∗
This paper shows that firms respond strategically to ENERGY STAR, a voluntary certification program for energy efficient products. Firms offer products that bunch at the certification requirement and charge a price premium on certified products. The second part of the paper develops and estimates a model of imperfect competition for the refrigerator market where firms optimize energy efficiency and prices. Policy simulations suggest that ENERGY STAR brings large welfare gains, but most of the gains come from firms’ profits. Firms exploit the certification to second-degree price discriminate and thus maintain higher markups on both certified and non-certified products.
JEL: D43, L13, L15, L68, Q48.
Keywords: Environmental certification, firm behavior, energy efficiency, imperfect competition.
∗ University of Maryland, e-mail: shoude@umd.edu.
I would like to thank James Sweeney, Wesley Hartmann, and Jonathan Levin for their valuable input throughout this project. I would also like to thank James Sallee for numerous suggestions. Thank you also to Christopher Timmins, Lucas Davis, Ginger Jin, Anna Spurlock, David Rapson, Katrina Jessoe, Kyle
Meng, Antonio Bento, Erich Muehlegger, and Shanjun Li, in addition of numerous seminar participants.
Daniel Werner provided excellent research assistance. Funding was in part provided by the Precourt Energy
Efficiency Center at Stanford and the Social Sciences and Humanities Research Council of Canada.

2
1.
In an effort to push consumers toward environmentally friendly products, governmental entities, non-profit organizations, and trade associations have inundated marketplaces with environmental certifications and ecolabels. When information about product attributes is hard to collect and process, or is shrouded, certifications are justified if they correct informational market failures. In practice, however, ecolabels tend to provide a limited amount of hard information, and resemble simple branding strategies. Even so, they may improve social welfare if they ultimately lead to reductions in negative externalities.
Complications arise when firms strategically respond to certification. If firms distort product design and prices to take advantage of a certification, this may have uncertain welfare effects. Theoretical work have shown (e.g., Amacher, Koskela, and Ollikainen 2004; Mason 2011) that when informational problems, externalities, and firm behavior interact, it is unclear whether environmental certifications improve welfare. Moreover, when ecolabels contribute to the creation of green markets, they may even have the unintended consequence of increasing environmental externalities
(Kotchen 2006). In practice, it thus remains mostly unknown whether environmental certification is a desirable policy tool (Mason 2013). More generally, empirical welfare analysis of certification programs have also remained surprisingly scarce (Dranove and Jin 2010). This paper aims to fill these gaps. I conduct a welfare analysis of the ENERGY STAR (ES) program with an emphasis on firm strategic behavior. The US Environmental Protection Agency (EPA) manages ES, one of the most well-know certifications in the US, with the goal to favor the adoption of energy efficient products by providing simple information to consumers.
The analysis proceeds in three steps. Using data from the US appliance market, I first show that
ES influences the provision of energy efficiency and prices. Firms tend to offer products that bunch at the certification requirement and charge a price premium on certified models. I then rationalize the stylized facts with a model of imperfect competition, and carry out a structural estimation with data from the US refrigerator market. This allows me to simulate the market with and without certification, demonstrate the distortions at play, and quantify the welfare effects associated with
ES.
I find that ES leads to energy savings, and brings significant welfare gains of the order of
$920 million annually for the US refrigerator market alone. This corresponds to an average of
$102 per refrigerator sold. The bulk of the gains comes from firms’ profits. Firms benefit from the certification because some consumers have a high willingness to pay for the label, which allows firms

3 to differentiate products in the energy dimension. In particular, the certification facilitates seconddegree price discrimination, which leads to higher markups on both certified and non-certified models relative to a market without certification. Consumers, on the other hand, would be almost as well off in a market without certification. When ES is not in effect, firms offer products that just meet the federal minimum energy efficiency standards, with only a few products exceeding the minimum. Products are therefore less differentiated, which increases competition, and reduces markups.
These results suggest that if the EPA were to stop managing ES, firms would respond by creating their own certification program. However, because consumers value the label itself highly, irrespective of the energy savings associated with certified products, firms would have an incentive to set a weak certification requirement.
The policy analysis also shows that firms would prefer a coarse certification accompanied by a label that acts as a strong brand to the provision of information that would make consumers fully informed about energy costs. New regulations that would require providing more information in the markets for energy-intensive durables could then face an uphill battle as they would lead to competing labeling schemes, which could have adverse effects (Fischer and Lyon 2012).
Although ES is one of the main policies used in the US to manage energy demand, this paper is the first to conduct a comprehensive welfare analysis of the program accounting for firm behavior.
1
Prior work on ES has focused on estimating how consumers value the certification (Houde 2014;
Ward, Clark, Jensen, Yen, and Russell 2011). These studies show that consumers’ willingness to pay for ES is large, and may even exceed the monetary value of the energy savings associated with certified products. An important conclusion from my analysis is that a high willingness to pay for the certification may facilitate second-degree price discrimination, and enable firms to extract more consumer surplus.
This paper complements a large body of work on instrument choice for environmental policy. My work is closely related to recent studies that have compared standards to market-based instruments in the car market accounting for firms’ strategic response (Holland, Hughes, and Knittel 2009;
Klier and Linn 2012; Jacobsen 2013; Whitefoot, Fowlie, and Skerlos 2011; Fischer 2010). The general consensus from these studies is that mandatory minimum standards reduce profits and are
1 Allcott and Sweeney (2014) recently conducted a large field experiment investigating how the ES certification performs when appliances are sold over the phone, and investigate the behavior of sales agents.

4 dominated by market-based instruments. The present paper focuses on a different market, but more importantly on a different use of standards. The fact that ES acts as a voluntary standard and induces innovation beyond a minimum standard is an important determinant of the welfare gains.
Finally, this paper also contributes to the estimation of games of strategic interactions with discrete-continuous strategy space. In the present context, product line decisions with respect to energy efficiency have a discrete component because of the binary nature of the ES certification, which complicates the identification and estimation of firms’ costs. I propose an estimator based on profit inequalities (Pakes 2010; Pakes, Porter, Ho, and Ishii 2011) to recover the primitives of the cost function. Prior applications of moment inequalities in industrial organization have been primarily for models of strategic entry (Holmes 2011; Ellickson, Houghton, and Timmins 2013) and auctions (Fox and Bajari 2013). The present paper proposes a novel, but natural application of such method. Using proprietary data on costs, I find that the structural estimator performs well.
The remainder of the paper is organized as follows. The next section discusses the rationale for product certification in the appliance market. Section II presents stylized facts demonstrating that firms respond to the ES certification. Sections III-V develop and estimate an oligopoly model of the US refrigerator market. Policy analysis follows in Section VI.
2.
The ES program was first established in 1992 by the US EPA. Historically, the program has been justified by two interacting market failures: consumers’ undervaluation of energy efficiency and negative externalities associated with energy use. The certification aims to address these market failures by providing simple information that identifies the most energy efficient products. As of
2013, more than sixty categories of products were covered by the program. Products that meet or exceed the ES requirement have the right to be marketed with the ES label, which is a brand logo that does not contain technical information (Figure 1(a)). For several energy-intensive durables, including most appliances, detailed energy information is also provided to consumers by the EnergyGuide label (Figure 1(b)), a mandatory program managed by the Federal Trade Commission.
This means that all appliance models offered in the marketplace display the EnergyGuide label prominently, but only the most efficient have the ES label. For certified products, ES is thus a coarse summary of readily available information.

5
From the consumer perspective, ES can be desirable in this context for three reasons. First, it provides a simple heuristic to account for energy efficiency. Formally, I consider that the certification has the ability to inform, albeit imperfectly, consumers about energy costs, while reducing their burden to process energy information. This explains why some consumers prefer to rely on the certification although more accurate information is available. Second, the certification may affect preferences directly by providing warm glow, conformity with social norms, or enacting purely altruistic motives. Third, in many regions of the US, consumers can claim rebates offered by utilities and local governments targeting specifically ES products.
There are good reasons to believe that imperfect competition is at play in the US appliance market. In the last decades, there have been notable mergers and acquisitions, which led to an increase in market concentration. The market for large appliances is now dominated by three manufacturers: Electrolux, Whirlpool, and General Electric. As of 2008, these three manufacturers had more than 80% of the market shares for refrigerators, dishwashers, and clothes washers (Table
1).
In imperfectly competitive markets, the welfare effects of an environmental certification are uncertain. First, firms can exploit the fact that some consumers have a high willingness to pay for
ES. They can thus engage in second-degree price discrimination, and benefit from the certification.
If ES were not in effect, firms would still have an incentive to distort the provision of energy efficiency, which could result in too little or too much investment in energy efficiency (Amacher,
Koskela, and Ollikainen 2004; Fischer 2010). A certification introduces distortions of its own, but could still lead to better market outcomes by increasing the provision of energy efficiency. As I will show, a coarse certification can also have the opposite effect, and crowd-out the provision of highly energy efficient products. In sum, whether society is better off with or without the ES program crucially depends on how firms respond to the certification.
3.
This section provides evidence that firms respond strategically to ES: they tend to offer appliance models that just meet the certification requirement and charge a price premium on products with the ES label. The main data used for the analysis are monthly national sales for four appliance types: refrigerator, air conditioner, clothes washer, and dishwasher. The data were provided by
NPD Group, a marketing firm that collected and aggregated transaction data from a large number of appliance retailers operating in the US. The sample consists of an unbalanced panel of appliance

6 models. In addition to the average monthly price, the data contain monthly sales and the main attributes of each appliance model. Ashenfelter, Hosken, and Weinberg (2013) and Spurlock (2013) use the same data and find evidence of price discrimination in the US appliance market. The present analysis complements these findings and shows how ES enables price discrimination in this market.
Figure 2 shows the empirical distribution of the energy efficiency offered (non-sales weighted) for each appliance type. Because of the nature of the NPD data, the choice set in a given year reflects both inventory and manufacturing decisions.
Energy efficiency is measured as the percentage reduction between the electricity consumption of each product offered and the federal minimum energy efficiency standard. For all these appliances, the ES certification requirement is set relative to the minimum standard, which is identified by the vertical line.
For refrigerators, the ES requirement was revised in April 2008, and the revision was announced exactly one year in advance. Before and after the revision, products bunch at the minimum standard, but a larger share of products just meet the ES requirement. Firms also appear to have the ability to adjust product lines quickly–not only they offered new models that met the revised requirement the same year it was announced, but they also discontinued models that were decertified soon after the revision. Finally, firms also offer a small share of highly energy efficient models that exceed the ES requirement.
The patterns are similar for air conditioners. Products bunch at the minimum standard, most products just meet the ES standard, and a few product exceed the certification requirement. There were no revisions in the requirement from 2005 to 2010, except in November 2005 when reverse cycle models, a particular type of air conditioner that can both heat and cool, were allowed to be covered by the ES program. As a result, a small number of models in the sample (N=6, Table 2) earned the ES certification at the end of 2005.
For dishwashers, products tend to bunch around the ES requirement, but the patterns are more idiosyncratic. This can be first explain by the fact that the minimum standard and ES requirement for dishwashers are defined by a combination of energy and water factors, which are inversely correlated. It is thus more difficult to manufacture a precise level of energy efficiency.
This appliance category was also subject to frequent revisions in the ES requirement.
It was revised (effective date) in January 2007, August 2009, and July 2011. The minimum standard was

7 also revised in January 2010. The revised standard relied on a different approach to compute the energy factor, which explains the important difference in the distribution between 2010 and other years.
2 Prior to 2010, the fact that the minimum standard had been in place for a long time could explain why the minimum standard was not binding. The cumulative effect of small innovations throughout the years may have enabled manufacturers to increase efficiency beyond the minimum required.
For clothes washers, new ES requirements became effective in January 2007, July 2009, and
January 2011. The revisions in 2007 and 2011 also coincided with changes in the minimum standard.
Figure 2 shows that for all years a large share of products just meet the minimum standard, with little bunching at the ES standard, especially in the most recent years. Instead, the distributions tend to be bimodal with a large share of products that just meet the minimum, and the remaining exceed the ES standard. It should be noted that the minimum standard and ES requirement for this appliance category are also defined by a combination of energy and water factors.
If some consumers value highly the ES label, firms should set prices above marginal costs and extract part of the willingness to pay associated with ES. I next show that this is the case. In particular, firms charge different prices for the exact same appliance models with and without the ES label. To measure the price premium associated with the ES certification, I first exploit the various revisions in the certification requirements that occurred during the 2005-2011 period.
Following a revision, the EPA requires that products that do not meet the new more stringent requirement be decertified.
3 This implies that the exact same appliance model can be found on the market with and without the ES label.
My empirical strategy is a simple difference-in-difference estimator that takes the following form:
(1) log ( P jt
) = ρEST AR jt
+ α t
+ γ j
+ βX jt
+ jt where α t and γ j are month-year and product fixed effects, respectively. The dependant variable is the log of the monthly price for appliance model j .
EST AR jt is a dummy variable that takes a
2 The previous minimum standard, effective in 1994, relied on an energy factor that was a function of capacity and energy consumption. Since 2010, it is simply set as a function of total energy consumption
(kWh/year).
3 In particular, the EPA requires that ES appliances that do not meet the new requirement have their ES logo removed or masked.

