contract
mitigates supply-side moral hazard. At the limit, it could even eliminate it, as complete
coverage (k = 1) would induce the contractor to offer socially optimal quality. But at the
same time, the contract gives rise to demand-side moral hazard: by lowering the homeowner’s
marginal value of energy service, it induces her to use more energy. At the limit, complete
coverage would drive the homeowner’s marginal energy expenditure to zero, thereby inducing
her to use energy service up to satiation. Complete coverage is therefore not optimal:
8GreenHomes America, Inc., NJ-PA Energy Group, LLC. and EcoWatt Energy, LLC. are the few examples
we have found of companies offering energy-savings insurance in the U.S. residential sector.
10
Proposition 3. Energy-savings contracts create demand-side moral hazard. As a conse-
quence, optimal insurance coverage is incomplete (0 < k∗θ < 1).
Note that if the building occupant were not adjusting her energy service (e.g., a tenant
subscribing to a utility-included rent contract, or an employee in a commercial building),
then the second moral hazard would not occur. The optimal contract would stipulate com-
plete coverage and bring the parties to the social optimum. This may explain why energy
performance contracts are common in the commercial sector but scarce in the residential
one.
Note also that by increasing both q and s, the energy-savings contract generates a rebound
effect. Unlike that induced by the simple energy efficiency contract discussed in Section 2.2,
this rebound effect can be interpreted as moral hazard.9
In practice, homeowner types, θ, cannot be observed. A uniform contract with coverage k¯
is therefore most likely to be offered to all homeowners. Such a contract generates additional
deadweight losses, as the coverage might be optimal to one homeowner, but is suboptimal
to all others.
Minimum quality standard. A number of voluntary certifications exist in the market-
place, most notably those provided by the Building Performance Institute (BPI) and the
Residential Energy Services Network (RESNET) in the United States. These programs typ-
ically ensure that professional workers and contracting companies are trained to the best
practices and that their performance is regularly verified.
We model such standards as a verifiable minimum quality input q¯.10 The instrument
generates two types of inefficiencies. First, compliance needs to be verified, which generates
enforcement costs M(q¯). These costs do not arise with energy-savings insurance, which
rely on a commonly-observed variable, namely, the energy use reported on utility bills.
Second, just like a uniform insurance contract, minimum quality standards do not account
for consumer heterogeneity.
9Formally, the rebound effect associated with the energy efficiency contract materializes as the positive
slope of the homeowner’s reaction function, which is not affected by the informational context. In contrast,
the rebound effect associated with the energy-savings contract corresponds to an upward shift of the home-
owner’s reaction function. This can be interpreted as demand-side moral hazard, just like the downward
shift of the contractor’s reaction function due to asymmetric information could be interpreted as supply-side
moral hazard.
10In practice, minimum standards could prescribe the grade of materials installed and the application
taken in the installation task.
11
We show in the Online Appendix that the value of q¯ which minimizes the deadweight
loss is such that the marginal disutilities of those homeowners for whom the standard is too
tight and the marginal utilities (net of marginal enforcement costs) of those willing to invest
beyond the standard are equalized.
To sum up, both insurance and standards can mitigate moral hazard, but not eliminate it.
Leaving aside the deadweight loss arising from consumer heterogeneity, which is equally faced
by a uniform insurance contract and a uniform quality standard, the comparison between
the two instruments boils down to how the deadweight loss due to demand-side moral hazard
induced by insurance compares with enforcement costs. This is a context-specific question,
which we examine numerically in Section 4.
Interaction with energy-use externalities. As we saw earlier, undoing moral hazard
has an uncertain effect on energy use due to the rebound effect. To the extent that it
backfires while energy-use externalities remain unpriced, implementing policy remedies to
moral hazard can have the unintended consequence of exacerbating deadweight losses. In
the Online Appendix, we uncover sufficient conditions for this not to occur. In sum, energy-
use externalities are important to consider when performing a welfare analysis of policy
instruments as they can influence the welfare ordering.
3. Empirical evidence of supply-side moral hazard
We now present three empirical facts which together suggest that contractors engage in
moral hazard by poorly installing energy efficiency measures when the quality of their work
is hard to verify. We do so using a rich dataset from a utility-sponsored retrofit program ran
in Florida.
We first find that realized energy savings after retrofit are below predicted savings for
several retrofit measures, a puzzle that is increasingly documented but still highly debated
(Metcalf and Hassett 1999; Fowlie, Greenstone, and Wolfram 2015). The discrepancy is
specifically large for those measures, the quality of which is deemed hard to observe ex post.
For other measures, the quality of which is deemed easy to observe, the magnitude and the
sign of the gap vary widely. We then find that the discrepancy varies as a function of the
day of the week on which a measure was installed and that this variation follows a particular
pattern—realized savings are lower toward the end of the week, notably on Fridays, but
only for those measures, the quality of which is hard to verify ex post. Lastly, we find that
this Friday effect is not driven by selection. That is, contractors do not choose to install
12
particular measures on a specific day of the week (though they might during the weekend).
Crucially, prices are not lower on Fridays when realized savings fall.
The second and third empirical facts are a novel contribution to the debate about predicted-
versus-realized energy savings. Together, they suggest that installation quality is under-
supplied on Fridays for measures specifically prone to moral hazard. To explain the Friday
effect, we argue that workers are more likely to experience negative productivity shocks
toward the end of the workweek and are thus more likely to shrink on quality if the infor-
mational context gives them the opportunity to do so.
3.1. Data
From 2006 to 2012, Gainesville Regional Utilities (GRU) ran rebate programs for home
energy retrofits. The programs targeted a variety of measures, including attic insulation,
duct sealing, air-conditioners, pool pumps, refrigerators and windows. Eligibility for rebate
required that measures be installed by pre-approved contractors. Prior to completion, each
project had to undergo an audit performed by a third party to assess the potential energy
savings associated with the measures, based on ex ante engineering calculations.11 We restrict
our analysis to homes where only one retrofit measure was undertaken.
We consider a sample of 4,099 projects for which the following information is available:
type of measure completed, predicted energy savings, rebate amount, price paid to the con-
tractor, and, crucially for our empirical exercise, the date on which a measure was installed.
Table 1 provides summary statistics. We match the program data with electricity and
natural-gas billing data recorded by GRU between 2002 and 2013. This procedure allows us
to link the characteristics of a retrofit measure to its impact on energy use.
3.2. Empirical strategy
A first challenge in detecting moral hazard is that quality is not directly observed in our
setting. Neither inputs to (e.g., hours worked and skills mobilized by installers, grade of
the products and materials installed) nor outputs of the measures (e.g., number and type of
defects) are documented. Our strategy to detect changes in quality then relies on estimating
realized energy savings, which is strongly correlated with the quality of installation.
