Bunching With the Stars: How Firms Respond to Product Certification Sébastien Houde∗ September 15, 2012 Preliminary and Incomplete, Do Not Circulate Abstract The Energy Star certification is a voluntary labeling program managed by the US Environmental Protection Agency (EPA) that favors the adoption of energy efficient products. This paper first shows that Energy Star influences firms decisions. Focusing on the refrigerator market, firms offer products that bunch exclusively at the minimum and Energy Star standards. I also find some evidence that consumers’ valuation of the Energy Star certification is factored in pricing decisions. The second part of this paper uses an oligopoly model calibrated for the US refrigerator market to investigate firms’ product lines and pricing decisions under various scenarios. I show that in a world without Energy Star firms might discriminate less in the energy efficiency dimension. This has important implications in determining the welfare effects of the program, and ultimately for the design of voluntary energy efficiency standards. ∗ I am indebted to Jim Sweeney, Wes Hartmann, Jon Levin and John Weyant for all their helpful advice in guiding me in this project. Department of Management Science and Engineering, Stanford University, Stanford, CA 94305. Email: shoude@stanford.edu. 2 The economics of energy efficiency policies is controversial. While engineering calculations have repeatedly suggested that the full technical potential for energy efficiency is far to be realized, economists have pointed out that moving closer to the full potential is desirable to the extent that it delivers net benefits to society. The point of contention is that many practitioners and researchers believe that the observed investments in energy efficiency are systematically below their socially optimal levels, a phenomenon known as the energy efficiency gap. It is however still much debated how big is the gap, why it arises, and whether it commands policy interventions. Among the policy tools proposed to address the energy efficiency gap, labels and voluntary certification programs are among the most popular. In several countries, it is mandatory to sell appliances with labels that displays energy information. The design of these labels vary widely, but often complex energy information is summarized with a star system that groups products in different efficiency classes (e.g., in Europe, India, Australia, and Japan). In the US and Canada, the Energy Star program uses such design to identify the most energy efficient products within product classes. Summarizing complex energy information with a simple graphical display has an obvious benefit, it eases consumers’ burden of processing information, which should increase the likelihood that consumers will account for energy efficiency in their purchasing decision. Certification systems that rely on coarse categorization bring however a number of perverse incentives. For instance, they make consumers insensitive to variations within category. In the energy context, this means that consumer may invest too much or too little in energy efficiency. Certification system may also distort equilibrium outcomes. If consumers rely heavily on a given certification, firms should respond by making product lines and pricing decisions that are consistent with consumers’ beliefs. Whether consumers are better off when the market is distorted by a certification remains uncertain. The goal of this paper is to first show how firms account for certification in their decisions to offer energy efficient products. The focus is on the Energy Star program. Using data from the US refrigerator market, I show that firms account for Energy Star in both their product lines and pricing decisions. I then propose a model that rationalizes firms’ responses and show that a simple static multi-product oligopoly model can explain various stylistic facts. After estimating the primitives of the model, I simulate various counterfactual scenarios that inform about the value of certification. In particular, I conduct a welfare analysis of the Energy Star certification program. In addition of bunching strongly at the minimum and Energy Star standards, I show that firms are charging higher markups on Energy Star refrigerators. Using various strategies, I find that the 3 additional markup on Energy Star models is on the order of 2-3%. This estimate comes remarkably close of my estimate of consumers’ willingness to pay for Energy Star (Houde 2012). For my preferred estimate of firms’ marginal cost of providing energy efficiency, I find that without Energy Star, firms will offer products that bunch almost exclusively at the minimum energy efficiency standard. Energy Star thus leads to significant energy savings. Interestingly, the largest benefits associated with Energy Star might however not come from the reduction in the environmental externalities due to electricity generation. The interactions between Energy Star and firms’ market power are at the source of the most important welfare gains. 