8 value one if product j is certified in month t , and zero otherwise. The dummy variable EST AR jt only varies for products that lose their certification.
X jt is a matrix with additional controls. In one specification, I control for the number of months a product has been on the market, which is a proxy for product age. I also consider linear time trends, and one-year pre-decertification linear time trends that can differ for decertified models and other models. The estimation is performed separately for the four appliance categories.
The coefficient ρ is the quantity of interest, and estimates the average impact of the ES label for a given appliance type. Because of the small number of decertified models for some events (Table 2), ρ does not vary across events.
Table 3 shows that the prices of decertified models decrease after a revision. For all three appliances categories with a least one decertification event, the decrease is statistically significant
(standard errors are clustered at the product level), and ranges from to 4.4-5.0%, 8.2%, and 7.6-
8.7%, for refrigerators, clothes washers, and dishwashers, respectively. Evaluated at the mean price observed in 2008 (Table 2), the premium is $69 (refrigerator), $58 (clothes washers), and
$49 (dishwasher). Controlling for pre-decertification linear time trends does not impact the results significantly, and I fail to reject that the pre-time trends are the same for decertified and nondecertified models.
4
The effect of product age is negative and significant–as products stay longer on the market the price tends to decrease. This is consistent with long-term inventory management.
As newer models enter the market, manufacturers and retailers may want to liquidate decertified models to free up inventory. The fact that the prices of decertified models decrease even after controlling for product age means that the estimates are not simply capturing end-of-life sales. As
I show next, even when the effect of inventories should be expected to have the opposite effect or can be completely ruled out, firms still set different prices for the exact same appliance models with and without the ES label.
First, consider the market for air conditioners. As mentioned earlier, the ES program was expanded in November 2005 to cover a particular type of AC models that were previously not eligible for the certification. In my sample, I thus observe a small number of AC models that earned the ES certification. In 2005-2006, firms were then presumably stocking up on these models, inventories should then have been increasing. Using the same strategy as before, I can then compare the price of these models before and after the certification, and use all other AC models as a counterfactual.
I find that earning the certification leads to a price of increase of 6.0% to 7.3% (Table 3), which
4
A pre-decertification linear time trend is specified for each decertification event for the 12 months preceding the revision. The appliance categories subject to three revisions have then 3 × 2 pre-decertification linear time trends estimated.

9 closely mirrors the effect of the decertification events. These estimates are, however, marginally significant, which is not surprising given that only six models in the sample earned the certification.
Second, I consider a decertification event that was completely unexpected by both manufacturers and retailers, and thus could have not impacted inventories and product lines. In January 2010, the EPA found that a number of ES refrigerator models (N=21) had undergone problematic testing procedures. As a result, their electricity consumption was underestimated. The EPA concluded that these refrigerator models were not meeting the ES requirement, and issued a public statement that they should be decertified immediately. Using a second data source provided by a large US appliance retailer, I observe the manufacturer suggested retail price and transaction price for 16 of these decertified models, in addition of hundred of other refrigerator models during the period 2009-
2010. Figure 3 shows the average normalized MSRP and promotional weekly prices for refrigerator models that lost their certification, and for models that were ES certified as of January 1, 2009.
The figure clearly shows that manufacturers responded by decreasing the MSRP of decertified models immediately following the EPA announcement. This sharp decrease is also reflected in the promotional prices set by the retailer. The difference-in-difference estimator (Online Appendix) suggests that the MSRP decreased by 4.5%, and the promotional prices decreased by 1.2%.
5
Altogether, the stylized facts show that firms operating in the appliance market are aware of the ES certification, believe that consumers value it, and optimize their product lines and prices consequently. Exploiting various decertification and certification events, I obtain estimates of a price premium associated with the ES certification that are highly consistent with each others.
In the US appliance market, firms have thus the ability to extract part of the consumer surplus associated with the high willingness to pay for the ES label by price discriminating.
The certification tends to cause information unravelling (Dranove and Jin 2010; Grossman 1981); as time passes more products become certified and thus ES becomes less informative. For some products, firms, however, maintain a clear bunching at the minimum and ES standards, suggesting
5
One caveat with the interpretation of the 2010 event is that decertified models not only lost the ES label, but were also required to have a new EnergyGuide label with accurate energy information. The change in labeling is then confounded with a change in technical energy information. The change in prices captures both consumers’ valuation of ES and energy costs. For the present purpose, these estimates are still relevant.
It shows that in this market firms charge prices above marginal costs, and rely on energy related information to set higher markups.

10 that such separating equilibrium is stable. When unraveling occurs, as for clothes washers, interestingly firms still differentiate their products in the energy dimension, suggesting that they believe that a share of consumers may value energy efficiency beyond the ES certification.
Why in some markets products bunch perfectly at the ES requirement but bunch less in others?
It is unlikely that consumer preferences differ substantially between these markets. On the other hand, the technologies to meet the ES requirement do. For dishwashers and clothes washers, firms need to optimize water and energy efficiency, which could explain that less than perfect bunching is observed. Furthermore, these products have been subject to frequent revisions in the minimum standards during the period 2005-2011, which may have driven more innovation. In the face of rapidly changing technologies, manufacturers may have more difficulty precisely optimizing energy efficiency.
4.
Two necessary conditions to rationalize bunching at the minimum standard and ES requirement, together with a price premium on certified products are: (1) firms have the ability to exercise some market power, and (2) consumers’ valuation of energy efficiency is heterogeneous. In this section, I develop a model of the appliance market that incorporates these features. In subsequent sections,
I estimate and simulate the model to illustrate how ES distorts firms’ decisions.
I characterize the appliance market with a static model of imperfect competition where firms strategically determine the energy efficiency level and the price of each product they offer. The model aims to represent a medium-run equilibrium where firms adjust to a policy change within 12 to 18 months. The decisions to enter and exit the market, to determine the size of product lines, and the quality of non-energy attributes are taken as given. My approach closely follows Jacobsen
(2013), Klier and Linn (2012), and Whitefoot, Fowlie, and Skerlos (2011), among others, who model car manufacturers’ product design decisions in response to fuel economy standards. The present model, however, has a different purpose and looks at the role of voluntary standards. Moreover, a number features specific to the appliance market motivate the following modeling assumptions.
Assumption 1: Firms Are Brand Managers
An important feature of the appliance market is that manufacturers offer similar products under different brand names. To circumvent the difficulties associated with modeling the manufacturers’ decision to rebrand products (e.g., product cannibalization and double marginalization), I will focus on modeling the behavior of brand managers. I will assume that each brand manager represents

11 a firm that maximizes the profits of his own brand. In this context, a product line decision consists of acquiring appliance models through procurement contracts with manufacturers; and brand managers’ product unit costs are simply the prices they pay to manufacturers (wholesale prices).
The model endogenizes markups set by brand managers, but not the manufacturers’ markups.
Assumption 2: Cost of Providing Energy Efficiency Is Separable
A second important feature of the appliance market is that within a relatively short time firms can change the energy efficiency level of their products, with little impact on their overall design.
This has been demonstrated during the various revisions in the ES requirement; manufacturers systematically managed to offer new products that were more energy efficient, but were otherwise similar to previous generations.
6
I take this as evidence that the cost of providing energy efficiency is separable from the cost of providing other attributes. I will further assume that the unit costs faced by brand managers, i.e., the wholesale prices, reflect this assumption.
Under these assumptions, consider that there are K brands, and brand manager k offers J k appliance models. Brand manager k maximizes profits by choosing the energy efficient levels, the vector f k
= { f k 1
, ..., f kJ k
} , and the prices, the vector p k
= { p k 1
, ..., p kJ k
} , of his J k models, taking the actions of rival firms as given. Firms face a population of heterogeneous consumers in which the demand for each product is Q kj
( f, p ), and depends on all energy efficiency levels and prices
( f = { f
1
, ..., f
K
} and p = { p
1
, ..., p
K
} ). The problem of brand manager k consists in solving: max f p k k
= { f k 1
= { p k 1
,...,f kJk
,...,p kJk
} ,
}
=
J k
X
( p kj j =1
− c kj
( f kj
)) · Q kj
( f, p ) s.t.
f ≥ f where c kj
( f kj
) is the unit cost of model j offered by brand k , and f is a vector of the minimum energy efficiency standards that each product must meet.
The demand model is borrowed from Houde (2014) and aims to capture the different mechanisms by which ES influences consumers. The purchase decision is modeled as a two-step process where
6
Interestingly, when the EPA announced in April 2007 that the ES requirement for refrigerators would be revised in April 2008, all but one manufacturer notified the EPA that they would not be able to offer new refrigerator models on time to meet the more stringent requirement. Ultimately, most manufacturers were, however, able to offer new models meeting the 2008 requirement within a year of the date that the EPA made the announcement.

12 consumers first update their beliefs about energy costs by collecting and processing different pieces of information, and then make a purchase decision. In this framework, consumers are heterogeneous in the costs of collecting and processing energy information, which leads to heterogeneity in the way consumers value energy efficiency. The existence of these costs rationalizes why some consumers either dismiss the energy efficiency attribute or rely on ES, although accurate information about energy costs is readily available via the EnergyGuide label.
7
The decision to collect and process energy information is treated as a latent decision unobserved by the econometrician. The choice model takes the following general form:
(2)
X
Q j
= e = { U,ES,I }
H ( e ) · P e
( j ) where e represents the level of knowledge about energy costs that each consumer acquires. Consumers fall into three mutually exclusive categories. They can be uniformed ( e = U ). In such case, they will not know the energy cost of each product and the meaning of the ES certification. They can be knowledgeable about ES ( e = ES ), but not about the exact energy cost of each product.
Finally, they can be fully informed ( e = I ) and know the energy cost of each product in their choice set.
H ( e ) is the probability that a consumer acquires knowledge e , and P e
( j ) is the choice probability conditional on the level of knowledge.
Q j is then the choice probability for product j .
8
In the empirical application, the choice probabilities are computed for each household i , and are region and time specific, which are denoted by r and t , respectively. The alternative-specific utilities that enter the conditional choice probabilities P e irt
( j ) for each type e are: e=I: U
I ijrt
= − ηP jrt
+ ψR rt
XD jt
+ τ
I
D jt
− θC jr
+ δ j
+
I ijrt e=ES: U
ES ijrt
= − ηP jrt
+ ψR rt
XD jt
+ τ
ES
D jt
− θESavings r
XD jt
+ δ j
+
ES ijrt e=U: U
U ijrt
= − ηP jrt
+ γ j
+
U ijrt
, where δ j is the quality of the product, P jrt is the price, D jt takes the value one if product j is certified ES at time t and zero otherwise, R rt is the rebate amount offered for ES products, and ijrt is an idiosyncratic taste parameter. Assuming that the idiosyncratic taste parameters extreme value distributed, the probabilities P e irt take the form of a multinomial logit.
are
7
Sallee (2013) develops a similar model and applies it to the car market to explain rational inattention to the energy efficiency attribute.
8
The subscript k that denotes that product j is offered by firm k is omitted to simplify the notation.