11Engineering estimates of energy savings are measure-specific. They take into account home-specific
features (building period, etc.) when necessary, namely for all measures except refrigerators and pool pumps.
13
Our empirical strategy consists to uncover heterogeneity in realized savings along two
dimensions. First, we classify the quality of installation as either easy or hard to observe by
the homeowners, and distinguish energy savings for these two categories of measures. We
consider a measure hard-to-observe if it meets two criteria: (i) the installation is an arduous
task that requires significant labor input and (ii) the quality of installation is difficult to
verify by a non-expert. Attic insulation and duct sealing, which both require significant
installation work and can hardly be verified after completion, belong to this category. Other
retrofit measures that mostly consist of replacing equipment, such as air-conditioners, pool
pumps or refrigerators, are deemed easy to observe. Window replacement, which requires
significant labor input but leaves few features hidden, could fall in either category. In our
preferred specification, we classify it as easy-to-observe. In a robustness check, we move it
into the hard-to-observe category and find little impact.
Second, we allow for heterogeneity in the estimates using interactions with day-of-the-week
dummies. We hypothesize that worker’s productivity is subject to systematic variations over
the week which are unaccounted for in the retrofit contract. In particular, we expect output
to be lower toward the end of the week, namely on Fridays and during weekends. The
motivation comes from labor studies showing that productivity tends to be lower on Fridays,
especially in the construction sector (e.g., Bryson and Forth, 2007). Workers’ fatigue is the
reason most frequently invoked.12 We then compare day-of-the-week estimates between easy
and hard-to-observe measures. Our testable hypotheses is that hard-to-observe measures
deliver fewer energy savings if they have been completed at the end of the workweek, but
that easy-to-observe measures are not subject to such a day-of-the-week effect.
We estimate realized energy savings using a difference-in-differences estimator where the
estimate can vary with respect to the two categories of measure and the day of completion.
The estimation follows the quasi-experimental approach of Fowlie, Greenstone, and Wolfram
(2015),Zivin and Novan (2016), and Maher (2016), only extended with interactions with day-
of-the-week dummies, DW , and dummies that identify hard-to-observe and easy-to-observe
measures, noted HTO and ETO, respectively. We consider the following regression model:
12Other explanations include: staff shortage, e.g., workers calling out ”sick” on Friday; quit-time, e.g.,
workers leaving early to start weekend; backlog, e.g., workers rushing to finish job to avoid having to revisit
site on weekend/next week.
14
(5)
log(kWhit) = αHTOit+γETOit+
D−1∑
d
αd ·DWid ·HTOit+
n−1∑
1
γd ·DWid ·ETOit+λim+αt+it,
where the dependent variable is the logarithm of the total monthly energy use (electricity
plus natural gas) of a particular household. The dummy HTOit turns from zero to one
the month household i invests in a hard-to-observe retrofit measure. The dummy ETOit is
defined similarly for investments in easy-to-observe measures. The terms λim and αt denote
household-calendar month fixed effects and month-of-sample fixed effects, respectively, and
implement the difference-in-differences estimator to estimate energy savings. Finally, DWid
is a dummy that identifies a specific day (or period) of the week d and estimates heterogeneity
in average realized energy savings with respect to the day of completion. Each dummy takes
a value of one if household i got the measure completed on day d and zero otherwise. For
instance, ”Friday” effects are estimated by having a dummy, DWiFriday, that turns on if the
measure was completed on a Friday; the omitted dummy corresponds to the remaining days
of the week (Monday to Thursday). The coefficient αd estimates the specific effect of day d
on realized energy savings for hard-to-observe measures; we expect it to be positive if moral
hazard exists. Coefficient γd estimates the same effect for easy-to-observe measures. This is
essentially a placebo test—we expect the coefficient to be zero.
3.3. Identification
The validity of the difference-in-differences estimator is guaranteed by two sets of fixed effects.
The dummies λim captures all household-specific characteristics that influence energy use.
Note that we allow for variation by calendar month m, which captures any seasonal pattern in
household-specific energy usage. Exploiting several years of monthly consumption allows us
to identify the coefficients λim. The month t-of-sample fixed effects (αt) control for weather
and any other contemporaneous shocks that may affect monthly consumption. Our large
sample of retrofits allows us to identify the coefficients αt. We thus effectively assume that
conditional on λim and αt, households were subject to similar trends in energy usage prior
to the retrofit measure. By the end of our time horizon, all households are treated.
The estimation is restricted to the sample of single-measure retrofit projects, which leads
us to exclude projects that combine several measures, such as low-income programs and
whole-home performance. The restriction ensures that we focus on households that had only
one interaction with the contractors, which facilitates how we classify a project as easy or
15
hard to observe. We further restrict the sample to projects that cost under $10,000. The
reasoning is that excluded projects are likely to take more than one day.13 Ultimately, the
working sample contains 4,099 projects, including 1,026 hard-to-observe measures and 3,073
easy-to-observe ones.
The validity of the test for moral hazard relies on the assumption that contractors do not
select the day of the week on which they complete a particular measure. Such selection can
be directly tested by comparing projects along key dimensions using observable attributes of
the retrofit contracts. Table 2 compares the average retrofit prices, average predicted energy
savings, average rebate amounts, and number of retrofit performed across day (or period) of
completion. For all four variables, we observe no statistically significant difference during the
days of the workweek. Importantly for our identification, Friday jobs are of the same size as
other jobs, as measured by both their price and predicted energy savings; any difference in
realized energy savings is therefore due to other factors. In contrast, weekend projects, which
happen to be very few, do exhibit some differences. This suggests that selection might be at
play during the weekend, which leads us to treat weekends separately in the estimation.14
3.4. Results
Columns 2 and 3 of Table 1 replicate the findings of Maher (2016). Most retrofit measures
subsidized by GRU exhibit a discrepancy between predicted and realized savings. For the
two main retrofit measures classified as hard-to-observe, attic insulation and duct repair,
realized savings are 60% and 32%, respectively, of predicted savings.15 For easy-to-observe
measures, the discrepancy varies widely in magnitude but also in sign. For three out of seven
of them, realized savings are indeed well above predicted ones.
Regression results that account for heterogeneity with respect to the day-of-the-week and
observability of the measure (Equation 5) are displayed in Table 3. Each column corresponds
to a specification that identifies a specific day-of-the-week effect on energy savings. The
interaction terms are the additional savings for the day identified. For instance, for the
model that estimates the Friday effect (column V), the coefficient represents the additional
13The data do not allow us to ascertain that completion of single measures takes exactly one day, that is,
less than eight hours. We nevertheless think this is a reasonable assumption for the technologies considered.