1. Related Literature forthcoming 2. A Primer on Energy Star and the US Refrigerator Market The Energy Star certification program was established by the US EPA in 1992. The goal of the program is to favor the adoption of energy efficient durables by residential and commercial consumers. The EPA sets voluntary energy efficiency standards and certifies products that meet or exceed the standards. Qualifying products can be labeled with an Energy Star logo (Figure 1). The logo does not provide detailed information about energy efficiency. The program complements minimum energy efficiency standards and the mandatory energy label EnergyGuide, which provides detailed technical information on energy consumption and costs (Figure 1). While minimum energy efficiency standards place a lower bound on efficiency, Energy Star aims to induce more efficient purchases by providing simple and salient information to consumers. The Energy Star standards for refrigerators are defined relative to the minimum energy efficiency standards. Since April 2008, full-size refrigerators must consume 20% or less electricity than the minimum energy efficiency standards to be certified Energy Star. The minimum energy efficiency standards for refrigerators vary as a function of size, door style (single door, top-freezer, bottomfreezer and side-by-side), and attributes (ice-maker and defrost technology). Table 1 summarizes the criteria for minimum and Energy Star standards. The periodic revisions of the Energy Star standards is a crucial feature of the program. The EPA revises the stringency of the standards using various criteria, such as the proportion of Energy Star products offered on the market, the market share of Energy Star products and the availability 4 of new technologies (McWhinney, Fanara, Clark, Hershberg, Schmeltz, and Roberson 2005). The stringency of the standards is ultimately determined by US EPA upon consultation with different stakeholders, such as manufacturers, part providers, retailers, analysts and environmental groups. Revisions in standards are usually announced by US EPA about one year in advance. When a revised Energy Star standard comes into effect, the EPA requires that manufacturers and retailers remove the Energy Star label on certified products that do not meet the more stringent standard. For full-size refrigerators, the Energy Star standard has been revised in 2001, 2004 and 2008. The minimum energy efficiency standard for refrigerators was last revised in 2001. The Department of Energy (DOE) has announced that a revised version of the minimum standard will be implemented in 2014 (Table 1). The new Energy Star standard has yet to be announced. An important institutional feature of the US electricity markets is that several electric utilities are subject to regulations that incentivize them to promote energy efficiency measures to consumers. Rebate programs tied to the purchase of Energy Star products have been arguably one of the most popular energy efficiency measures pursued by US electric utilities. Thus, apart from the informational component, financial incentives may play a role in favoring the adoption of Energy Star products (Datta and Gulati 2009). I account for this in my empirical strategy and I distinguish between the information effect and the rebate effect associated to the program. The refrigerator market, like other appliance markets, is an important contributor to the manufacturing and retailing sectors of the US economy. Almost all American households own at least one refrigerator, and the fraction of households owning two or more refrigerators has been steadily increasing. In 2010, 9.01 million of full-size refrigerators were shipped to the US market (AHAM, DOE). An important feature of the US refrigerator market for the present analysis is that it has becomes increasingly concentrated in recent years. As of 2008, the market was dominated by three manufacturers: Electrolux, General Electric (GE), and Whirpool. Together they held about 85% of the market share for full-size refrigerators (Table 2). This high concentration is notably the result of several mergers and acquisitions that have taken place since the early eighties. Figure 2 summarizes the market structure; the notable mergers and acquisitions are depicted by circles and the most popular brands and their relationship with the main three manufacturers are shown in the dotted boxes. A particular institutional detail of the market is that manufacturers compete under various brand names, and some brands, such as Kenmore, are not own by a particular manufacturer. This feature of the market is believed to be an important factor that limits manufacturers’ market power. 5 The distribution of products across brands is still however fairly concentrated, especially in recent years; most products are offered by the major brands associated to the top three manufacturers and Kenmore (Table 3). Beyond its significance for the US economy and its interesting market structure, the refrigerator market also provides a good laboratory to develop and test my framework, because modeling the purchase decision for refrigerators has a particular methodological appeal. Refrigerators are one of the few energy intensive durables for which the utilization decision does not need to be explicitly modeled to have an accurate estimate of how much each consumer will pay to operate its refrigerator. For this particular good, the utilization decision is then not, presumably, at the source of important unobserved heterogeneity. 3. Preliminary Evidence This section investigates whether firms, namely manufacturers, brand managers, and retailers account for the Energy Star program when deciding which products to offer and how to set prices. Using different data sources, I provide evidence that firms participating in US the refrigerator market are well aware of the Energy Star program, and ultimately the program is an important determinant of equilibrium outcomes. 3.1. Data The California Energy Commission (CEC), the Federal Trade Commission (FTC) and the EPA maintain databases for the US appliance market.1 I use these three data sources to reconstruct the choice set for the whole US refrigerator market and the timing of product entry and exit for the period 2003-2010. Using the data from a large retailer (see appendix for a full description), I also look at the pricing strategies of manufacturers and retailer(s). For each refrigerator model in the sample, I observe three types of price: the manufacturer’s suggested retail price (MSRP), the promotional price, which is set by the retailer, and the manufacturer price, which is the price paid by the retailer to the manufacturers. 