13
In this framework, there are three mechanisms that lead to the adoption of ES products. First, there are the expected energy savings, i.e., the fact that ES products have lower energy costs on average. When e = I , consumers can perfectly forecast the energy operating costs of each product.
The variable C jr is the product of the annual energy consumption of refrigerator j and the average energy price in region r . When e = ES , the term ESavings r is the difference between the average energy operating costs of ES and non-ES products in region r . For these consumers, the certification serves as a heuristic; instead of computing the exact difference in energy costs, they rely on a proxy that reflects the average savings associated with the certification. For both e = I and e = ES , the parameter θ captures the sensitivity to energy operating costs, and thus determines how the monetary value of the energy savings influences choices.
The second mechanism is the effect of the label itself, which is captured by the parameters τ
I and τ
ES
. The label effect rationalizes a number of behavioral effects, which cannot be identified separately in the data. It may represent a true preference parameter that captures warm glow, social norms, or altruistic motives. It is also possible that it captures biases on how consumers perceived certified products. The label effect is differentiated for e = I and e = ES because different levels of consumer sophistication may led to a different perception of the label.
The third mechanism is the effect of rebates, which provide an explicit monetary incentive to purchase ES products. This is captured by the parameter ψ .
In Houde (2014), I develop a fully structural approach to model the latent probabilities H irt
( e ).
For the present application, they are specified as follows:
(3) e V irt
( e )
H irt
( e ) =
P k e V irt
( k ) with
(4) V irt
( e = I ) = − K
I
− β
F
X i
+ γ
I
1
M eanElec rt
+
γ
I
2
V arElec rt
+ γ
I
3
N bM odels rt
V irt
( e = ES ) = − K
ES
− β
ES
X i
+ γ
ES
1
M eanES rt
+ γ
ES
2
V arES rt
+
γ
ES
3
N bM odels rt
V irt
( e = U ) = 0 where K e is a constant, and X i is a vector of consumer demographics. The other variables aim to capture factors that could influence consumers’ decision to collect energy information, and are specific to the choice set faced by each consumer.
M eanElec rt and V arElec rt are the mean and

14 variance in electricity costs for all products offered in region r at time t .
M eanES rt and V arES rt are the mean and variance of the proportion of ES models offered. Finally, N bM odels rt is the number of models in the choice set in a given region.
The Nash equilibrium of the game is given by the vectors f
∗ and p
∗ that solves a following system of 3 × ( J
1
+ J
2
+ ...J
K
) equations. For each firm k , the first order conditions are:
(5)
(6)
(7)
(8)
Q kl
( f
∗
, p
∗
) +
J k
X
( p
∗ kj j =1
− c kj
( f
∗ kj
)) ·
∂Q kj
( f
∗
∂p
∗ kl
, p
∗
)
1 { π ( f kl
, f
∗ k, − l
, p
∗ k
) > π ( f
ES kl
, f
∗ k, − l
, p
∗ k
) |∀ f kl
}×
= 0 ,


Q kl
( f
∗
, p
∗
) dc kl
( f
∗ kl
) df kl
−
J k
X
( p
∗ kj j =1
− c kj
( f
∗ kj
)) ·
∂Q kj
( f
∗
, p
∗
)
∂f kl

= 0

,
1 { π ( f kl
, f
∗ k, − l
, p
∗ k
) ≤ π ( f
ES
, f
∗ k, − l
, p
∗ k
) |∀ f kl
}× h f
∗ kl
= f
ES i
, for all l ∈ J k and k f kl
≥ f for all l ∈ J k and k
In the second and third first order conditions, the indicator function arises because the demand function is not continuous at f
ES
; the derivative of the profits with respect to energy efficiency level f kl is then not defined at this point. The discontinuity at the certification requirement implies that it may not be optimal to equate the marginal cost of providing energy efficiency with the marginal valuation. In the presence of ES, firms’ strategies then become a discrete-continuous choice where firms decide whether or not to bunch at the certification requirement and which price to set. The existence and uniqueness of an equilibrium in this game are not guaranteed.
Note that the equilibrium outcomes in the presence of certification is rather uninformative about what would happen in a market not subject to certification. To illustrate, consider the monopoly

15 case.
Under certification, the monopolist faces three consumer types: consumers that dismiss energy efficiency, consumers that rely on ES, and perfectly informed consumers. A separating equilibrium with three different products may exist where one product would just meet the minimum standard, one product would bunch at ES, and one product, the one offered to the perfectly informed consumers, would be located at the point where the marginal cost of providing energy efficiency equals the marginal valuation. This energy efficiency level could be below or above the certification requirement. Everything else being equal, the smaller the fraction of perfectly informed consumers, the more difficult it would be for the monopolist to support this equilibrium.
In the absence of certification, the share of perfectly informed consumers may increase relative to a market with certification. In this framework, this occurs because some consumers prefer to rely on
ES instead of being perfectly informed to economize on the cost of processing energy information, but in the absence of certification they would prefer to be informed instead of remaining completely uniformed.
9
In a market without certification, the monopolist may then have a greater incentive to meet the demand of the perfectly informed consumers as they represent a larger share of the population. As a result, it may be optimal to offer a highly energy efficient product in market not subject to certification, but not in a market with certification. In sum, the coarse nature of the ES can crowd out firms’ incentive to offer highly energy efficient models. The effect of certification on the provision of energy efficiency is thus uncertain.
5.
The focus of this section is on the cost estimation, and especially the identification of the marginal cost of providing energy efficiency. I refer the readers to Houde (2014) for further details on the demand estimation.
The estimation is carried with transaction level data of the US refrigerator market for the year
2008. The data were provided by an appliance retailer with non-marginal market shares in the US appliance market. For each transaction, I observe the refrigerator model purchased, the price and taxes paid, attribute information, and the location of the store. For a large subset of transactions,
I also observe demographic information. Only transactions that can be imputed to a household are
9
In Houde (2014), I provide a formal proof of this result.

16 considered. The MSRP and promotional prices are provided for each transaction, which allows me to construct price time series that are model and region specific.
The retailer also provided attribute information for all refrigerator models that were offered at the store between 2007 and 2011. In addition of product characteristics, I observe the manufacturer’s suggested retail price (MSRP) and the wholesale price for all these refrigerator models.
The retailer data are complemented with data on electricity prices (Energy Information Administration, Form EIA-861), and rebates (DSIRE database). Data from the Federal Trade Commission and the EPA were also used to determine which refrigerator models were on the market during the period 2007-2011.
The goal of the cost estimation is to identify the unit cost of each product offered and the marginal cost of providing energy efficiency. For the estimation, I make a functional form assumption and define the unit cost of product l offered by brand manager k as follows:
(9) c ( f kl
) = e
( α kl
+ φf kl
)
I estimate α kl
, a constant specific to each product, and the parameter φ , which determines how the unit cost varies as a function of energy efficiency. Energy efficiency is defined as the inverse of energy consumption. This is a convenient modeling assumption that allows me to link supply and demand; the energy costs faced by consumers are simply the inverse of the energy efficiency levels set by firms multiplied by the energy price. Under this functional form, the costs are increasing and convex with respect to energy efficiency if φ > 0.
I also assume that there are five brand managers operating in the US refrigerator market: the four main brands, and a generic brand that includes all other brands. Brand managers compete by placing their products at appliance stores. To reduce the dimensionality of the problem, I model the game for only one appliance store, which is representative of the US refrigerator market. The choice set is fixed at 78, and the number of refrigerators offered by each firm is held constant.
10
To ensure that the choice set is representative of the US market, the distribution of the 78 products
10
The size of the choice set corresponds approximately to the number of models offered in a store in my sample. In my sample, the average number of refrigerator models offered by a store is 99 (Table 2). I set the size of the choice set to 78 for computational reasons. Although I was able to solve the model for larger choice sets, the simulation results were more robust and converged faster with smaller choice sets. Qualitatively, the results were the same with a larger choice set.

17 in terms of brand, style, size, and energy efficiency was selected to fit the distribution observed nationally in the year 2010
11
. To illustrate, suppose that 5% of the full-size refrigerators available on the US market in 2010 were GE top-freezer refrigerators with a size between 12 cu.ft. and 21 cu.ft., and certified ES. I sample 4 products in my sample (5% X 78 ≈ 4%) that fit this description.
In the presence of ES, the provision of energy efficiency is a location decision with a discrete choice component: firms must decide if each product must meet or not the certification requirement.
Because of the discrete nature of the decision, point identification of the marginal cost of providing energy efficiency is not possible when all products bunch exclusively at the minimum and ES standards. The first-order conditions with respect to energy efficiency point identify the marginal cost only when some products do not bunch at either of the standards. When such a condition is not met, set identification is however still possible using profit inequalities (Pakes 2010; Pakes,
Porter, Ho, and Ishii 2011). My empirical strategy for the estimation of the parameter φ then consists of using moment inequalities combined with additional moments provided by a subset of the first order conditions with respect to energy efficiency.
For the moment inequalities estimator, I follow Fox and Bajari (2013); Ellickson, Houghton, and Timmins (2013) and use profit inequalities with the pairwise maximum-score approach proposed by Fox (2007). Put simply, the approach consists of creating small perturbations in the observed strategies of each firm, holding constant the strategies of other firms, and make a pairwise comparison between observed and counterfactual profits computed at each perturbation. The maximum-score objective function simply rewards the value of φ consistent with profit maximization where observed profits are greater than counterfactual profits. The estimation procedure is described in full details in the Online Appendix.
The estimation is performed by forming the individual choice probabilities for each consumer, Q ijrt
, and estimating the model via maximum likelihood. I allow heterogeneity with respect to income by estimating the model separately for three different income groups. Three large sub-samples of transactions were randomly drawn from the universe of transactions made by households that belong to a particular income group. I distinguish between households with income of less than
≥ $50,000, households with income between $50,000 and $100,000, and households with income of
11
The year 2010 was chosen because it represents a year where firms seemed to have fully adjusted to the change in certification requirement that occurred in 2008.