14The problem could come from the way contractors either favor certain projects or elect to work during
the weekend.
15The estimate of realized saving for duct repair is not statistically different from zero, while predicted
saving are 107.5 kWh/month.
16
savings relative to the average savings for Monday to Thursday. A positive estimates means
that realized savings on that particular day are lower than for other days of installation.
For measures classified as hard-to-observe, we find a positive, economically large and
statistically significant Friday effect, but we do not find any such effect for other days of
the workweek. All models include an interaction for weekend effect, which is always positive
but not statistically significant. This effect is still present if we group Friday and weekend
estimates together (column VI). The magnitude of the estimates, obtained by adding the
coefficients HTO and HTO · DHFriday imply that the realized energy savings for hard-to-
observe measures completed at the end of the week are close to zero and not statistically
significant. For other days of the week, the savings are of the expected sign, economically
large and statistically significant. Importantly, the Friday-effect only holds for hard-to-
observe measures. For easy-to-observe measures, we do observe large savings on each workday
and slightly larger savings on Wednesdays.
The magnitude of the Friday effect is large. Table 4 displays estimates of the realized
savings with and without controls for the Friday effect, and compares them to the pre-
dicted savings. After controlling for the Friday effect, the discrepancy between predicted
and realized energy savings shrinks by 65% for hard-to-observe retrofit measures. It remains
unaffected for easy-to-observe measures, regardless of the controls (Table 7, Appendix).
These results are robust to a number of specification tests presented in the Appendix.
Excluding low-income weatherization and home performance projects have little impact.16
Performing the estimation in level instead of logarithm still produces a positive Friday effect
for hard-to-observe measures. The coefficient is, however, marginally statistically signifi-
cant.17
Altogether, the results give support to our hypothesis that under-treatment occurs on
Fridays when quality is hard to verify. Though essentially a positive test of existence of
moral hazard, our analysis additionally indicates that the problem is important, perhaps
enough to undo all energy savings.
16Classifying windows replacement as a hard-to-observe measure does not qualitatively change the results
either.
17The difference in the Friday effect for easy versus hard-to-observe measures has a p-value of 0.0793.
17
4. A numerical illustration
We now specify and parameterize the model developed in Section 2 to quantify the energy
efficiency gap due to moral hazard and compare policy solutions in the presence of energy-
use externalities. Our numerical exercise focusses on attic insulation, which is frequently
advocated as the most cost-effective measure to reduce energy use in the residential sector.
Parameterization relies on the GRU program data, results from the econometric analysis
conducted in the previous section, and the U.S. Residential Energy Consumption Survey
(RECS).
4.1. Model specifications
Supply side. The impact of quality on energy use is modeled with the following functional
form:
(6) E(s, q) ≡
(
1−Gmin − (Gmax −Gmin)
(
1− e−ωq
))
E0(s),
where the quality of installation q is assumed to be dimensionless and ranging from 0% to
100%. Parameters Gmin and Gmax capture the minimum and maximum efficiency improve-
ment that the investment can deliver, respectively. We set these values to 1% and 15%
based on the range of predicted energy savings estimated in the GRU program. Parameter
ω captures the sensitivity of the reduction in energy use with respect to the level of quality
provided. We set ω so that the efficiency evaluated at q = 95% is the 95th percentile of the
efficiency estimated for attic insulation in Section 3 (12.6%).
The cost of supplying quality is specified as follows:
(7) C(q) ≡ K + φ2 q
2,
with K > 0 and φ > 0. These parameters are calibrated using two moments derived from
the GRU data so that (i) the equilibrium quality under perfect information generates the
95th percentile of the predicted energy savings estimated for attic insulation in Section 3
(7.5%) and (ii) the cost of this equilibrium quality matches the median cost estimate, $678.
Demand side. The homeowner’s utility and energy use are specified with the following
functional forms:
(8) V (s) ≡ Vmax
(
1− e−α(s−smin)
)
,
18
(9) E0(s) ≡ β(s− smin)γ,
with α > 0, β > 0 and γ > 1. Both functions vary with indoor temperature s. They are
parameterized using information on indoor temperature, energy use, energy expenditure and
income contained in the RECS online database for 2009. We extract a preliminary sample
of 4,306 U.S. households who own and occupy their house and pay for natural gas for space
heating. We then remove households who declare a winter daytime temperature below 60◦F
or above 80◦F and thereby obtain a working sample of 4,266 households.18 This sample
covers 35% of the complete dataset.
Calibration of the parameters relies on moments derived for the median homeowner,
assumed to be of type θ = 1. It proceeds as follows:
• The maximum value derived from space heating, Vmax, is set equal to $2,816. This
value is obtained by assuming that 4.3% of the annual median income ($65,000) is
used for space heating. This fraction corresponds to the 95th percentile of the income
share dedicated to space heating in the RECS sample.
• Parameters α, β and γ are calibrated so that: (i) the optimal temperature to the
median homeowner is 69◦F, the median of the RECS sample; (ii) at this tempera-
ture, the median homeowner’s annual use of natural gas for space heating is 50 MCF,
the median of the RECS dataset; (iii) the price-elasticity of energy use evaluated at
s = 69◦F is −0.4, the middle value of the range found in the literature by Gilling-
ham, Newell, and Palmer (2009) for short-term price-elasticities of natural gas use
([−0.03;−0.76]).
• Non-energy attributes  are calibrated so that the median homeowner is the marginal
participant.19 That is, she is indifferent between not investing and investing under
asymmetric information: NPV AI(θ = 1) = 0. This leads to non-energy net benefits
of $441.
18We set thus set smin = 60◦F≤ s ≤ 80◦F.
19Participation is therefore set to 50% under asymmetric information. In the RECS sample, the annual
fraction of homeowners investing in insulation is 3.4% (6.8% of the population declare having insulation
installed in the last two years), a rate close to 2.9%, which would be the annual rate if investment occurred
once every 35 years. Therefore, a participation of 100% in our model can be interpreted as an annual
insulation rate of 6.8% in the population.
19
The price of energy p is set equal to $10.5 per thousand cubic feet (MCF), which is the
variable part of the GRU residential rates for natural gas in 2015. This is close to the average
price paid for natural gas use for space heating in the RECS dataset ($11.14/MCF) and the
$11.45/MCF reported by Davis and Muehlegger (2010). The discount rate r is set equal to
7%, the value recommended by the U.S. Office of Management and Budget (OMB 2009) to
assess private investment. The investment horizon is set equal to 35 years, the conventional
lifetime of attic insulation
Finally, to model heterogeneity in θ, the distribution of temperatures observed in the
RECS sample (mean 69.0◦F, standard deviation 3.4◦F) is fitted with a log-normal distribu-
tion of homeowners’ types θ with parameters µ = 0 and σ = 1 (see Figure 1).