1 The three agencies provide similar data, but with some important differences. The CEC has data for the Californian refrigerator market for the period 1978-2011. The FTC has data for the whole US refrigerator market for the period 2003-2011. The EPA has data for the whole US refrigerator market and information about which models are certified Energy Star. 6 3.2. Product Line In several appliance markets, there is clear evidence that manufacturers and brand managers2 account for Energy Star when deciding the energy efficiency levels of their products.3 This is illustrated by Figure 3, which shows the distribution of refrigerator models in the energy efficiency dimension for the year 2010. Refrigerator models offered either just meet the federal minimum standard or the Energy Star standard; few models are located between the standards, or exceed the Energy Star threshold. As a result, the distribution bunches strongly at the minimum standard and the Energy Star standard. Looking at the dynamics in the choice set, the data also suggest that manufacturers have the ability to change their product lines quickly, and adjust to revision in Energy Star within one to two years. Figure 4 shows how product lines evolved during the period 2003 to 2010 for one particular type of refrigerator, the full-size bottom-freezer without an ice-maker. We observe that following the revision of the Energy Star standard in 2004 and 2008, manufacturers responded not only by offering new models that met the revised standard, but also by discontinuing models that were decertified.4 Note that the 2008 revision was announced exactly one year in advance by the EPA, and most manufacturers were able to offer products that meet the more stringent standard within a year. In 2010, about eighteen months after the standard was revised, most decertified products had exited the market. Altogether, the distribution of refrigerators in the energy efficiency dimension, and adjustments to revisions in the Energy Star standard show that manufacturers and brand managers optimize their product lines to account for Energy Star. As I will show, these decisions are consistent with profit maximizing behaviors where firms differentiate their products to extract the maximum surplus from consumers. I next turn to the analysis of pricing decisions, and provide further evidence consistent with this theory. 3.3. Pricing If a fraction of consumers value Energy Star products, and products can be differentiated, firms may be able to charge larger markups on Energy Star products. I now investigate how firms 2 Because some brands are not owned by any manufacturers, product line decisions can be attributed to manufacturers and brand managers. 3 Sallee (2011) provides such evidence for several appliance markets. He also shows that certification has a similar effect in the building market. 4 This pattern holds for other types of full-size refrigerators, as well. 7 set prices to account for Energy Star, and provide estimates of the size of markups associated to certification. I propose three strategies, each relying on a different quasi-experiment. First, I propose a matching estimator that exploits the existence of refrigerator models that have similar attributes, but different energy efficiency levels. Second, I look at the prices of refrigerators before and after the 2008 revision in the Energy Star and use a difference-in-difference estimator to arrive at the change in price due to the decertification. The third strategy is similar to the second, but relies on a different natural experiment: in 2010, the EPA decertified twenty-one refrigerator models because of a problematic test procedure. Matching Estimator. In the US refrigerator market, manufacturers commonly offer product lines with several models. Within a product line, models may be differentiated with respect to their color, their door material (stainless or not), the ice-maker option, and their energy efficiency level. To isolate the effect of Energy Star on markups, I exploit this particularity of the market, and propose a matching estimator that compares, within product line, the markups of refrigerators that differ with respect to their energy efficiency level and Energy Star certification, but have otherwise similar attributes. In particular, I match all refrigerators that have a similar design (top freezer, bottom-freezer and side-by-side), brand, size, door material (stainless or not), ice-maker option, and defrost technology. Matched refrigerators may differ in color and energy efficiency levels; all other observed attributes are identical. In my sample, I observe 265 refrigerators that have at least one match, and I was able to recover 56 product lines where at least one model is certified Energy Star and one model is not certified. Figure 5 compares the (average) percentage markups of Energy Star and non-Energy Star models for each product line. Markups are computed using two approaches. The first approach compares the percentage difference between the manufacturers’ suggested retail prices (MSRPs) and the the prices that manufacturers sell their products to retail stores and brand managers, which are referred as the manufacturer prices. The second approach compares the promotional prices with the manufacturer prices. For each refrigerator models, the markups vary over time because MSRPs and promotional prices vary from week to week. However, for each product, I observe only one manufacturer price. To construct Figure 5, I compute the average markups over the period 20082010, for all models that are certified Energy Star within a product line, and all models that are not certified. 