18 more than $100,000. Given the large number of transactions, the lost of efficiency from sub-sampling is small.
The estimation is performed with a two-step maximum likelihood procedure to ease the computation due to the large number of fixed effects. In the first step, product fixed effects are recovered by estimating a simple conditional logit, with no unobserved heterogeneity, for each income group.
The second step estimates the mixed logit model, but uses the product fixed effects from the conditional logit as data. The standard errors in the second step are corrected using the Murphy and
Topel (1985)’s approach.
The choice model does not contain an outside option. The purchase decision being modeled is conditional on the decisions to replace a refrigerator and go shopping at a particular store.
Therefore, the price coefficient corresponds to a short-run elasticity.
The following sources of variation identify the various behavioral parameters. The sample contains large and frequent variations in prices. I exploit the temporal variation in prices to identify consumers’ sensitivity to the retail price of a particular model.
Following the revision of the ES standard for refrigerators in April 2008, a large number of refrigerator models lost their ES certification (Table 2). Using data that cover a time period before and after the revision in the standard, it is possible to observe the same refrigerator model being sold at the same store, with and without the ES label. This variation in labeling can then be used to identify how consumers are influenced by the label.
I observe the same refrigerator models being sold at stores located in different electric utility territories. This allows me to control for product fixed effects and to use cross-sectional variation in average electricity prices and rebates across regions to identify the sensitivity to electricity prices and rebates. Average electricity prices are state specific, and rebates are county-specific.
6.
Performing a grid search over the possible values of the parameter φ , I found that φ = 429 yields the largest value for the maximum-score estimator. At this value of φ , a 1% increase in energy efficiency for a refrigerator model consuming 550 kWh/year, increases the unit cost by 0.78%. Is this a reasonable value? In the Online Appendix, I present an alternative estimator that relies on

19 minimal structural assumptions, and does not require information about demand, and obtain an estimate of similar magnitude. With this reduced-form estimator, the cost elasticity with respect to energy efficiency is 0.66.
Table 4 presents the estimates of the demand for all three income groups. Focusing on the price coefficients, we observe an inverse correlation between consumers’ sensitivity to prices and income levels, i.e., the marginal utility of income decreases with income. Meanwhile, lower-income consumers are also less sensitive to electricity costs. To interpret the magnitude of the estimate of the sensitivity to electricity costs, we need to compare η and θ . Instead of simply reporting the ratio,
I make additional assumptions
12
Assuming a refrigerator lifetime of 18 years, the implied discount rate is 6% for households with an income larger than $100K, 11% for households with income between $50K and $100K, and
8% for households with income of less than $50K. These discount rates rationalize the purchase decision of the share of perfectly informed consumers, but do not represent aggregate demand.
The behavioral estimates of the simple conditional logit yield implied discount rates of 178%, 82% and 37% for the high, medium, and low income groups, respectively. These rates are high, which suggests that consumers, on average, undervalue energy operating costs. The estimates from the mixed logit model makes it clear that this undervaluation is due to unobserved heterogeneity in information acquisition costs, which explains the existence of a consumer type that dismisses the energy efficiency attribute altogether.
Looking at the latent probabilities, households in the lowest income group have a high probability of being uninformed ( e = U ), while households in the highest income groups are the most likely to be perfectly informed about energy costs ( e = I ). Medium and high income households have a high probability to rely on ES, but for low income households this probability is smaller. ES products are thus being purchased by the most affluent consumers, and energy efficiency tends to be dismissed altogether by households with the lowest income.
For consumers that rely on the certification ( e = ES ), the effect of the ES label is positive, relatively large, and varies across income levels. The estimate of the label effect τ
ES translates into a willingness to pay ( τ
ES
/η ) for the certification itself that ranges from $95 to $152. For the
12
I assume that consumers form time-unvarying expectations about annual electricity costs, and do not account for the effect of depreciation. See Online Appendix for further details on the calculations.

20 perfectly informed consumers, the label effect ( τ
I
) is small and even negative. This suggests that for consumers that rely on an accurate measure of energy costs, the ES products are not valued beyond the monetary value of the energy savings.
7.
The main analysis consists in simulating the US refrigerator market with and without the ES certification, and comparing welfare and other metrics. For all scenarios, the demand is simulated with a random sample of 500 households taken from the whole sample of transactions. Therefore, households differ with respect to demographic information and the region where they live. The price of electricity faced by each household is the average electricity price at the state level for the year 2010. All households have access to a $50 rebate for ES products. Unless otherwise indicated, the ES requirement is set to the level in effect in 2010 (20% relative to the minimum standard).
The policy analysis is performed by bootstrapping the model 100 times. For each bootstrap iteration, the demand parameters are sampled from their estimated distribution, and the model solves for the Nash equilibrium using the Gauss-Siedel algorithm. The results reported are the mean and standard deviation of the difference in metrics between policy scenarios.
13
The metrics considered to compute welfare are the consumer surplus, producer surplus, externality costs, and change in government revenue.
Consumer Surplus.
In the demand model, there is a discrepancy between the notions of decision and experience utility, because consumers that rely on the ES certification ( e = ES ) or remain uninformed ( e = U ) have imperfect knowledge about electricity costs. Therefore, the utility experienced after the purchase decision may not be the same as the one anticipated at the time of purchase.
14
To measure the change in consumer surplus, I simply assume that the expected utility faced by the fully informed consumer type ( e = I ) coincides with the experienced utility that each consumer will receive over the lifetime of a refrigerator. Put another way, fully informed consumers have rational and unbiased expectations. Under this assumption, the expected consumer surplus
13
That is, I report the mean of the differences, and not the difference of the means.
14
Allcott, Mullainathan, and Taubinsky (2014) refer to such discrepancy as an internality. In the present model, the costs of acquiring and collecting energy information and, possibly, the ES label are at the source of the internality.

21
( ESC irt
), for consumer i that lives in region r and makes a purchase at time t , is then given by:
(10)
1
ECS irt
=
η i
J
X X
H irt
( e ) · P e irt
( j ) · ( γ j j e
− η i
P jrt
+ ψ i
R rt
XD jt
− θ i
C jr
) ,
The term
1
η i is the inverse of the marginal utility of income of household i , and is used to convert utils into dollars. All estimated paramaters vary across income groups. The above measure corresponds to the consumer surplus obtained over the entire lifetime of a refrigerator. To obtain an annual measure, I assume that consumers will own (and believe they will own) their refrigerators for 18 years, and compute the corresponding annuity using the implied discount rate specific to each income group. Whether the label effect should be part of the consumer surplus is open to discussions. I will report the results without the label effect (parameters τ
ES and τ
I
) assuming that the label affects decisions, but not how a consumer experiences the various services provided by a refrigerator.
Producer Surplus.
The producer surplus is the sum of the profits made by each brand manager. Profits are computed by multiplying markups with the simulated market shares, where the markups are the difference between prices and unit costs. The change in manufacturers’ profits are not accounted for in the producer surplus. I conjecture that under most circumstances, the sign of the change in profits, in a given scenario, would be the same for brand managers and manufacturers. When this hold, the magnitude of the change in profits reported can be considered a lower bound.
Externality Costs.
The externality costs associated with electricity generation account for the emissions of carbon dioxide ( CO
2
), sulphur dioxide ( SO
2
), and nitrous oxide ( N O x
). The dollar damages of the externality costs under each scenario are computed by taking the product of the average electricity consumption purchased, the emission factors, and the damage costs per unit of emissions. The average electricity consumption purchased is the average of the electricity consumption of the refrigerators sold, weighted by market shares. Table 4 presents the emission factors and the damage costs. For the analysis, the high estimates of the externality costs are used, which translates to a cost of $0.079/kWh.
Government Revenues.
According to the US Government Accountability Office (GAO), the
EPA and the DOE have spent $57.4 M/year, on average, during the period 2008-2011 to run the
ES program. If we assume that more than 60 product categories are covered by the ES program, a naive estimate of the administrative costs of the program for the refrigerator market alone can

22 be obtained by assuming that the overall costs are distributed proportionally across all product categories. Under this assumption, the administrative costs would be $0.96 M/year.
To obtain an estimate of the welfare effects for the whole US refrigerator market, I scale all of the above measures by the market size. I use the annual shipments of refrigerators in the US for the year 2010, which is 9.01 million units (DOE 2011).
Before presenting the welfare analysis, I first demonstrate that the model has good internal validity.
Figure 4(a) shows the distribution in energy efficiency predicted by the model under ES. Compared to the distribution observed in 2010 for the US refrigerator market (Figure 2), the model replicates well the bunching at the minimum and ES standards, although the observed distribution is more idiosyncratic. These discrepancies can be explained by the fact that the model is static, while in practice product line decisions and the certification requirement are determined dynamically.
Clearly, products located between the minimum and ES standards correspond to models that met the previous certification requirements, but did not exit the market. These previous requirements and the anticipation effect of future revisions are not captured by the model.
By strategically determining energy efficiency levels, firms differentiate their products and can charge prices above marginal costs. The ability of the model to replicate pricing strategies is crucial for the welfare analysis. Figure 5(a) compares the markups predicted by the model and the observed markups.
15
For proprietary reasons, the markups are normalized. Each normalized markup corresponds to the ratio of the percentage markup of a product and the lowest percentage markup observed in the retailer data. Qualitatively, the patterns are similar. Predicted markups are, however, higher on average, which suggests that the model tends to overestimate the degree of market power. Nonetheless, the model does well in predicting which refrigerators should have the highest markups. The model thus provides a good approximation of how firms price discriminate in this market.
The main estimates of the overall effects of ES are presented in Column I of Table 5. Each metric represents the average difference in equilibrium outcomes between the scenario without certification and the scenario with certification.
15
The observed markups are the retailer prices minus wholesale prices divided by the retailer prices.

23
Focusing on energy savings, the model predicts that by removing ES, the average electricity consumption of a refrigerator purchased
16 would increase by 50 kWh/year, or about 10%. Without certification, firms offer products that mostly bunch at the minimum standard, with a few products that exceed the minimum (Figure 4(b)). The most energy efficient products are, however, amongst the most popular, and attract significant market shares ( ∼ 35%). A first important conclusion is that in a market not subject to certification, energy efficiency will be provided beyond the certification requirement and sought by a fraction of consumers that value this attribute. The certification requirement thus crowds out the most energy efficient products from the marketplace.
Removing the certification leads to a large loss in profits; about $883 million per year, or $98 per refrigerator sold. A reduction of $98 in retail price is economically significant, but appears to be reasonable in this market. Considering that the average price paid for a full-size refrigerator was $1,325 in 2008 (Table 2), $98 corresponds to an average reduction in retail price of 7.4%. To further put this number in perspective, the 2010 DOE’s national impact analysis
17 of minimum standards for refrigerators assumes that the retailer markup was 37% of the final price. This percentage markup translates into a markup of $242 for top-freezer, $375 for bottom-freezer, and
$462 for side-by-side. An average markup of $98 attributable at the certification alone thus appears plausible in this market.
Figure 5(b) compares the estimated markups for each product in a market with and without certification. Markups decrease sharply for most products. An important conclusion is that the certification affects all prices in equilibrium, even products that were not certified in the first place.
Without ES, products are less differentiated in the energy efficiency dimension, which increases competition, and leads to lower markups. The fact that some consumers value highly the ES label thus allows firms to screen consumers in the energy efficiency dimension and extract consumer surplus from all consumers, even the ones that do not value the certification.
Interestingly, when ES is not in effect, the products with the highest markups are the ones with the highest energy efficiency levels (Figure 5(c)). Without certification, firms thus still distort energy efficiency levels to extract rents from consumers. Note also that the products with the highest markups are also amongst the cheapest (Figure 5(d)). On one hand, this implies that when
ES is not in effect, energy efficiency can be accessible to all income groups. On the other hand, the
16
The average electricity consumption of a refrigerator purchased is the sales-weighted average of the annual electricity consumption of all refrigerator models offered.
17
The models used by the DOE for the national impact analysis can be requested to the author.