Energy-use externalities. As we saw in Section 2, the comparison of policy solutions to
moral hazard will be affected by energy-use externalities. In our policy scenarios, we assume
that the externalities associated with natural gas are valued at $1.69/MCF, reflecting a social
value of carbon of $33/tCO2.
4.2. Quantification of the energy efficiency gap
Figure 3 maps different equilibra in the framework proposed by Jaffe, Newell and Stavins20
(1994), so as to visualize the trade-offs between economic efficiency (measured as present dis-
counted welfare) and energy efficiency. In Jaffe and Stavins’ words, the perfect-information
equilibrium with unpriced externalities corresponds to the ”Narrow economists’ optimum.”
Further improvements along both the energy efficiency and welfare dimensions can occur if
externalities are internalized through a Pigouvian price, establishing the ”True social opti-
mum”. The welfare gains from undoing moral hazard ($507) are 2.4 times larger than those
from internalizing energy-use externalities ($210). The difference is in part due to the small
size of the marginal externality, $1.69/MCF, relative to the price of energy, $10.50/MCF,
which approximates the marginal loss due to inefficient investment under moral hazard.21
Sensitivity analysis. The key parameters of the model are varied according to assump-
tions outlined in Table 5. Scenario variants are meant to mimic different barriers to energy
20This is the ultimate version of a diagram which first appeared in .
21The moral-hazard market failure can also be restated as an average implied discount rate of 18%. This
value is computed by solving and averaging the θ-specific discount rates that match the quality offered under
perfect information with the net present value enjoyed by the homeowner under asymmetric information,
discounted at 7% by assumption. It is consistent with the estimates reported in the empirical literature on
energy efficiency investments ().
20
efficiency, which, according to the taxonomy referred to in the introduction, are categorized
either as market failures, nonmarket failures or behavioral anomalies. Figure 4 displays the
size of the deadweight losses due to moral hazard across scenarios together with the elasticity
of the deadweight loss with respect to each parameter.
Parameterization of the efficiency function is the most sensitive determinant of the dead-
weight loss. This underlines the importance of reducing errors in the measurement of technol-
ogy efficiency to perform accurate welfare analysis of energy efficiency programs. The extent
of distortions in energy markets and in the capitalization of energy savings are the next
important determinants. This illustrates the importance of potential interactions between
market failures. In contrast, under-valuation of energy savings and errors in the measure-
ment of technology cost have little impact. Lastly, the parameters of homeowners’ utility
have no influence, a result we formalize below.
The relative magnitude of moral hazard and externalities varies widely, from around 10:1
in the “High efficiency” scenario to around 1:3 in the “High discount rate” scenario. This wide
range of variation stresses the critical aspect of parameterization, but nonetheless suggests
that moral hazard deserves at least as much consideration as externalities.
A sufficient statistic of the deadweight loss. To formalize parameter sensitivity, we
show in the Online Appendix that the deadweight loss can be approximated by the following
sufficient statistic:
(10) ∆W = −p∆E|sΓ(r, l)−∆C
The formula weighs the cost of quality against its benefits in terms of gross energy savings.
This corresponds to a net present value calculation that only takes into account technological
information. It does not require knowledge of the utility function V (·) nor the effect of the
energy service s on energy use E(·, ·). This explains why parameters Vmax and σ are found
to have no influence on the deadweight loss. It also means that the direct rebound effect
can be ignored. Still, the formula contains the key parameters of the market and behavioral
environment p, l and r. In our simulations, we find that it never underestimates the exact
deadweight loss by more than 7% across scenarios.
21
4.3. Insurance versus standard
Figure 3 displays the welfare effects of energy-savings insurance and minimum quality stan-
dards, with and without externalities. Optimal insurance contracts, which are homeowner-
specific, have an average coverage of 18%. Since homeowner types may not be perfectly
observable, it is worth also considering uniform insurance contracts. As depicted in the
figure, incremental coverage initially improves both energy efficiency and welfare, up to a
point where efficiency becomes so expensive that it starts deterring participation, hence a
backward bend in both efficiency and welfare trajectories. The best such contract is located
farthest to the right of the horizontal axis. Absent externalities, the associated coverage is
22%. Otherwise, coverage is 10%. The welfare gains with the best uniform insurance are
only slightly below those provided by optimal homeowner-specific contracts. This means
that in the context considered here, ignoring consumer heterogeneity in policy design has
little implication. The point also applies to standards. Just like uniform insurance, mini-
mum quality standards draw an ellipse driven by increasing stringency. The best standard
mandates a quality of 43% if externalities are absent and 49% otherwise. Anyway, it brings
market agents very close to the social optimum.
Overall, insurance mitigates almost half of the deadweight loss due to moral hazard if
externalities are absent, but it has little effect otherwise. The best standard appears sub-
stantially more efficient in the figure. Yet unlike insurance, standards require verification,
monitoring and enforcement. In practice, this would require an audit post-installation. As
reported by Palmer, Walls, Gordon, and Gerarden (2013), the cost of an audit for retrofits
is on average $347 in the United States. Accounting for this as enforcement cost, the net
effect of the best standard vanishes and policy comparison is more ambiguous.
An important takeaway is that unlike that of insurance, the performance of standards is
relatively unaffected by energy-use externalities. This is due to the nature of the rebound
effect induced by each instrument. Both policies induce a direct rebound effect, as they
increase energy efficiency and thus lower the marginal cost of energy services. In addition,
insurance coverage further reduces marginal energy expenditures. This causes demand-side
moral hazard—an over-use of energy which exacerbates externalities. Failure to internalize
energy-use externalities thus favors minimum quality standard over energy-savings insurance.
22
5. Conclusion
Many energy efficiency measures can be thought of as credence goods, the performance of
which is never fully revealed to the buyer. This characteristic is conducive to a variety of
information asymmetries which can cause under-investment in energy efficiency, a phenom-
enon known as the energy efficiency gap. In this paper, we were interested in the existence
of and solutions to moral hazard in the quality of installation of energy efficiency measures.
We provide empirical evidence of this market failure in the U.S. home retrofit market.