8 We observe that Energy Star models tend to have higher markups (more points above the 45◦ line), especially when markups are computed with promotional prices. Regressing the percentage markups on a product line fixed effect, and a dummy variable that takes the value one if a product is certified Energy Star and zero otherwise, I find that the markups on Energy Star refrigerators are 1.9 percentage points higher (p = 0.025), when MSRP are used to define markups, and 3.0 percentage points higher (p = 0.001), when promotional prices are used (Table 5). In sum, the matching estimator suggests that manufacturers and retailers charge larger markups on Energy Star models. 2008 Revision of the Energy Star Standard. When the EPA revises the Energy Star standard, it requires that manufacturers and retailers remove the Energy Star label on certified products that do not meet the more stringent standard. Following the 2008 revision, it is then possible to observe the same refrigerator model being sold with and without the Energy Star label. In 2008, there were 2,762 full-size refrigerator models offered on the US market, 1,278 of which were decertified.5 My sample contains 1,196 refrigerator models that lost their certification in 2008. Figure 6 compares the average normalized prices (MSRP and promotional price) of models that lost their certification to prices of models that were not certified Energy Star as of January 1st 2008 (N=482), and prices of Energy Star models that did not loose their certification following the revision in standard (N=227). Normalized prices are computed by dividing the price of each refrigerator model by its average price before the revision of the Energy Star standard, i.e., the first seventeen weeks of 2008. Only models that entered the market before the standard are revised are then considered. Figure 6 plots the mean and the standard errors from a regression spline where the normalized price is fitted on week fixed effects and allows for a discontinuity in the seventeenth week. For both MSRP and 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, we however observe that the MSRPs of non-Energy Star models, and Energy Star 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 Energy Star 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 5This number was obtained from the US EPA. According to the FTC data, there were 2,693 full-size refrigerators offered on the US market. 9 trends for promotional prices are similar. In addition, we also observe that the promotional prices of decertified models are much less variable relatively to other models. To estimate the size of the price change, I use the following difference-in-difference estimator: Pj,t (1) = αT af tert + βDecertif iedj + ρT af tert × Decertif iedj + j,t P̄T af terj =0 where the dummy variable T af tert takes the value one for all weeks after the decertification and is zero otherwise, and the dummy variable Decertif iedj takes the value one for refrigerator models that lost their Energy Star certification. PT af terj =0 is the average price (MSRP or promotional) for the period before the decertification (T af tert = 0). The quantity of interest is ρ, the difference-indifference estimator. I also estimate a more flexible version of the model that allows for week fixed effects: X Pj,t αi W eeki,t + j,t = ρEnergyStarj,t + (2) P̄T af terj =0 i where the dummy variable EnergyStarj,t takes the value one if refrigerator model j is certified Energy Star in week t and is zero otherwise. Table 6 presents the estimates. According to the simpler difference-in-difference estimator, the suggested retail price of decertified models decreases by 2.5%. For promotional price, the decrease is 1.9%, and is not statistically different from zero. Controlling for week fixed effects, the decrease in MSRP is 0.3%, and the decrease in promotional prices is not statistically significant. These estimates should be interpreted with caution. The validity of the difference-in-difference estimator requires that both types of refrigerator models be subject to similar time trends in the period before the revision. Figure 6, however, suggest that prices of non-Energy Star models had started increasing before the revision, but decertified models were not subject to such a trend. To summarize, both manufactures and retailers appear to have responded to change in decertification. Perhaps surprisingly, there is limited evidence that they responded by decreasing the prices of decertified models in the post-revision period. Instead, manufacturers appear to have responded by increasing the prices of other models, which led to a relative price decrease for decertified models. Promotional prices follow similar trends, it also appears that retailers responded by excluding the decertified models from their dynamic pricing strategy. 2010 Decertification. In January 2010, the EPA found that a number of Energy Star refrigerator models had undergone problematic testing procedures. As a result, their electricity consumption was underestimated (ref). The EPA concluded that these refrigerator models were not meeting 10 the Energy Star standard, and issued a public statement that 21 refrigerator models would be decertified. I observe 16 of these refrigerators in my sample. Figure 7 shows, respectively, the average normalized suggested retail and promotional prices for refrigerator models that lost their certification, and for models that were certified Energy Star as of January 1st 2009. I now use the models that keep their Energy Star certification as the counterfactual. Clearly, we see that following the decertification, prices (MSRP and promotional) dropped sharply. Using the same difference-in-difference estimators as above, the size of the decrease is 6.5% for the suggested retail price (Model V) and 2.5% (Model VI) for promotional price. Using the estimator that controls for weekly fixed effect (Models VII and VIII), the size of the estimates is about half, respectively 3.6% and 1.2%, but still statistically significant. Overall, the estimates suggest that the manufacturers and retailers responded to the decertification by lowering prices. 