24 fact that these products have the largest markups also suggests that the exercise of market power falls disproportionately on lower income households.
Consumers are slightly worst off without ES, but the change in consumer surplus is not significantly different than zero. Refrigerators are cheaper, by $60 on average, although less energy efficient. Therefore, the reduction in prices is large enough to compensate for the increase in lifetime electricity costs. Note that the change in consumer surplus reported in Table 5 excludes the effect of the ES label. That is, the welfare measure takes the stand that the label does not affect utility, although it impacts decisions. If the label were to truly provide a positive utility shock, removing the certification would further decrease the consumer surplus.
Accounting for the changes in externality costs, consumer surplus, and profits, welfare decreases when ES is not in effect. The bulk of the welfare loss comes from the loss in profits. This suggests that if the ES program were to be removed, firms would respond by developing their own certification. A world without ES might, therefore, not be a world without certification.
In the Online Appendix, I present various sensitivity tests.
For non-marginal variations in the cost of providing energy efficiency and the effect of the ES label in the demand model, the predictions of the model remain relatively robust (Figure 2).
Alternative Scenarios.
Columns III and IV in Table 5 present the results of a simulation that compares a scenario with ES to a scenario without certification where all firms offer refrigerators that just meet the minimum standard. This latter scenario is of particular interest because previous studies that have provided estimates of energy savings attributable to ES (e.g.,Sanchez,
Brown, Webber, and Homan 2008) made this implicit assumption. The present results show that ignoring firms’ incentive to provide energy efficiency in the absence of certification leads to grossly overestimate energy savings associated. The electricity consumption of a refrigerator purchased increases by 97 kWh/year, on average, when the certification is not in effect. This estimate is almost twice as large as the one obtained previously.
An important finding from the demand estimation is that consumers that rely on ES value certified products well beyond the value of their energy savings. In particular, some consumers appear to have a high willingness to pay for the label itself. In Columns V and VI, I show what would happen if consumers did not value the ES label itself at the moment of making a purchase decision. In particular, I compare scenarios with and without certification using the demand model with τ I = 0 and τ ES = 0. Figure 4(c) shows that without the label effect, firms would still offer products that bunch at the ES standard, but most products would just meet the minimum

25 standard. As result, the energy savings induced by the certification is much lower, only 13 kWh/year per refrigerator sold. Firms still benefit from the certification, but by much less. Removing ES reduces profits by $489 million per year, or about $54 per refrigerator sold, compared to $97 in the main scenario. This implies that the label effect alone allows firms to increase markups by $43, on average. This estimate closely matches the ES price premium identifies by the hedonic regressions
(Section II).
The coarse nature of the ES certification makes it arguably a second-best policy. Ideally, we would like to provide a labeling scheme that makes all consumers perfectly informed about energy costs. Although such policy might not exist in practice, I investigate how it would compare with
ES. Columns VII and VIII present the result of a scenario where all consumers make their purchase decision fully informed about energy efficiency to the basecase scenario with certification. This simulation is in essence an estimate of the opportunity cost associated with imperfect information in the appliance market. The first important result is that under perfect information, firms offer several highly energy efficient models (Figure 4(d)), which lead to large energy savings, 71 kWh/year per refrigerator sold, relative to a market with ES and imperfectly informed consumers. Consumers are also better off. If the regulator were to care only about externality costs and consumer surplus, there would then be a rationale for a more informative certification program. Firms are, however, still better off under ES. This suggests that even if a regulator were to provide better information, firms would have an incentive to undo this policy by responding with their own, coarser, certification.
Overall, the change in profit dominates the change in externality costs and consumer surplus. Total welfare is then still larger when the ES certification is in effect. This is surprising, but illustrates the important role played by imperfect competition in this market. In this context, both a coarse certification, and a labeling scheme that provides perfect information become a second-best policy in the presence of multiple interacting market failures.
8.
This paper develops an economic framework to perform a welfare analysis of the ES certification program accounting for firm strategic behavior. The various mechanisms by which the certification can influence consumers are also explicitly modeled. The welfare analysis offers a cautionary tale on how certification, and more generally, nudges and information-based policies should be used, especially when it comes to address environmental externalities. Historically, ES has been managed like a marketing program where a strong branding effect has been sought, and deemed a successful

26 metric. I show that consumers’ high willingness to pay for the ES label favors the adoption of more energy efficient products, and induces firms to offer more of these products. The certification thus contributes to reducing environmental externalities. However, the fact that consumers value the ES certification highly allows firms to charge higher prices. In particular, the certification facilitates second-degree price discrimination, and markups on both certified and non-certified products are larger when ES is in effect. As a result, firms benefit from the certification at the expense of the consumers.
Without certification, most products offered just meet the minimum energy efficiency standard, but firms find optimal to offer a few highly energy efficient products. This illustrates that the coarseness of the certification crowds out the provision of high levels of energy efficiency.
Firms’ ability to strategically adjust energy efficiency levels is crucial to the analysis. The existence of imperfect competition in the market for energy intensive durables implies that the provision of energy efficiency is sub-optimal, and neither a certification nor a Pigouvian instrument adequately addresses this market failure.
An important feature of the market not addressed by the oligopoly model is how the certification induces innovation. Future research should aim to understand how the certification’s incentive to screen consumers translates into dynamic product line and pricing decisions.
Allcott, H., S. Mullainathan, and D. Taubinsky
(2014): “Energy policy with externalities and internalities,” Journal of Public Economics , 112(0), 72 – 88.
Allcott, H., and R. Sweeney
(2014): “Information Disclosure through Agents: Evidence from a Field Experiment,” Working Paper 20048, National Bureau of Economic Research.
Amacher, G. S., E. Koskela, and M. Ollikainen
(2004): “Environmental quality competition and eco-labeling,” Journal of Environmental Economics and Management , 47(2), 284 – 306.
Ashenfelter, O. C., D. S. Hosken, and M. C. Weinberg
(2013): “The Price Effects of a Large Merger of Manufacturers: A Case Study of Maytag-Whirlpool,” American Economic
Journal: Economic Policy , 5(1), 239–61.
DOE
(2011): “Energy Conservation Program: Energy Conservation Standards for Residential
Refrigerators, Refrigerator-Freezers, and Freezers,” Federal Register , 76(179).

27
Dranove, D., and G. Z. Jin
(2010): “Quality Disclosure and Certification: Theory and Practice,”
Journal of Economic Literature , 48(4), 935–63.
Eichholtz, P., N. Kok, and J. M. Quigley
(2010): “Doing Well by Doing Good? Green Office
Buildings,” American Economic Review , 100(5), 2492–2509.
Ellickson, P. B., S. Houghton, and C. Timmins
(2013): “Estimating network economies in retail chains: a revealed preference approach,” The RAND Journal of Economics , 44(2), 169–193.
Fischer, C.
(2010): “Imperfect Competition, Consumer Behavior, and the Provision of Fuel
Efficiency in Light-Duty Vehicles,” RFF discussion paper 10-60, Resources for the Future.
Fischer, C., and T. P. Lyon
(2012): “Competing Environmental Labels,” Discussion paper,
University of Michigan.
Fox, J. T.
(2007): “Semiparametric estimation of multinomial discrete-choice models using a subset of choices,” The RAND Journal of Economics , 38(4), 1002–1019.
Fox, J. T., and P. Bajari
(2013): “Measuring the Efficiency of an FCC Spectrum Auction,”
American Economic Journal: Microeconomics , 5(1), 100–146.
Greenstone, M., E. Kopits, and A. Wolverton
(2011): “Estimating the Social Cost of
Carbon for Use in U.S. Federal Rulemakings: A Summary and Interpretation,” NBER Working
Papers 16913, National Bureau of Economic Research, Inc.
Grossman, S. J.
(1981): “The informational role of warranties and private disclosure about product quality,” Journal of Law and Economics , pp. 461–483.
Holland, S. P., J. E. Hughes, and C. R. Knittel
(2009): “Greenhouse Gas Reductions under Low Carbon Fuel Standards?,” American Economic Journal: Economic Policy , 1(1), pp.
106–146.
Holmes, T. J.
(2011): “The Diffusion of Wal-Mart and Economies of Density,” Econometrica ,
79(1), 253–302.
Houde, S.
(2014): “How Consumers Respond to Environmental Certification and the Value of
Energy Information,” Working Paper 20019, National Bureau of Economic Research.
Jacobsen, M. R.
(2013): “Evaluating U.S. Fuel Economy Standards in a Model with Producer and Household Heterogeneity,” American Economic Journal: Economic Policy , 5(2), 148–187.
Klier, T., and J. Linn
(2012): “New-vehicle characteristics and the cost of the Corporate Average
Fuel Economy standard,” The RAND Journal of Economics , 43(1), 186–213.

28
Kotchen, M. J.
(2006): “Green markets and private provision of public goods,” Journal of
Political Economy , 114(4), 816–834.
Manski, C. F.
(1975): “Maximum score estimation of the stochastic utility model of choice,”
Journal of Econometrics , 3(3), 205 – 228.
Mason, C.
(2011): “Eco-Labeling and Market Equilibria with Noisy Certification Tests,” Environmental and Resource Economics , 48(4), 537–560.
Mason, C. F.
(2013): “The Economics of Eco-Labeling: Theory and Empirical Implications,”
Working paper, Center for Energy and Environmental Economics.
Muller, N. Z., and R. Mendelsohn
(2012): “Efficient Pollution Regulation: Getting the Prices
Right: Corrigendum (Mortality Rate Update),” American Economic Review , 102(1).
Newell, R. G., and J. Siikam¨
(2013): “Nudging Energy Efficiency Behavior, The Role of
Information Labels,” RFF discussion paper 13-17, Resources for the Future.
Pakes, A.
(2010): “Alternative Models for Moment Inequalities,” Econometrica , 78(6), 1783–1822.
Pakes, A., J. Porter, K. Ho, and J. Ishii
(2011): “Moment Inequalities and Their Application,” Working paper, Harvard University.
Sallee, J. M.
(2013): “Rational Inattention and Energy Efficiency,” Working Paper 19545, National Bureau of Economic Research.
Sanchez, M. C., R. E. Brown, C. Webber, and G. K. Homan
(2008): “Savings Estimates for the United States Environmental Protection Agency’s ENERGY STAR Voluntary Product
Labeling Program,” Energy Policy , 36(6), 2098 – 2108.
Spurlock, C. A.
(2013): “Appliance Efficiency Standards and Price Discrimination,” Working paper, Lawrence Berkeley National Laboratory.
US Government Accountability Office (GAO)
(2011): “ENERGY STAR Providing Opportunities for Additional Review of EPA’s Decisions Could Strengthen the Program,” Report to congressional requesters, GAO-11-888.
Walls, M., K. Palmer, and T. Gerarden
(2013): “Is Energy Efficiency Capitalized into Home
Prices? Evidence from Three US Cities,” RFF discussion paper 13-18, Resources for the Future.
Ward, D. O., C. D. Clark, K. L. Jensen, S. T. Yen, and C. S. Russell
(2011): “Factors influencing willingness-to-pay for the ENERGY STAR label,” Energy Policy , 39(3), 1450 – 1458.
Whitefoot, K., M. Fowlie, and S. Skerlos
(2011): “Product Design Response to Industrial
Policy: Evaluating Fuel Economy Standards Using an Engineering Model of Endogenous Product

Design,” Haas Working Papers 214, Energy Institute at Haas.
29

30
9.
(a) ES Label (b) EnergyGuide Label

(a) Refrigerators
2005
(b) ACs
2005
(c) Dishwashers
2005
(d) Clothes Washers
2005
0 10 20 30 40 50 60
2006
0 5 10 15 20 25 30
2006
0 20 40 60 80 100
2006
0 50 100 150 200
2006
0 10 20 30 40 50 60
2007
0 5 10 15 20 25 30
2007
0 20 40 60 80 100
2007
0 50 100 150 200
2007
0 10 20 30 40 50 60
2008
0 5 10 15 20 25 30
2008
0 20 40 60 80 100
2008
0 50 100 150 200
2008
0 10 20 30 40 50 60
2009
0 5 10 15 20 25 30
2009
0 20 40 60 80 100
2009
0 50 100 150 200
2009
0 10 20 30 40 50 60
2010
0 5 10 15 20 25 30
2010
0 20 40 60 80 100
2010
0 50 100 150 200
2010
0 10 20 30 40 50 60
2011
0 5 10 15 20 25 30
2011
0 10 20 30 40 50
2011
0 50 100 150 200
0 10 20 30 40 50 60
Percentage Better Minimum Standard
0 5 10 15 20 25 30
Percentage Better Minimum Standard
0 10 20 30 40 50
Percentage Better Minimum Standard
0 50 100 150 200
Percentage Better Minimum Standard
Notes: Each panel plots the empirical density of the energy efficiency levels offered (nonsales weighted) for four appliance types. The x-axis is the percentage decrease in electricity consumption (kWh/y) relative to the federal minimum energy efficiency standards. The ES requirement is identified by the dark vertical line. Source: NPD Group.
31

32
Decertification
Energy Star Models
Decertified Energy Star Models
2009w13 2009w40
Weeks
2010w13
(a) MSRP
Energy Star Models
Decertified Energy Star Models
Decertification
2010w40
2009w13 2009w40
Weeks
2010w13 2010w40
(b) Promotional Price
Notes: Each panel displays average normalized weekly prices, with 5% confidence intervals, of refrigerators that belong to different efficiency classes. The normalized price for each model is the weekly price (MSRP or promotional) divided by its average weekly price. The average normalized price and standard errors in each efficiency class are computed by fitting a cubic spline on normalized prices. In panels a) and b), prices of the 16 models that were decertified are compared to the prices of ES models that did not lose their certification.