Using data from a utility-sponsored retrofit program ran in Florida, we find that for mea-
sures such as attic insulation and duct sealing, the quality of which is hard to observe ex
post, energy savings are significantly lower when the retrofit was completed on a Friday—a
day particularly prone to negative shocks on workers’ productivity—than on any other week-
day. We interpret this outcome as supply-side moral hazard and show that it can explain a
large fraction of the discrepancy between predicted and realized energy savings, an increas-
ingly documented manifestation of the energy efficiency gap. In theory, the problem can be
addressed by private interventions such as energy-savings insurance or public interventions
such as minimum quality standards. We show that neither intervention can eliminate the
loss: insurance induce demand-side moral hazard, as lower marginal expenditures encourage
over-use of energy; standards incur enforcement costs. The comparison between the two is
therefore context-specific. Our numerical model suggests that while energy-savings contracts
with small insurance coverage (typically 10-20%) can significantly mitigate the moral hazard,
they also amplify energy-use externalities. Minimum quality standards therefore seem more
desirable if externalities are left unpriced.
We see several interesting extensions to our analysis. On the theoretical front, attention
should be focused on reputation strategies. The finding that energy efficiency providers do
offer some quality during most of the workweek suggests the incentive might be important.
On the empirical front, new experiments allowing for direct observation of quality should be
designed, in order to directly investigate the link between installation defects and missing
energy savings. On the policy front, efforts should be devoted to ex post evaluation of energy
performance contracts, quality certifications and other existing remedies to moral hazard.
23
Acknowledgements
The authors started this research at the Precourt Energy Efficiency Center (PEEC), Stan-
ford University. We thank PEEC for funding and James Sweeney for encouragements and
comments. We are grateful to Tiger Adolf, Rick Diamond, Matt Golden, Evan Mills and
Carol Zabin for early discussions that motivated this work. We thank Philippe Bontems,
Fre´de´ric Branger, Lawrence Goulder, Christian Huse, Franck Lecocq, Marie-Laure Nauleau,
Philippe Quirion, and Adrien Vogt-Schilb for comments on earlier drafts. Lastly, we thank
participants to the various seminars and conferences where the paper was presented for their
feedback.
24
References
Abbring, J. H., J. J. Heckman, P.-A. Chiappori, and J. Pinquet (2003): “Adverse
Selection and Moral Hazard in Insurance: Can Dynamic Data Help to Distinguish?,”
Journal of the European Economic Association, 1(2-3), 512–521.
Allcott, H., and M. Greenstone (2012): “Is There an Energy Efficiency Gap?,” Jour-
nal of Economic Perspectives, 26(1), 3–28.
Allcott, H., S. Mullainathan, and D. Taubinsky (2014): “Energy policy with ex-
ternalities and internalities,” Journal of Public Economics, 112, 72–88.
Balafoutas, L., A. Beck, R. Kerschbamer, and M. Sutter (2013): “What Drives
Taxi Drivers? A Field Experiment on Fraud in a Market for Credence Goods,” The Review
of Economic Studies, 80(3), 876–891.
Brounen, D., and N. Kok (2011): “On the economics of energy labels in the housing
market,” Journal of Environmental Economics and Management, 62(2), 166–179.
Bryson, A., and J. Forth (2007): “Are There Day of the Week Productivity Effects?,”
Discussion paper, Manpower Human Resources Lab.
Campolieti, M., and D. E. Hyatt (2006): “Further Evidence on the “Monday Effect”
in Workers’ Compensation,” Industrial & Labor Relations Review, 59(3), 438–450.
Cooper, R., and T. W. Ross (1985): “Product Warranties and Double Moral Hazard,”
The RAND Journal of Economics, 16(1), 103–113.
Davis, L. W. (2012): “Evaluating the Slow Adoption of Energy Efficient Investments: Are
Renters Less Likely to Have Energy Efficient Appliances?,” in The Design and Implemen-
tation of U.S. Climate Policy, ed. by D. Fullerton, and C. Wolfram, pp. 301–316. University
of Chicago Press.
Davis, L. W., A. Fuchs, and P. Gertler (2014): “Cash for coolers: evaluating a large-
scale appliance replacement program in Mexico,” American Economic Journal: Economic
Policy, 6(4), 207–238.
Davis, L. W., and E. Muehlegger (2010): “Do Americans consume too little natural
gas? An empirical test of marginal cost pricing,” The RAND Journal of Economics, 41(4),
791–810.
de Wilde, P. (2014): “The gap between predicted and measured energy performance of
buildings: A framework for investigation,” Automation in Construction, 41, 40–49.
25
Dulleck, U., and R. Kerschbamer (2006): “On Doctors, Mechanics, and Computer
Specialists: The Economics of Credence Goods,” Journal of Economic Literature, 44(1),
5–42.
Dulleck, U., R. Kerschbamer, and M. Sutter (2011): “The Economics of Credence
Goods: An Experiment on the Role of Liability, Verifiability, Reputation, and Competi-
tion,” American Economic Review, 101(2), 526–555.
Fowlie, M., M. Greenstone, and C. Wolfram (2015): “Do Energy Efficiency Invest-
ments Deliver? Evidence from the Weatherization Assistance Program,” Working Paper
21331, National Bureau of Economic Research.
Fuerst, F., P. McAllister, A. Nanda, and P. Wyatt (2015): “Does energy effi-
ciency matter to home-buyers? An investigation of EPC ratings and transaction prices in
England,” Energy Economics, 48, 145–156.
Gillingham, K., M. Harding, and D. Rapson (2012): “Split Incentives in Residential
Energy Consumption,” The Energy journal, 33(2), 37–62.
Gillingham, K., R. G. Newell, and K. Palmer (2009): “Energy Efficiency Economics
and Policy,” Annual Review of Resource Economics, 1(1), 597–620.
Gillingham, K., and K. Palmer (2014): “Bridging the Energy Efficiency Gap: Pol-
icy Insights from Economic Theory and Empirical Evidence,” Review of Environmental
Economics and Policy, p. ret021.
Houde, S., and J. E. Aldy (2014): “Belt and Suspenders and More: The Incremental
Impact of Energy Efficiency Subsidies in the Presence of Existing Policy Instruments,”
Discussion paper, National Bureau of Economic Research.
Jaffe, A. B., and R. N. Stavins (1994): “The energy-efficiency gap: What does it
mean?,” Energy Policy, 22(10), 804–810.
Krishnamurthy, C. K. B., and B. Kristro¨m (2015): “How large is the Owner-Renter
Divide in Energy Efficient Technology? Evidence from an OECD cross-section,” The
Energy Journal, 36(4).
Levinson, A., and S. Niemann (2004): “Energy use by apartment tenants when landlords
pay for utilities,” Resource and Energy Economics, 26(1), 51–75.