3.4. Discussions In the US refrigerator market, product lines and price patterns indicate that manufacturers, brand managers, and retailers account for Energy Star. From the present analysis, five stylistic facts emerge: (1) manufacturers/brand managers offer refrigerator models that just meet the Energy Star standard or the minimum standard; (2) after the 2008 revision, decertified models exited the market within 18 months; (3) Energy Star models have higher markups; (4) in 2008, decertification did not lead to an absolute price decrease, but a relative price decrease; and (5) in 2010, decertification led to an absolute price decrease. Altogether, these suggest that firms strategically optimize product lines and prices. In particular, these facts are consistent with the hypothesis that firms have some market power, which allows them to differentiated their products and maintain positive markups. Interestingly, the size of the markup associated to Energy Star is consistent with my estimate of consumers’ valuation of the Energy Star label. In my demand estimation, I found that consumers value Energy Star models beyond energy savings, which translates to a willingness to pay for Energy Star that is consistent 11 with a price premium of 1.5% to 3%.6 The matching estimator suggests that markups on Energy Star models is about 1.9-3.0% higher than non Energy Star models, and the 2010 decertification led to a price decreasing of about 1.2% to 3.6%. Higher markups for Energy Star models implies that firms extract part of the consumer surplus brought by better energy efficiency. Consumers then face a longer payback period to recoup their investment with energy savings. For instance, if we compare the refrigerator models used for the matching estimator, the difference in prices for Energy Star and non-Energy Star models is about $125 (for MSRP, Table 6). For this price differential, assuming a discount rate of 7%, and an average electricity price of $0.11, the paypack period for an Energy Star refrigerator model is 9.3 year. If we assume that the markup for Energy Star models leads to an increase in price of 3 percentage points relative to price of non Energy Star models, removing this additional markup would decrease the payback period by 4.2 years. One element particularly puzzling among the five stylistic facts is the apparent inconsistency in the price dynamics following the revision of the Energy Star standard in 2008, and the decertification of the 21 models in 2010. It seems at first intuitive to think that if consumers value the Energy Star label itself removing the label should lead to a price decrease. However, if what matters for firms’ market shares is the relative change in prices, increasing the prices of non-decertified models can be equivalent to a decrease in prices in decertified models. In theory, both stylized facts can then be reconciled. A market equilibrium where the prices of non-decertified models increase, i.e., where all prices are weakly larger, should however lead to larger profits than a market equilibrium where prices are weakly smaller. Therefore, if in 2008, firms were able to adjust to the decertification by increasing prices, they should have done similarly in 2010. It should however be noted that in 2008 a large number of models got decertified (1,278 models). In 2010, only 21 models lost their certification. To support a new market equilibrium with higher prices should require an important change in market conditions, which the 2008 decertification might have provided. The 2010 decertification was a marginal event, affecting only two brands, and thus probably of not enough importance to affect the market equilibrium. 6My demand estimation suggests that the Energy Star label influences the probability of selling a given refrigerator model by 7-15%. For an own price elasticity of about 6, a 2% decrease in price will lead to a 12% increase in sale. 12 4. Model This section presents a model of firms’ product lines and pricing decisions focusing on how firms strategically determine energy efficiency levels and prices of refrigerators. To rationalize the stylize facts reported above, a necessary feature of the model is to allow for product differentiation and imperfect competition. I then propose to characterize the US refrigerator market with a static multi-product oligopoly. The model aims to represent a medium-run equilibrium (12-18 months) where the decisions to enter and exit the market, and to determine the size and characteristics of product lines are taken as given. My approach closely follows Bento, Goulder, Jacobsen, and von Haefen (2009), Klier and Linn (2012), and Whitefoot, Fowlie, and Skerlos (2011), who investigate how car manufacturers respond to mandatory fuel economy standards. The present model, however, has a different purpose and looks at the role of voluntary standards. Additionally, a number of institutional features specific to the refrigerator market motivate particular modeling assumptions. 