0 10 20 30 40
Percentage Better than Minimum Standard
50
(a) w ES, φ = 429
0 10 20 30 40
Percentage Better than Minimum Standard
50
(b) w/o ES, φ = 429
0 10 20 30 40
Percentage Better than Minimum Standard
50
(c) w ES, τ I = 0, τ ES = 0
0 10 20 30 40 50
Percentage Better than Minimum Standard
(d) w/o ES, All Informed Consumers
Notes: Empirical density of the energy efficiency offered for different scenarios. For each product, the energy efficiency level offered in equilibrium is the mode across the 100 bootstrap iterations. The ENERGY STAR requirement is set at 20% above the minimum standard.
33

34
Predicted, certified products
Predicted, non−certified products
Observed
With ENERGY STAR, certified products
With ENERGY STAR, non−certified products
Without ENERGY STAR
Products
(a) Markups, Observed vs. Predicted
Products
(b) Markups w vs. w/o ES, Predicted
With ENERGY STAR, certified products
With ENERGY STAR, non−certified products
Without ENERGY STAR
0
With ENERGY STAR, certified products
With ENERGY STAR, non−certified products
Without ENERGY STAR
10 20 30 40
Energy Efficiency: Percentage Better than Minimum Standard
50
(c) Markups vs. Energy Efficiency, Predicted
Price Rank
(d) Markups vs. Prices, Predicted
Notes: Each observation represents the normalized markup of a product included in the representative choice set. In Panel a, all percentage markups are normalized by dividing by the lowest observed markup. The observed percentage markup is the difference between the promotional and wholesale prices, divided by promotional price. Predicted markups are obtained from the simulation model. A value of one means that the percentage markup is exactly equal to the lowest percentage markup observed in the retailer data. In Panels b, c, and d, the percentage markups are normalized by dividing by the lowest markup predicted when ES is in effect. Panel b shows that predicted markups in a market without ES are smaller than in a market with ES. Panel c shows that when ES is not in effect, the three products with the largest markups are also the most energy efficient. In Panel d, lower price ranks correspond to lower prices. Panel d shows that the cheapest products have the highest percentage markups.

Dishwasher Washer Room AC
Market Share (Year 2008)
Whirlpool/Maytag
GE
49
27
64
16
13
-
Electrolux
LG
Haier
2007
2008
2009
2010
2011
18
-
-
Other: 6
Average Retail Price ($USD)
2005
2006
598
646
574
608
671
634
618
6
6
-
8
644
657
663
704
734
714
667
13
32
8
34
294
330
348
329
352
337
365
Refrigerator
33
27
23
3
6
8
1,401
1,511
1,454
1,440
1,511
1,471
1,470
Sources.
Market shares: Appliance Magazine and Department of Energy. Average prices:
NPD Group.
35

36
Decertification/Certification Events
Refrigerators 04/2008
1 Year 1 Year
Total Energy Star
Total Non Energy Star
Before After
601
290
Total Unique Models 891
Nb of Decertified Models 359
424
737
1292
Clothes Washers
Total Energy Star
Total Non Energy Star
Total Unique Models
Nb of Decertified Models
01/2007 07/2009 01/2011
1 Year 1 Year 1 Year 1 Year 1 Year 1 Year
Before After Before After Before
100 143 235 311 375
27
127
25
116
259
99
334
5
76
387
67
442
29
After
359
77
436
Dishwashers
Total Energy Star
Total Non Energy Star
Total Unique Models
Nb of Decertified Models
01/2007 08/2009 07/2011
1 Year 1 Year 1 Year 1 Year 1 Year 1/2 Year
Before After Before After Before
185 344 594 644 693
After
532
96
281
12
421
765
191
785
89
234
878
99
792
218
269
801
Air Conditioners
Total Energy Star
Total Non Energy Star
Total Unique Models
Nb Certified Models
11/2005
1 Year 1 Year
Before After
72
127
199
6
66
107
173

37
Refrigerators
(I) (II)
Dependent Variable: Log(price)
Decertified -0.044
∗∗∗
(0.016)
Certified
-0.050
∗∗∗
(0.017)
Clothes Washers
(I) (II)
-0.082
∗∗∗
-0.082
∗∗
(0.031) (0.035)
Dishwashers
(I) (II)
-0.076
∗∗∗
(0.013)
-0.087
∗∗∗
(0.015)
Linear Time Trend
Different Pre-Time Trend
Months On Market
Month-Year FEs
Product FEs
No
No
-0.003
∗∗∗
(0.0004)
Yes
Yes
Yes
Yes
Yes
Yes
No
No
-0.008
∗∗∗
(0.0006)
Yes
Yes
Yes
Yes
Yes
Yes
No
No
-0.005
∗∗∗
(0.0004)
Yes
Yes
Yes
Yes
Yes
Yes
Air Conditioners
(I) (II)
0.060
0.073
∗
(0.042) (0.041)
No
No
-0.002
∗
Yes
Yes
(0.0011)
Yes
Yes
Yes
Yes
Nb of Obs
R
2
Nb of Models
Nb of Certification Changes
77,922
0.954
4,873
359
77,922
0.954
4,873
359
20,507
0.897
955
59
20,507
0.897
955
59
36,381
0.934
2,059
319
36,381
0.935
2,059
319
Notes: Standard errors are clustered at the product level.
∗ p < 0 .
05,
∗∗ p < 0 .
01,
∗∗∗ p < 0 .
001
11,248
0.835
642
6
11,248
0.835
642
6

38
Income
< $50,000
Income
≥ $50,000 &
< $100,000
Income
≥ $100,000
Behavioral Parameters Purchase Decision:
Price (ˆ ) -0.543
∗∗∗
(0.001) -0.448
∗∗∗
ES τ
I
ES τ
ES
Rebate ( ˆ )
-0.312
∗∗∗
0.825
∗∗∗
0.282
∗∗
(0.065)
(0.123)
0.008
0.429
(0.112) 0.132
∗∗∗
∗∗
Elec. Costs (ˆ
K
K h m
) -5.197
∗∗∗
-11.116
∗∗∗
-2.368
∗∗
(0.407) -3.359
∗∗∗
(1.245) -8.124
∗∗∗
(0.846) -3.928
∗
Educ College (
Educ Graduate ( β
I
)
FamSize (
Age ( β
I
)
β
I
)
β
I
)
Political Democrat (
Political Others ( β
Educ College ( β
ES
Educ Graduate ( β
I
)
)
ES
β
)
I
)
0.155
0.205
-0.325
∗∗∗
0.185
∗∗∗
-0.560
∗
-0.490
-0.052
1.439
∗∗∗
(0.189) 0.273
(0.426) 1.941
∗∗∗
(0.060) -0.477
∗∗∗
(0.018) 0.207
∗∗∗
(0.332) -0.958
∗∗
(0.328) -1.992
∗∗∗
(0.268) 0.695
∗
(0.369) 2.477
∗∗∗
FamSize ( β
ES
)
Age ( β
ES
)
Political Democrat (
Political Others ( β
ES mean-ElecCost var-ElecCost
Nb Models ( γ I
)
Variance Price ( γ I )
Variance FE ( γ
I
)
Proportion-Estar
Nb Models ( γ ES )
Variance Price ( γ
ES
)
Variance FE ( γ
ES
)
β
)
ES
)
0.081
0.042
∗∗∗
-1.310
∗∗∗
-1.581
∗∗∗
-2.663
∗∗
-3.426
0.013
0.034
-0.197
2.667
∗∗∗
-0.090
∗∗∗
0.067
-0.089
(0.064) -0.020
(0.009) 0.061
∗∗∗
(0.332) -1.368
∗∗
(0.333) -2.041
∗∗∗
(0.969) 3.209
∗∗
(4.983) -30.029
∗∗∗
(0.008)
(0.061)
0.004
-0.185
(0.125) 0.156
(0.805) 6.817
∗∗∗
(0.017) -0.048
∗∗∗
(0.066)
(0.135)
0.245
-0.293
(0.001) -0.357
∗∗∗
(0.020) -0.064
∗
(0.044) 0.513
∗∗∗
(0.047) 0.054
(0.147) -3.861
∗∗∗
(1.634) -6.041
∗∗∗
(0.001)
(0.030)
(0.054)
(0.043)
(0.224)
(1.548)
(1.864)
(0.278)
-1.336
0.288
(0.469) 0.578
(0.077) -0.180
∗∗
(0.020) 0.166
∗∗∗
(0.409) -0.785
∗∗
(0.420) -1.057
∗∗
(1.544)
(0.235)
(0.262)
(0.056)
(0.015)
(0.333)
(0.329)
(0.359) 0.301
(0.502 ) 0.405
(0.088) -0.166
∗∗
(0.015) 0.056
∗∗∗
(0.457) -1.235
∗∗∗
(0.449) -1.313
∗∗∗
(1.417)
(8.911)
-1.294
3.275
(0.256)
(0.291)
(0.061)
(0.012)
(0.333)
(0.318)
(1.066)
(6.569)
(0.005) 0.005
(0.003)
(0.135 ) -0.040
(0.131)
(0.384) -0.603
(0.463)
(1.350) 2.677
∗∗∗
(0.009) -0.024
∗∗∗
(0.702)
(0.005)
(0.161)
(0.454)
0.127
(0.126)
-0.255
(0.432)
Own-Price Elasticity
Implicit Discount Rate
WTP ES Label, e = I ($)
WTP ES Label, e = ES ($)
Prob. Taking Rebate
Q ( e = I )
Q ( e = ES )
Q ( e = U )
-7.1
0.08
-57.4
151.8
0.52
0.19
0.16
0.65
-5.8
0.11
1.9
95.8
0.29
0.34
0.24
0.41
-4.6
0.06
-18.0
143.8
0.15
0.30
0.37
0.33
Nb Obs.
49,279 76,115 76,115
Notes: In the estimation, estimated product fixed effects enter the individual choice probabilities and are treated as data. Standard errors are obtained using the Murphy and Topel’s approach (1985).
∗
( p < 0 .
05),
∗∗
( p < 0 .
01),
∗∗∗
( p < 0 .
001)