Maher, J. (2016): “Measuring the Accuracy of Engineering Models in Predicting Energy
Savings from Home Retrofits: Evidence from Monthly Billing Data,” Working paper.
26
Maruejols, L., and D. Young (2011): “Split incentives and energy efficiency in Canadian
multi-family dwellings,” Energy Policy, 39(6), 3655–3668.
Metcalf, G. E., and K. A. Hassett (1999): “Measuring the Energy Savings from Home
Improvement Investments: Evidence from Monthly Billing Data,” Review of Economics
and Statistics, 81(3), 516–528.
Mills, E. (2003): “Risk transfer via energy-savings insurance,” Energy Policy, 31(3), 273–
281.
Murphy, L. (2014): “The influence of energy audits on the energy efficiency investments
of private owner-occupied households in the Netherlands,” Energy Policy, 65, 398–407.
Myers, E. (2013): “Asymmetric Information in Residential Rental Markets: Implications
for the Energy Efficiency Gap,” EI @ Haas WP 246.
OMB, U. S. O. o. M. a. B. (2009): “Informing regulatory decisions: report to Congress
on the costs and benefits of federal regulations and unfunded mandates on state, local,
and tribal entities.,” Discussion Paper Appendix D, OMB Circular A-4.
Palmer, K., M. Walls, H. Gordon, and T. Gerarden (2013): “Assessing the energy-
efficiency information gap: results from a survey of home energy auditors,” Energy Effi-
ciency, 6(2), 271–292.
Schneider, H. S. (2012): “Agency Problems and Reputation in Expert Services: Evidence
from Auto Repair,” The Journal of Industrial Economics, 60(3), 406–433.
Scott, S. (1997): “Household energy efficiency in Ireland: A replication study of ownership
of energy saving items,” Energy Economics, 19(2), 187–208.
Sorrell, S. (2004): “Understanding barriers to energy efficiency,” in The Economics of
Energy Efficiency: Barriers to Cost-Effective Investment, ed. by S. Sorrell, E. O’Malley,
J. Schleich, and S. Scott, pp. 25–93. Edward Elgar, Cheltenham.
Sunikka-Blank, M., and R. Galvin (2012): “Introducing the prebound effect: the gap
between performance and actual energy consumption,” Building Research & Information,
40(3), 260–273.
Zabin, C., T. W. Lester, and J. Halpern-Finnerty (2011): “Chapter Four: Case
studies of the HVAC, residential retrofit, and commercial lighting control sectors,” in
California workforce education and training needs assessment: for energy efficiency, dis-
tributed generation, and demand response, pp. 89–118. Institue for Research on Labor and
Employment, University of California, Berkeley.
27
Zivin, J. G., and K. Novan (2016): “Upgrading Efficiency and Behavior: Electricity
Savings from Residential Weatherization Programs,” The Energy Journal, 37(4).
28Table 1. Summary statistics by energy efficiency measure
Predicted Predicted Maher (2016)’s DiD Project Rebate Day of Observability
energy savings energy savings energy savings price amount completion
(MWh/year) (kWh/month) (kWh/month)
mean (SD) mean (SD) coefficient (SE) mean (SD) mean (SD) median # Obs.
Attic Insulation 1.55 129.17 78.65∗∗∗ $761 $199 Wednesday Hard 575
(0.04) (3.33) (26.91) ($501) ($77)
Duct Repair 1.29 107.50 33.72 $863 $360 Wednesday Hard 367
(0.08) (6.67) (33.77) ($911) ($99)
Low-Interest - - 73.04 - $1,089 Wednesday Hard 84
Loan - - (79.79) - ($505)
Super-SEER 1.93 160.83 208.9∗∗∗ $7,291 $555 Wednesday Easy 623
Central AC (0.53) (44.17) (27.11) ($1,467) ($62)
SEER-15 0.55 45.83 161.3∗∗∗ $5,672 $295 Wednesday Easy 297
Central AC (0.12) (10.00) (37.60) ($1,412) ($44)
Pool Pump 1.76 146.67 101.6** $1,452 $284 Wednesday Easy 394
(0.16) (13.33) (47.05) ($505) ($97)
Refrigerator 1.53 127.50 47.57∗∗∗ - $72 Wednesday Easy 1,162
Removal (0.40) (33.33) (17.47) - ($12)
Low-Income 1.12 93.33 52.87 $301 $2,170 Wednesday Easy 227
Grant (0.82) (68.33) (38.52) ($339) ($1,528)
Whole Home 2.49 207.50 131.1∗∗∗ $654 $967 Wednesday Easy 342
Performance (1.00) (83.33) (43.88) ($239) ($289)
Low-E Window 0.66 55.00 155.2** $3,511 $159 Wednesday Easy 28
(0.36) (30.00) (72.66) ($2,205) ($81)
Total 4,099
Notes: The estimates in the column labelled “Maher (2016)’s DiD” are technology-specific estimates of energy savings taken
directly from Maher (2016). They are computed using a difference-in-differences estimator similar to that used to estimate the
savings with day-of-the-week effects. The regression model does not consider heterogeneity with respect to the day of installation
and estimates average savings for each measure. Standard errors are clustered at the household level (* p <.10 ** p <.05 ∗∗∗
p <.01). The column “Observability” classifies the quality of a measure as “Easy” or “Hard” to observe ex post.
29
Table 2. Summary statistics by day of the week
Monday Tuesday Wednesday Thursday Friday Weekend All Days # Obs.
A. Predicted energy savings (MWh per year)
HTO measures 1.38 1.28 1.14 1.06 1.34 1.33 1.25 943
(.43) (.63) (.82) (.91) (.5) (.67) (0.69)
ETO measures 1.57 1.55 1.59 1.45 1.55 1.41 1.54 2,801
(.65) (.59) (.58) (.71) (.71) (.62) (0.64)
B. Project price (dollars)
HTO measures 777 791 760 621 770 603 728 734
(654) (776) (1,002) (544) (669) (619) (738)
ETO measures 2,973 1,567 1,355 2,988 3,040 895 2,003 1,060
(3,271) (2,720) (2,660) (3,252) (3,211) (1,974) (2,976)
C. Rebate amount (dollars)
HTO measures 286 335 373 413 295 254 329 1,026
(181) (250) (369) (375) (243) (236) (291)
ETO measures 622 404 351 632 552 372 474 3,073
(825) (628) (570) (841) (687) (700) (703)
# Obs. 590 842 1,008 643 657 359 4,099 4,099
Notes: Mean and standard deviation (in parentheses) for the three main variables reported on con-
tracts. Differences in the number of observations across variables are due to missing information.