4.1. Assumptions Assumption 1: Firms are Brand Managers. An important feature of the refrigerator market is that manufacturers offer similar products under different brand names. Offering products under different brand names raises a number of issues for manufacturers, such as product cannibalization, and double marginalization. To circumvent the difficulties associated with the manufacturers’ decision to rebrand products, I will focus on modeling the behavior of brand managers. I will assume that each brand manager represents a firm that aims to maximize the profit of his own brand. In this context, the product line decision consists of acquiring refrigerator models through procurement contracts with manufacturers, and brand managers’ costs are simply the prices they pay to manufacturers. Assumption 2: Costs of Providing Energy Efficiency are Separable. A second important feature of the refrigerator market is that, within a relatively short time, manufacturers can change the energy efficiency level of their refrigerators, with little impact on their overall design. This has been demonstrated by the various revisions in the Energy Star standard, which has shown that manufacturers managed to offer new products that were more energy efficient, but were otherwise similar to previous generations.7 I take this as evidence that the cost of providing energy efficiency 7Interestingly, when the EPA announced in April 2007 that the Energy Star standard 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 revised standard (ref). Ultimately, most manufacturers were, however, able to 13 are separable from the cost of providing other attributes. I will further assume that the costs faced by brand managers, i.e., the manufacturers’ prices, reflect this assumption. 4.2. A Static Multi-Product Oligopoly Model Consider that they are K brands, and brand manager k offers Jk refrigerator models. Each brand manager maximizes his profit by choosing the energy efficient levels, the vector fk = {fk1 , ..., fkJk }, and the prices, the vector pk = {pk1 , ..., pkJk }, of his Jk refrigerator models, taking the strategies of his competitors as given. Firms face a population of consumers, where the demand for each product j is a function of consumers’ valuation of energy efficiency and Energy Star. The problem of the brand manager k consists in solving: (3) max fk ={fk1 ,...,fkJk }, pk ={pk1 ,...,pkJk } = Jk X (pkj − ckj (fkj )) · Qkj (f, p|θ, τ ) j=1 where f = {f1 , ..., fK }, p = {p1 , ..., pK }, ckj (fkj ) is the brand-model specific cost of model j offered by brand k, and Qkj (f, p|θ, τ ) is the demand. The demand is a function of all prices and energy efficiency levels, and depends on consumers’ valuation of energy efficiency (θ) and Energy Star certification (τ ). The first order conditions for firm k are given by: Qkl (f ∗ , p∗ |θ, τ ) + Jk X ∗ (p∗kj − ckj (fkj )) · j=1 ∂Qkj (f ∗ , p∗ |θ, τ ) = 0, ∂p∗kl offer new models meeting the 2008 Energy Star standard within a year of the date that the EPA made the announcement. 14 ∗ ES ∗ 1{π(fkl , fk,−l , p∗k ) > π(fkl , fk,−l , p∗k )|∀fkl }× Qkl (f ∗ , p∗ |θ, τ ) ∗) dckl (fkl dfkl − Jk X ∗ (p∗kj − ckj (fkj )) · ∂Qkj (f ∗ , p∗ |θ, τ ) j=1 ∂fkl = 0 , ∗ ∗ 1{π(fkl , fk,−l , p∗k ) ≤ π(f ES , fk,−l , p∗k )|∀fkl }× h i ∗ fkl = f ES , for all l ∈ Jk and k where f ES is the energy efficiency level corresponding to the Energy Star standard, and f ∗ and p∗ correspond to the Nash equilibrium that solves the system of 3 × J1 × J2 × ... JK equations. Note that the second and third conditions arise because the demand function is not continuous at f ES ; the derivative of the profit with respect to energy efficiency level fkl is then not defined at this point. This discontinuity at the Energy Star threshold implies that the firm may find optimal to not equate the marginal cost of providing energy efficiency level fkl with its marginal benefit. Whether firms find optimal to offer products that meet the Energy Star standard depends crucially on consumers’ valuation of energy efficiency, the marginal cost of providing energy efficiency, and firms’ market power. Moreover, heterogeneity in consumers’ valuation of energy efficiency is also an important element that will determine whether firms can differentiate their products with respect to energy efficiency. To illustrate how these different elements come into play, I first solve the model analytically for a simple case: a monopoly facing only two consumer types. The monopoly case allows me to provide a graphical analysis, which will be extended to explain the various stylistic facts previously reported. 4.3. Illustration: Monopoly with Two Consumer Types Consider a market with only one firm selling to two potential consumers, each with different valuation of energy efficiency. Consumer i’s utility of purchasing refrigerator j is given by: (4) Uij = δj + θi fj − pj where δj is the overall quality of the product, θi = {θL , θH } corresponds to consumer i’s valuation of energy efficiency, f , and p is the purchase price. 15 The monopolist must choose the energy efficiency level and price of each refrigerator model that he offers. Assume that the quality is the same for all models, i.e., δj = δ, and is fixed. With only one consumer type, the monopolist will offer one model with an energy efficiency level such that the marginal cost of providing energy efficiency equals the consumer’s marginal valuation of energy efficiency. Formally, if the cost of providing energy efficiency is c(f ), the monopolist sets f s.t. c0 (f ) = θ. He then sets a price to extract the maximum surplus from the consumer, i.e., Uj (p) = 0 ⇒ p = δj + θfj . With two types of consumers, offering two different models might be optimal. The monopolist faces the canonical screening problem; he can sell two different products if the individual rationality (IR) constraints and incentive compatibility (IC) constraints are respected. Formally, the problem of the monopolist is: (5) max fH ,fL ,pH ,pL pH + pL − c(fL ) − c(fH ) s.t. IRH: δ + θH fH − pH ≥ 0 IRL: δ + θL fL − pL ≥ 0 ICH: θH fH − pH ≥ θH fL − pL ICL: θL fL − pL ≥ θL fH − pH It can be shown that having θH > θL , ensures that the incentive compatibility constraint of the low type (ICL) is non-binding at the optimum. Moreover, if ICL is not binding at the optimum, the individual rationality constraint of the low type (IRL) must be binding, otherwise the firm could increase profit by slightly increasing the price pL . By a similar argument, the incentive