Main Estimates Firms
Bunch at
Minimum
No Label
Effect
All
Informed
Q ( e = I ) = 100%
I
Mean
II
Std
III IV V VI VII
Mean Std Mean Std Mean
VIII
Std kWh Purchased
(kWh/year)
50.3
Externality Costs 36.0
(M$/year)
Consumer Surplus -1.7
(M$/year)
Profits -882.9
(M$/year)
Total Welfare
(M$/year)
Change in Price
($)
-920.6
-60.2
25.2
18.0
13.8
369.2
379.8
32.9
97.0
69.4
-25.0
-1003.7
-1098.1
-81.4
20.6
14.7
15.3
336.8
343.8
27.1
14.7
10.5
8.6
-615.5
-617.4
-32.2
17.0
12.2
19.8
398.4
394.0
33.1
-70.9
-50.7
62.7
-418.4
-304.9
-7.7
23.7
17.0
30.2
412.3
419.3
36.4
Notes: The table reports the difference between a market without and with ES. The counterfactual scenario is the market without ES. A negative sign implies that removing ES leads to a decrease in a particular metric relative to a market with ES. The model is bootstrapped 100 times. The mean and standard deviation of the difference in various metrics are reported. Columns I and II are the main estimates. They suggest that removing ES decreases total welfare, but most of the decrease comes from a reduction in profits. Columns III and IV report estimates assuming that firms will bunch exclusively at the minimum standard when ES is not in effect. Columns V and VI report on a scenario where the label effect is set to zero, i.e., τ
I
= 0 and τ
ES
= 0. Without the label effect, profits decrease and the energy savings brought by ES are much lower relative to the main scenario (Columns I and II). Columns VII and VIII compare a scenario where all consumers are perfectly informed to the basecase scenario where
ES is in effect. Under perfect information, consumers are better off, energy savings are larger, but profits are lower.
39

40
A.
Alternative Regressions: ENERGY STAR Price Premium . The data used for this analysis are the weekly prices of all refrigerator models sold at a major US appliance retailer. The data from this retailer are not included in the NPD sample, and thus presents an alternative view of the pricing strategy in the appliance market. I exploit two events that occurred in the refrigerator market that led to the decertification of refrigerator models. As mentioned in Section II.B, in 2008 the EPA revised the certification requirement for full-size refrigerators. A large number of products then lost their ES certification. In January 2010, a similar event occurred, but it was unexpected by the firm.
For each refrigerator two prices are observed, the manufacturer’s suggested retail price (MSRP) and the promotional price. The data cover the period January 2007 to December 2011. I estimate the impact of the change in decertification using a difference-in-difference estimator. For the 2008 decertification, I use refrigerator models that were never certified ES or met the new certification as a counterfactual. For the 2010 decertification, I restrict the sample to refrigerator models that were certified as of January 2010. In my sample, 16 of these refrigerator models lost their certification because of the problematic testing procedure. Models that were certified ES and did not lose their certification serve as counterfactuals.
The effect of decertification is estimated with the following model:
(11) log ( P jt
) = ρEST AR jt
+ α t
+ γ j
+ jt where α t and γ j are week and product fixed effects, respectively. The dependant variable is the log of the weekly price. The estimation is performed using the MSRP or the promotional price.
EST AR jt is a dummy variable that takes a value one if product j is certified in week t , and zero otherwise. The dummy variable EST AR jt only varies for products that lose their certification in 2008 or 2010. The dummy γ j captures all time-invariant product attributes specific to each refrigerator model. The coefficient ρ is then the quantity of interest, and estimates the impact of removing the ES label on a given refrigerator model.

41
2008 Decertification I estimate the effect of the 2008 decertification with data from January
2007 to December 2009. During that period, there were 2,337 different refrigerator models sold at the retailer, and 1,347 lost their ES certification in April 2008.
Figure 1 (panels a and b) provides graphical evidence of the impact of decertification on prices.
The average normalized prices (MSRP and promotional) for three efficiency classes are shown: models that lost their certification, models that were not certified ES as of January 1 2007, and models that did not lose their certification following the revision. Normalized prices are computed by dividing the price of each refrigerator model by its average price. Figure 1 plots the mean and the standard errors of a flexible regression spline fitted on the normalized price, and allows for a discontinuity in the last week of April 2008.
For both the MSRP and the promotional price, there is no clear evidence that the prices of decertified models changed around the time of the revision. In the post-revision period, however, we observe that the MSRPs of non-ES models, and ES models that remained certified, have a strong upward trend cumulating with a large price increase in the first week of the year 2009. The prices of decertified ES models have a similar trend, but it is much less pronounced. In relative terms, decertified models thus became less expensive in the post-revision period. The trends for promotional prices are similar, although they are subject to larger weekly variations.
The difference-in-difference estimator provides a valid estimate of the effect of decertification, if we believe that decertified models would have been subject to a similar trend than non-decertified models in the post-revision period had the decertification not occurred. This is a strong assumption.
There are a number of factors that could explain the large increase in prices in 2008. I conjecture that the Great Recession might have played a role. Firms anticipating lower demand for durables might have increased prices, which will be consistent with theories of countercyclical markups.
Table 1 presents the estimates. Consistent with the graphical evidence, they both suggest that the decertification led to a small but significant relative price decline.
18
The reduction is 4.6% for
MSRP, and 2.0% for promotional prices.
18
A positive coefficient for ρ means that prices are higher when products are certified EST AR = 1.

42
Revision Energy Star
Revision Energy Star
2008w1
Non Energy Star Models
Decertified Energy Star Models: 15% Better than Minimum
Energy Star Models: 20% Better than Minimum
2008w27 2009w1
Weeks
2009w26
(a) MSRP: 2008 Revision
2010w1 2008w1
Non Energy Star Models
Decertified Energy Star Models: 15% Better than Minimum
Energy Star Models: 20% Better than Minimum
2008w27 2009w1
Weeks
2009w26
(b) Promotional Price: 2008 Revision
2010w1
Notes: Each panel displays average normalized weekly prices, with 5% confidence intervals, of refrigerators that belong to different efficiency classes. The normalized price for each model is the weekly price (MSRP or promotional) divided by its average weekly price. The average normalized price and standard errors in each efficiency class are computed by fitting a cubic spline on normalized prices. Three efficiency classes are considered: models that were not certified ES before and after the revision (less than 15% more efficient than the minimum standard), models that lost the ES certification (15-19% more efficient than the minimum standard), and models that met the revised standard before and after the revision
(at least 20% more efficient than the minimum standard.)

43
EST AR
2008 Revision in ES Standard
MSRP Promo.
(I)
Dependent Var.:
(II) log(price)
0.046
∗∗∗
(0.0039)
Yes
Yes
0.020
∗∗∗
(0.0038)
Yes
Yes
MSRP
2010 Revision in Certification
Promo.
(III)
Dependent Var.:
(IV) log(price)
0.045
∗∗∗
(0.0051)
Yes
Yes
0.012
(0.0073)
Yes
Yes
Week FE
Product FE
Nb of Models 2,623
Nb of Decertified 1,334
Models
Nb of
Observations
Adjusted R 2
130,914
0.041
2,623
1,055
130,914
1,367
16
57,601
1,367
16
57,601
0.013
0.031
0.004
Notes: For specification I and II, refrigerator models that were not certified ES as of January 1 st 2008, models that were certified and met the more stringent standard, and models that lost their ES certification are considered. For models V-VIII, only refrigerator models that were certified ES as of January 1 st are considered. For 2008, I assume that the decertification of ES models occurred in the seventeenth week. For 2010, I assume that the decertification occurred in the fifth week. The dummy variable EST AR jt
Standard errors are clustered at the product level.
∗ p < 0 .
takes the value one if a refrigerator is certified in week
05,
∗∗ p < 0 .
01,
∗∗∗ p < 0 .
001 t .

44
2008 2010
Choice Set
Nb of Refrigerator Models, US Market
Nb of Decertified Refrigerator Models, US Market
Nb of Refrigerator Models, Retailer’s Sample
Nb of Decertified Refrigerator Models, Retailer’s Sample
Average Size of Choice Set at Retailer (Store-Trimester)
SD Size of Choice Set (Store-Trimester)
Price
Avg Promotional Price
SD Promotional Price Offered
Avg Promotion
Avg Duration of a Promotion
Avg Promotional Price ENERGY STAR
Avg Promotional Price Non-ENERGY STAR
Avg Promotional Price Decertified Model
Electricity Costs: US Refrigerators
Avg Elec. Price (County)
Max Elec. Price
Min Elec. Price
Avg Elec. Consumption
Avg Elec. Consumption ENERGY STAR
Avg Elec. Consumption Decertified ENERGY STAR
Avg Elec. Consumption Non-ENERGY STAR
Choice Set: Cost Estimation/Policy Simulation
Nb of Refrigerator Models
Brand
Brand A
Brand B
Brand C
Brand D
Generic Brand
2,524
1,278
1,483
674
99
44
$1,325
$639
11%
1.4 week
$1,426
$1,240
$1,416
1,496
21
1,174
14
92
37
$1,411
$693
15%
3.1 week
$1,368
$1,472
$2,468
0.113$/kWh 0.113$/kWh
0.03$/kWh 0.03$/kWh
0.420$/kWh 0.368$/kWh
520 kWh/y 506 kWh/y
500 kWh/y 501 kWh/y
520 kWh/y 547 kWh/y
568 kWh/y 525 kWh/y
78
(%)
10.0
17.3
41.3
9.3
24.0
Size
AV < 23 .
5 cu. ft.
AV ≥ 23 .
5 cu. ft.
Refrigerator Style
Top Freezer
Side-by-Side
Bottom-Freezer
ES Certification
No
Yes
(%)
49.3
50.7
(%)
16.0
25.3
58.7
(%)
29.3
70.7

45
Structural Estimator: Marginal Cost of Providing Energy Efficiency . The unit cost of product l offered by brand manager k is given by:
(12) c ( f kl
) = e
( α kl
+ φf kl
)
The parameters to be estimated are α kl
, a constant specific to each product, and the parameter φ that determines how the unit cost varies as a function of energy efficiency.
The estimation procedure combines moment inequalities with the first order conditions of the firm’s problem that hold with equalities. The goal of the estimator is to select the parameters values the most consistent with profit maximization. More formally, define the maximum profits of firm k by π
∗ k
. The observed product lines and prices are assumed to be the ones maximizing profits, and correspond to the vectors f
∗ k and price of product l and set f
∗ kl and p
∗ k
. Suppose that firm k changes the energy efficiency level
0
→ f kl and p
∗ kl
→ p
0 kl
, but all other products remain the same, including the products offered by competitors. The counterfactual profits for a small change in the value of product l are noted π l k
0
≡ π ( f kl
, p
0 kl
, f
∗
− l
, p
∗
− l
), and π
∗ k
≥ π l k according to profit maximization.
I do not observe all components of profits, therefore the quantity π
∗ k
− π l k cannot be computed exactly. In the present application, I do not observe the fixed (sunk) costs of developing each product line, and the cost of certifying a product. Market shares and unit costs, however, are observed (i.e., estimated), which allows me to compute brand manager markups and profits net of fixed and certification costs. I distinguish between the observed and unobserved components of the profit function as follows:
(13) π k
= r k
( f, p | φ ) − ξ k
J k
−
J k
X
ζ kj
1 { f kj j
≥ f
ES
}
As in Pakes, Porter, Ho, and Ishii (2011), I distinguish between two types of unobservables.
ξ k represents unobserved fixed costs that firm k faces to develop each product offered. This includes
R&D activities, inventory management, and marketing effort. These fixed costs should be economically significant and correlated with firms’ characteristics. I then treat ξ k as an anticipated shock, which plays the role of the structural error term at the source of endogeneity.
ζ kj represents the additional fixed cost that firm k incurs to develop a product that meets the certification requirement.
I assume that ζ kj is an idiosyncratic and unanticipated cost, which plays the role of measurement error in the estimation. Finally, note that the indicator function identifies the certified products.