The number of observations for a given day include all contracts with information for at least one
variable. The label HTO refers to hard-to-observe measures and ETO refers to easy-to-observe
measures.
30
Table 3. Day-of-the-week effects
Dep. Variable: Monday Tuesday Wednesday Thursday Friday Friday+WE
Log(kWh+Nat.
Gas/month)
Effect Effect Effect Effect Effect Effect
HTO=1 -0.0254∗ -0.0299∗ -0.0292∗ -0.0400∗∗ -0.0507∗∗∗ -0.0507∗∗∗
(0.0152) (0.0165) (0.0168) (0.0165) (0.0169) (0.0169)
ETO=1 -0.0561∗∗∗ -0.0617∗∗∗ -0.0419∗∗∗ -0.0613∗∗∗ -0.0586∗∗∗ -0.0586∗∗∗
(0.00928) (0.00970) (0.00977) (0.00941) (0.00916) (0.00916)
HTO=1 X -0.0508
Monday (0.0469)
ETO=1 X -0.00151
Monday (0.0183)
HTO=1 X -0.0266
Tuesday (0.0336)
ETO=1 X 0.0223
Tuesday (0.0159)
HTO=1 X -0.0271
Wednesday (0.0307)
ETO=1 X -0.0480∗∗∗
Wednesday (0.0159)
HTO=1 X 0.0259
Thursday (0.0334)
ETO=1 X 0.0303
Thursday (0.0186)
HTO=1 X 0.0720∗∗
Friday (0.0318)
ETO=1 X 0.0149
Friday (0.0218)
HTO=1 X 0.0581∗∗
Friday + WE (0.0274)
ETO=1 X -0.0528
Friday + WE (0.0328)
HTO=1 X 0.00916 0.0136 0.0132 0.0238 0.0345
WE (0.0391) (0.0399) (0.0401) (0.0399) (0.0400)
ETO=1 X -0.0157 -0.0102 -0.0296 -0.0105 -0.0132
WE (0.0286) (0.0287) (0.0287) (0.0286) (0.0285)
Observations 684,015 684,015 684,015 684,015 684,015 684,015
R-squared 0.647 0.647 0.647 0.647 0.647 0.647
Notes: Standard errors are clustered at the household level (* p <.10 ** p <.05 *** p <.01).
The label HTO refers to hard-to-observe retrofit measures, and ETO refers to easy-to-observe
measures. The first two rows of estimates correspond to the energy savings for the days of the
week not captured by the interaction effect in the lower rows. The last column groups Fridays
and weekends together.
31
Table 4. Gap Realized-Predicted Savings Without Friday-Effect
Realized Realized Predicted Gap
Savings Savings Savings Predicted-Realized
kWh/month MWh/year MWh/year Savings
HTO Mon-Th 90.65 1.088 1.21 0.122
HTO Mon-Fri 74.29 0.891 1.24 0.349
Difference (%) -22.0 -22.0 2.42 64.9
Notes: The realized savings in the first column are computed after con-
trolling for the Friday effect. The realized savings in the second are not
controlled for the Friday effect. Estimates are reported for hard-to-observe
measures only. Other regression estimates are provided in the Appendix
(Table 7). Controlling for the Friday effect reduces the discrepancy between
realized and predicted savings by 65%.
32Table 5. Scenarios for sensitivity analysis
Scenario Parameter Reference Scenario Scenario Nature of the barrier Included in
varied value value interpretation to efficiency sufficient
statistic?
”High efficiency” Gmax 15% 30% Upper range of the po-
tential for energy effi-
ciency claimed by engi-
neering studies, manu-
facturers, etc.
Nonmarket failure Yes
”Low energy price” p $10.5/MCF $6.3/MCF Removal of a 40%
mark-up associated
with average cost
pricing of natural gas,
as estimated by Davis
and Muehlegger (2010)
Market failure (price dis-
tortion)
Yes
”Low capitalization” l 35 years 10 years Homeowners do not
value energy savings
beyond the typical
length of ownership of
a dwelling
Market failure (informa-
tion asymmetry)
Yes
”High discount rate” r 7% 20% Typical under-
valuation found in
empirical analyses
of energy efficiency
decisions
Behavioral anomaly Yes
”High technology cost” δ 1,024 3,000 The total cost for
the highest quality is
$3,500, which matches
that faced by the 99th
percentile of the GRU
dataset
Nonmarket failure Yes
”High heating valua-
tion”
Vmax $2,816 $4,000 6% of median income
dedicated to heating
Nonmarket failure No
”Low heterogeneity” σ 1 0.25 Narrow distribution,
with thermostat set-
tings ranging from
68◦F (5th percentile)
to 70◦F (95th per-
centile)
Nonmarket failure No
33
Table 6. Simulation results, averaged over the population of total mass 1
Without Pigouvian price With Pigouvian price
Model output Unit Before
investment
Asymmetric
informa-
tion
Perfect in-
formation
Optimal in-
surance
Before
investment
Asymmetric
informa-
tion
Perfect in-
formation
Welfare improvement, without
externalities
$ 0 11 364 176 -196 -185 180
Welfare improvement, with ex-
ternalities
$ 0 24 531 35 222 247 741
Homeowners’ equilibrium tem-
perature
◦F 68.9 68.9 69.2 69.8 67.7 67.7 68.0
Annual natural gas use for
space heating
MCF 50.2 50.0 47.4 52.6 43.1 42.9 40.7
Annual natural gas expendi-
ture
$ 527 525 498 552 453 451 427
Annual CO2 emissions tCO2 2.6 2.6 2.4 2.7 2.2 2.2 2.1
Annual external cost of CO2
emissions
$ 85 84 80 89 73 73 69
Contractor’s equilibrium qual-
ity
0% 41% 16% 0% 46%
Energy efficiency of insulation 0.5% 8.1% 4.0% 0.7% 8.6%
Rebound effect 9% 31% 219% 21% 34%
Cutoff type of the marginal
participant
1.00 0.16 0.32 0.67 0.18
Participation rate 50% 97% 87% 66% 96%
Zero-profit insulation price $ 509 690 540 509 739
Homeowner’s net present value $ 22 377 202 37 542
Insurance premium 1,725
Insurance optimal coverage 18%
Notes: ”Energy efficiency” is averaged over the whole population. The average over participants is obtained
by dividing ”Energy Efficiency” by ”Participation rate”. Welfare improvements are measured against present
discounted welfare before investment, without a Pigouvian price ($49,433 without externalities, $46,463 with
externalities).
34
Figure 1. Model fit to RECS temperature data. The simulated prob-
ability distribution is calculated with the triangle method.