compatibility constraint of the high type (ICH) must be binding at the optimal, i.e., the consumer with a high valuation of energy efficiency may have an incentive to purchase the product offered to the consumer with a low valuation of energy efficiency (all proofs can be found in appendix). The firm must then distort the prices and energy efficiency levels to ensure that the high type consumer does not find optimal to purchase a less energy efficient product. Using the fact that IRL and ICH are binding, we can solve for prices as a function of energy efficiency levels, and the relaxed form of the monopolist’s problem is given by: 16 max θH (fH − fL ) + 2δ + 2θL fL − c(fL ) − c(fH ) (6) fH ,fL The first order conditions yield: ∗ θH = c0 (fH ) 2θL − θH = c0 (fL∗ ), and the optimal prices are given by: p∗L = δ + θL fL∗ p∗H = θH (fH − fL ) + δ + θL fL At the optimal solution, the firm will distort the energy efficiency level of the less energy efficient product only. If the cost function is increasing with respect to energy efficiency, the monopolist will have to reduce fL to screen between consumers. Relative to the social optimum, the low type consumer then purchases less energy efficient products. Interestingly, if θL is small and θH is large, there might not be a solution where fL∗ is positive. In such case, the firm might simply set fL∗ = 0, or offer only one product (pooling equilibrium). In a market without Energy Star, two possible equilibria might then be sustained: a separating equilibrium corresponding to the solution of 6 (Figure 8(a))8, or a pooling equilibrium where f ∗ solves c0 (f ∗ ) = θL and p∗ = δ + θL f ∗ (Figure 8(b)). When the market is subject to Energy Star, this complicates the matter. The Energy Star program creates a discontinuity in the valuation of energy efficiency, which increase the number of possible equilibria. To model Energy Star, I assume that consumer i’s utility of purchasing refrigerator j is given by: (7) Uij = δj + θi fj + τ Dj − pj where Dj = 1 if the refrigerator is certified Energy Star. The parameter τ subsumes different behavioral effects; it might capture the effect of rebates, bias in the perception of the quality of 8A special case of the separating equilibrium exists when fL∗ = 0 17 Energy Star products, warm glow effect, or energy savings associated with Energy Star products. I will be agnostic on the causes of the discontinuity, and simply assume that both consumers respond to the certification in the same manner, i.e., τ = τ L = τ H . Graphically, this means that the demand curves have a notch of size τ at the Energy Star standard: f ES . With Energy Star, the monopolist might find profitable to depart from the optimality conditions dictated by the first order conditions, and will offer products that bunch at the Energy Star standard. In particular, it can be shown (see appendix) that for each energy efficiency level that solves the problem 6, there is a range of values [f i , f ES ], i = {L, H} for which it is always optimal for the firm to depart from the optimality conditions and bunch at the Energy Star standard. When Energy Star is in effect, there are now four possible equilibria. In addition to the separating and pooling equilibria that can be sustained without Energy Star, a separating equilibrium with one product bunching at Energy Star (e.g., Figure 9(a)), or a pooling equilibria with both products bunching at Energy Star can arise. Before using the model to explain the stylized facts, I make a small detour to discuss how competition will impact the equilibrium outcomes. 4.4. Oligopoly Case When more than one firm is active in the market, firms can set prices above their marginal costs as long as they have some market power. Therefore, if products can be differentiated, notably because of brands, positive markups can be sustained in equilibrium. Rearranging the first order condition for price, we obtain the classic formula for the Lerner index applied to the multi-product oligopoly case (Tirole, 1988): (8) X (pkj − c(fkj )) Qkj kl,kj 1 pkl − c(fkl ) = − pkl kl,kl j6=l Rkl kl,kl where kl,kl ≡ −(∂Qkl /∂pkl )(pkl /Qkl ) is the own elasticity of demand, kl,kj ≡ −(∂Qkj /∂pkl )(pkl /Qkj ) is the cross-elasticity of demand of good kj w.r.t. price kl, and Rkl ≡ pkl Qkl is the revenue associated to good kl. Note that when all products offered by a given firm are substitutes, the cross-elasticity of demand is negative, kl,kj < 0. A firm with multiple competing products will then want to charge higher markups. 18 The effect of competition from other firms can be formalized via the size of the own elasticity of demand. When the market has multiple firms, each firm will face a lower residual demand, which will lead to a larger own elasticity of demand. As expected, competitive forces will then lead to lower markups. Whether firms will still find optimal to differentiate their products in the energy efficiency dimension under competition is uncertain. Clearly, if the market were perfectly competitive, firms could only charge prices that are equal to marginal costs. In such case, firms will see no benefits in differentiating their products in the energy efficiency dimension. Bunching at the minimum and Energy Star standards could then only be rationalized by arguing that standards are focal points that firms converge to.9 Returning to the two consumer types example, when subject to competition, firms will not be able to extract the full surplus of the low valuation consumer (i.e., IRL will not be binding). In a separating equilibrium, both prices will be lower, and consumers will be better off. 4.5. Explaining the Stylized Facts Bunching At the Standards. In the US refrigerator market, we observe a special case of a separating equilibrium, where some products bunch at the minimum energy efficiency standard, but most products just