46
The observed component of profits, r k
( f, p | φ ), is the sum of brand manager markups weighted by market shares:
(14) r k
( f, p | φ ) =
J k
X
( p kj
− c kj
( f kj
| φ )) · Q kj j =1
For ease of notation, I only express the dependency of r k on the parameter φ . The profits also depend on the constant α kl and the demand parameters, which are both estimated separately, as discussed below.
The objective function of the maximum-score estimator is constructed using the following approach. For the five brands active in the market, I compute the observed component of profits, r k for all k , using the demand estimates, observed prices, and the cost function evaluated at φ . I then perturb the energy efficiency level and price of each product offered, and compute the counterfactual profits at each perturbation. For products that bunch at ES, the perturbation consists of setting the energy efficiency level and price to their optimal level if ES was not in effect. I do this simply by using the first-order conditions of the oligopoly model assuming that there is no certification. For products that do not bunch at ES, the perturbation consists of setting the energy efficiency level equal to the certification requirement. I then find the optimal price associated with this energy efficiency level. Pairwise comparison between optimal and counterfactual profits for a change in the energy efficiency level of product l offered by firm k then yields:
(15) π
∗ k
− π l k
= r
∗ k
( f
∗
, p
∗
| φ ) − r i k
0
( f k,i
, p
0 k,i
, f
∗
− i
, p
∗
− i
| φ ) + ζ kj where the fixed costs of developing product lines cancel out because the number of products offered by each firm is the same at each perturbation. The unobserved certification cost ζ kj remains because for each perturbation a product either gains or loses the certification. I treat ζ kj as an error term with an unknown distribution. I assume that ζ kj are i.i.d. for each firm k and has support equal to the real line. As shown by Manski (1975), those are sufficient conditions for the rank ordering property to hold,
19 which in turn allow me to construct the pairwise maximum-score objective function proposed by Fox (2007). In the present application, an estimate of φ is obtained
19
In this application, the rank property holds if r
∗ k
( f
∗
, p
∗ | φ ) > r i k
0
( f k,i
, p
0 k,i
, f
∗
− i
, p
∗
− i
| φ ) implies
F ( r
∗ k
( f
∗
, p
∗ | φ )) > F r i k
0
( f k,i
, p
0 k,i
, f
∗
− i
, p
∗
− i
| φ ) , where F is the probability that a given level of profits is chosen.

47 by maximizing:
(16) Q ( φ ) =
X
J k
X
1 [ r
∗ k
( f
∗
, p
∗
| φ ) − r l k
0
( f kl
, p
0 k,l
, f
∗
− l
, p
∗
− l
| φ ) > 0] k l where the indicator function takes the value one if the optimal profits are greater that the counterfactual profits. In practice, Q ( φ ) takes the forms of a step function, and thus only identifies a range of possible values for φ .
Point identification is possible, however, when some products do not exactly meet the minimum and ES standards. For these products, the optimal energy efficiency level is a solution of:
(17) Q kl
( f
∗
, p
∗
) dc kl
( f
∗ kl
) df kl
−
J k
X
( p
∗ kj j =1
− c kj
( f
∗ kj
)) ·
∂Q kj
( f
∗
, p
∗ | θ )
∂f kl
= 0 .
The above condition can be used to form the following moment:
(18) M ( φ ) =
X
J k
X
1 [ f kl
= f ∨ f
ES
] × k l


Q kl
( f
∗
, p
∗
) dc kl
( f
∗ kl
| φ ) df kl
−
J k
X
( p
∗ kj j =1
− c kj
( f
∗ kj
| φ )) ·
∂Q kj
( f
∗
, p
∗
)


∂f kl
= 0 , where the indicator function takes the value one if f kl does not equal the minimum standard (f) or the ES standard ( f
ES
).
For the estimation of the constant α kl specific to each product, I rely on the first-order conditions with respect to prices. Because there are P k
J
K first-order conditions with respect to prices, and the vector of unit costs, c ( f ), has P k
J
K system.
20
For a given value of φ elements, unit costs are the solution of a just-identified and a vector of estimates of unit costs, I can then recover each constant α kl by simply inverting the cost function (Equation A).
To summarize, the cost estimation proceeds as follows:
(1) Guess a value of the parameter φ .
20
In matrix notation, the first-order equations with respect to prices are: D ( f, p )+ D p
( f, p ) × [ p − c ( f )] = 0, where D is a vector of demand for a given firm, and D p is the Jacobian matrix with respect to prices. When
D , D p and p are known, the vector of unit cost, c ( f ), can be obtained simply by matrix inversion.

48
(2) Using the first-order conditions with respect to prices, solve a system of P k
J
K equations to recover the P k
J
K kl
).
(3) Compute the constants of each cost function using: α ˆ = log ( ˆ ) − φf kl
.
(4) Compute the pairwise maximum-score objective function: Q ( φ ).
(5) Compute the moment condition for products do not exactly meet the minimum and ES standards: M ( φ )
(6) Return to step 1.
I perform a grid search over the possible values of the parameter φ , and select the value that maximizes Q ( φ ) and sets M ( φ ) the closest to zero, with equal weight to the two objective functions.
I do not estimate the standard error for φ , but in the simulations I will provide sensitivity tests with respect to this parameter.
Alternative Estimator: Marginal Cost of Providing Energy Efficiency . In the refrigerator market, manufacturers commonly offer product lines that consist of a group of three to ten refrigerator models with similar overall design, such as the size and style (top freezer, side-by-side, bottom-freezer), but that differ with respect to less important attributes, such as the ice-maker option, the finish option (stainless or not), and the energy efficiency levels. In my sample, I observe
78 product lines offered between 2007 and 2011 with very similar refrigerator models that have different energy efficiency levels, but otherwise similar attributes. With these product lines, I am able to match 155 refrigerator models with at the least another model that has the same size, style, brand, door material (stainless or not), ice-maker option, defrost technology, and entered the market the same year (Table 3). Matched refrigerators differ only in color and energy efficiency levels; all other observed attributes are identical.
For all matched refrigerator models (N=155), I simply exploit variation in energy efficiency within product line, and estimate the marginal cost of providing energy efficiency by regressing the log of the wholesale price on a product line fixed effect, dummies for color, and a proxy for energy efficiency:
(19) ln ( price j,r
) = α + γ j,j
0
+
K
X
γ k
C k j k
+ φEf f iciency j
+ j,r
,

49 where γ j,j
0 is a product line fixed effect that is common to the matched refrigerator models j and j
0
, and C k j are dummy variables that are equal to one if refrigerator j is of a given color. For the proxy for energy efficiency, I use the same functional form than for the structural estimator, where energy efficiency is defined as the inverse of the annual electricity consumption. The estimate of the parameter φ is 362, which corresponds to a cost elasticity of 0.66% (Table 3).
φ
Structural Estimator
429
Matching Estimator
362 †
(208)
0.66
Cost Elasticity:
∂c ( f )
∂kW h kW h c ( f )
0.78
Nb of Product Lines
Nb of Products
78
78
78
155
Notes: For the structural estimator, a grid search is performed to find the value that optimizes the objective function. The matching estimator takes the wholesale price as the dependent variable.
For each matched refrigerator model, only one wholesale price is observed. The dependent variable is the inverse of the annual electricity consumption of each product. Product line fixed effects are included. Dummies for color are also included, none of them are statistically significant. Models that have the same size, style, brand, door material (stainless or not), ice-maker option, defrost technology, and entered the market the same year, share the same product line fixed effect.
† p < 0 .
10,
∗ p < 0 .
05,
∗∗ p < 0 .
01,
∗∗∗ p < 0 .
001
Demand Estimation . To compute the discount rate implied by the demand parameters η and
θ , I first assume that consumers form time-unvarying expectations about annual electricity costs.
Moreover, I assume that consumers do not account for the effect of depreciation, and compute the lifetime electricity costs ( LC r,j
) by summing and discounting the expected annual electricity costs
( C r,j
) over the lifetime of the durable:
(20) LC r,j
=
L
X
ρ t
C r,j t =1
=
1 − ρ
L
C r,j
,
1 − ρ where L is the lifetime of the durable, ρ = 1 / (1 + r ) is the discount factor, and r is the (implied) discount rate.

50
Consistent with the fact that | η | corresponds to the marginal utility of income, the coefficient for the sensitivity to annual energy costs in the demand model ( θ ) is a reduced form parameter that corresponds to:
(21) θ = η
1 − ρ L
1 − ρ

51
Emission Factors .
Non-baseload Output Emission Rates (U.S. Average)
Pollutant Estimate Source
CO 2
CH 4 a
N 2 O a
1,583 lb/MWh
35.8 lb/GWh
19.9 lb/GWh USEPA, eGRID2007
SO 2
N Ox
6.13 lb/MWh
2.21 lb/MWh
Damage Cost (2008 $)
Pollutant Low Estimate High Estimate Source
CO
SO
2
2
N Ox
$21.8/t
$2,060/t
$380/t
$67.1/t
$6,700/t
$4,591/t
Greenstone, Kopits, and Wolverton (2011) low: Muller and Mendelsohn (2012), high: USEPA b low: Muller and Mendelsohn (2012), high: DOE c
Notes: (a) Externality costs associated to CH 4 and N 2 O are assumed to be the same than for CO 2.
CH 4 and N 2 O are converted in CO2 equivalent using estimates of global warming potential (GWP). The GWP used for CH 4 is 25, and the
GWP used for N 2 O is 298. Source: IPCC Fourth Assessment Report: Climate Change 2007. (b) Estimate used in the illustrative analysis of the 2012 regulatory impact analysis for the proposed standards for electric utility generating units.
(c) Higher value of the estimate used in the Federal Rule for new minimum energy-efficiency standards for refrigerators
(1904-AB79).

52
Sensitivity Tests . Table 5 present various sensitivity tests. For a low marginal cost, where φ =
150, removing the certification leads to a smaller decrease in electricity consumption. Put another way, ES is associated with smaller energy savings. This result is explained by the fact that with a lower cost of providing energy efficiency, firms offer more models that exceed the minimum standard
(Figures 2(b) and 2(a)), and they can do so at a lower cost when the certification is not in effect.
For a larger marginal cost of providing energy efficiency, where φ = 650, the model predicts, as expected, more bunching at the minimum standard when ES is in effect (Figures 2(d) and 2(c)).
Removing the certification leads to slightly larger energy savings, but the results are qualitatively similar to the main scenario.
The third sensitivity test investigates the effect of the label effect. For this simulation, the label effect for consumers that rely on ES was reduced by half for all income groups, and set to zero for perfectly uniformed consumers. As a result, firms offer less products that bunch at ES (Figures
2(e) and 2(f)). Results are otherwise similar to the main scenario.
kWh Purchased (kWh/year)
Externality Costs (M$/year)
Low Marginal High Marginal Low Label
Cost Cost Effect
I II III IV V VI
Mean Std Mean Std Mean Std
39 27 60 28 37 22
28 19 43 20 26 16
Consumer Surplus (M$/year) -15
Profits (M$/year) -828
31
369
4
-860
21
341
7 17
-846 376
Total Welfare (M$/year)
Change in Price ($)
-871
-39
363
30
-899
-73
366
30
-865 380
-54 31

0 10 20 30 40 50
Energy Efficiency: Percentage Better than Minimum Standard
(a) w ES, φ = 150
0 10 20 30 40 50
Energy Efficiency: Percentage Better than Minimum Standard
(b) w/o ES, φ = 150
0 10 20 30 40
Energy Efficiency: Percentage Better than Minimum Standard
50
(c) w ES, φ = 650
0 10 20 30 40
Energy Efficiency: Percentage Better than Minimum Standard
50
(d) w/o ES, φ = 650
0 10 20 30 40
Energy Efficiency: Percentage Better than Minimum Standard
50 0 10 20 30 40
Energy Efficiency: Percentage Better than Minimum Standard
50
(e) w ES, φ = 429, Small Label Effect (f) w/o ES, φ = 429, Small Label Effect
53