35
Figure 2. Energy efficiency gap in the reference scenario. The hor-
izontal axis represents average present discounted welfare for different equi-
libria, with energy-use externalities valued at $33/tCO2. The vertical axis
represents average energy efficiency, with non-participating homeowners get-
ting 0%.
.
36
Figure 3. Impact of policies in the reference scenario. Uniform insurance draw a parametric
curve with coverage increasing counter-clockwise by 1%, from 0% to 100%. Similarly, uniform
standards draw a parametric curve with mandated quality increasing counter-clockwise by 1%,
from 0% to 100%. The best uniform insurance has a coverage of 22% without externalities and 10%
otherwise. The best uniform standard mandates a quality of 43% without externalities and 49%
otherwise.
37
Figure 4. Sensitivity analysis. The deadweight loss from moral hazard
is calculated as the welfare difference between perfect and asymmetric infor-
mation, both with unpriced externalities; the externalities are calculated as
the welfare difference between priced and unpriced externalities, both under
perfect information. Elasticies, which measures each parameter’s influence,
are calculated as the percentage change in the deadweight loss due to moral
hazard divided by the percentage change in parameter value. They are all posi-
tive, except for the ”High discount rate” and ”High technology cost” scenarios.
They are reported in the figure in absolute value. Scenario assumptions are
detailed in Table 5.
38
6. Appendix: Additional Regression Result
Table 7. Realized Energy Savings: Controlling for Friday-Effect
Dep. Variable: Controlling for Not Controlling for
kWh+Nat.
Gas/month
Friday-Effect Friday-Effect
MH=1 -90.65∗∗∗ -74.29∗∗∗
(27.80) (23.33)
NoMH=1 -103.3∗∗∗ -107.8∗∗∗
(13.46) (12.59)
MH=1 X 74.34
Friday (46.29)
NoMH=1 X -29.71
Friday (37.15)
MH=1 X 120.9∗∗ 104.8∗∗
WE (49.90) (47.28)
NoMH=1 X -63.52 -58.86
WE (58.67) (58.50)
Constant 1,565∗∗∗ 1,565∗∗∗
(7.883) (7.882)
Observations 692,879 692,879
R-squared 0.679 0.679
Notes: Standard errors are clustered at the household level (*
p<.10 ** p<.05 *** p<.01).
39
Table 8. Day-of-the-Week Effect: No Low-Income Grant and Home Perfor-
mance Projects
Dep. Variable: Monday Tuesday Wednesday Thursday Friday Friday+WE
Log(kWh+Nat.
Gas/month)
Effect Effect Effect Effect Effect Effect
MH=1 -0.0203 -0.0247 -0.0241 -0.0349∗∗ -0.0455∗∗∗ -0.0455∗∗∗
(0.0155) (0.0167) (0.0169) (0.0166) (0.0171) (0.0171)
NoMH=1 -0.0502∗∗∗ -0.0552∗∗∗ -0.0285∗∗ -0.0523∗∗∗ -0.0521∗∗∗ -0.0521∗∗∗
(0.0108) (0.0114) (0.0116) (0.0109) (0.0105) (0.0105)
MH=1 X -0.0507
Monday (0.0469)
NoMH=1 X 0.0134
Monday (0.0223)
MH=1 X -0.0266
Tuesday (0.0336)
NoMH=1 X 0.0266
Tuesday (0.0190)
MH=1 X -0.0270
Wednesday (0.0307)
NoMH=1 X -0.0598∗∗∗
Wednesday (0.0187)
MH=1 X 0.0259
Thursday (0.0334)
NoMH=1 X 0.0276
Thursday (0.0238)
MH=1 X 0.0719∗∗
Friday (0.0318)
NoMH=1 X 0.0265
Friday (0.0273)
MH=1 X 0.0580∗∗
Friday + WE (0.0274)
NoMH=1 X -0.0414
Friday + WE (0.0350)
MH=1 X 0.00896 0.0134 0.0130 0.0236 0.0342
WE (0.0391) (0.0399) (0.0401) (0.0399) (0.0400)
NoMH=1 X -0.000968 0.00402 -0.0223 0.00115 0.00101
WE (0.0326) (0.0328) (0.0328) (0.0326) (0.0325)
Observations 540,229 540,229 540,229 540,229 540,229 540,229
R2 0.648 0.648 0.648 0.648 0.648 0.648
Notes: Standard errors are clustered at the household level (* p<.10 ** p<.05 *** p<.01).
40
Table 9. Day-of-the-Week Effect: Level
Dep. Variable: Monday Tuesday Wednesday Thursday Friday Friday+WE
kWh+Nat.
Gas/month
Effect Effect Effect Effect Effect Effect
MH=1 -63.09∗∗∗ -69.09∗∗∗ -70.30∗∗∗ -79.05∗∗∗ -90.65∗∗∗ -90.68∗∗∗
(23.40) (26.01) (27.03) (26.60) (27.80) (27.80)
NoMH=1 -104.8∗∗∗ -118.5∗∗∗ -99.46∗∗∗ -112.6∗∗∗ -103.3∗∗∗ -103.4∗∗∗
(13.50) (14.37) (14.79) (14.01) (13.46) (13.46)
MH=1 X -59.26
Monday (82.87)
NoMH=1 X -20.63
Monday (32.09)
MH=1 X 93.50∗∗ 99.45∗∗ 100.9∗∗ 109.5∗∗ 120.9∗∗
Tuesday (46.70) (48.65) (49.42) (49.09) (49.90)
NoMH=1 X -61.86 -48.22 -66.97 -53.98 -63.52
Tuesday (58.77) (58.96) (58.97) (58.78) (58.67)
MH=1 X -27.53
Wednesday (60.69)
NoMH=1 X 44.84
Wednesday (27.67)
MH=1 X -18.93
Thursday (50.94)
NoMH=1 X -27.49
Thursday (27.61)
MH=1 X 24.42
Friday (54.46)
NoMH=1 X 29.64
Friday (32.15)
MH=1 X 74.34
Friday + WE (46.29)
NoMH=1 X -29.71
Friday + WE (37.15)
MH=1 X 91.75∗∗
WE (39.67)
NoMH=1 X -133.0∗∗∗
WE (51.41)
Constant 1,565∗∗∗ 1,565∗∗∗ 1,565∗∗∗ 1,565∗∗∗ 1,565∗∗∗ 1,565∗∗∗
(7.881) (7.885) (7.884) (7.881) (7.883) (7.883)
Observations 692,879 692,879 692,879 692,879 692,879 692,879
R-squared 0.679 0.679 0.679 0.679 0.679 0.679
Notes: Standard errors are clustered at the household level (* p<.10 ** p<.05 *** p<.01).