meet the Energy Star standard. This equilibrium is consistent with a market where there is a share of consumers that do not place a high value on energy efficiency, but others value energy efficiency, and especially Energy Star, highly (e.g., Figure 9(b)). Market power combined with consumer heterogeneity can thus explain bunching at the standards. Higher Markups for Energy Star Models. The fact that we observe models bunching at the standards corresponds to a separating equilibrium and reveals that the profit margins on Energy Star models are at least as large as the margins on non-Energy Star models. Otherwise, this would lead to the following contradiction. If firms were making larger profits on low efficiency models, they could increase their profits by offering only such models. This would correspond to a pooling 9Note that the analysis of the perfect competition case can lead to sharp predictions if we assume that firms face a small fixed cost to maintain each product on the market. In this scenario, the equilibrium prices will be set equal to total costs, and only products that maximize the consumer surplus would be sold. This is due to fact that firms would undercut each others by changing prices and energy efficiency levels to capture the market, and ultimately only products that benefit the most to consumers will remain. In particular, only products that maximize the difference between firms’ total costs and consumers’ valuation of energy efficiency will be offered. 19 equilibrium with bunching at the minimum standard. Remember that consumers that value highly energy efficiency would purchase a low efficiency model if Energy Star models were not offered. This is true because, in equilibrium, their incentive compatibility constraint (ICH) must be binding, i.e., they are indifferent between low efficiency and high efficiency models. In sum, higher markups on Energy Star models is a necessary condition to support a separating equilibrium. Decertified Products Exit the Market. When the Energy Star standard is revised and a product looses it Energy Star certification, the discontinuity in the demand curve due to Energy Star will be at the new Energy Star threshold (Figures 10). For models that just meet the old standard, this may imply that the product has no potential market, unless prices are reduced (Figure 10(a)), or the consumer surplus associated with this product becomes much smaller (Figure 11(b)). Even after adjusting prices, it still unlikely that it would be optimal for the firm to maintain a significant number of models that meet the previous standard. As shown earlier, it can be shown that there is a range of values [f i , f ES ], i = {L, H} for which the firm will always prefer to offer an Energy Star model, instead of maintain an interior solution. Given that change in the Energy Star standard usually follows small increment, i.e., 5%, models that meet the old standard are likely to fall in the range [f i , f ES ]. Decertification Leads to a Price Decrease. If following its decertification, a product does not meet the individual rationality constraints of any agents, it will have no potential market and remain unsold (Figure 10(a)). To liquidate its remaining inventory, its price will have to decrease so that at least the high valuation consumers may want to purchase the product. Decertification Leads to a Price Increase. When markups are not too large, decertified models may still bring a positive surplus for some consumers, even before prices are adjusted (Figure 11(b)). Post-certification, it is however unlikely that a separating equilibrium could be sustained without some re-optimization in prices. Following the decertification, consumers with a high valuation of energy efficiency will have a lower surplus associated with the decertified models, and might thus prefer to purchase a low efficiency model. Note that an increase in the equilibrium prices of the low efficiency models could ensure that the decertified models keep their market, and a separating equilibrium would be sustained. Following the 2008 revision, we observe something akin of this scenario. One question however remains. Why were firms able to increase prices of the low efficiency models in the post-revision period? If firms could increase prices of low efficiency models, they should have done so in the pre-revision period. To sustain a separating equilibrium, firms 20 should always try to set the highest price possible for the low efficiency models. Doing so, allow them to relax the incentive compatibility constraint of the high valuation consumers, and charge larger prices to them. To understand the price dynamics in post-revision period, we need also to account for the fact that firms also offered new models that meet the revised standards. To ensure that the new products has a market, firms needed to set prices so that the new models brought surplus that were at least as large that the ones brought by the lower efficiency models and the decertified models. In sum, the addition of new models led to new conditions for a separating equilibrium, which may have required an increase in the price of lower efficiency models. 5. Estimation In this section, I estimate the various primitives the model, namely the cost functions, and the demand system. The focus is however mostly on the cost estimation. The demand estimation is discussed in great length in a companion paper (Houde, 2012). 5.1. Cost Estimation A common approach to estimate firm costs is to use equilibrium conditions and observed prices to infer cost functions. Using the first order condition of the oligopoly problem with respect to price, we can solve for the costs c(fkl ), ∀kl, given an estimate of the demand function. Because we have Jk products, and c(fkl ) costs to estimate, no estimation is formally required, and all c(fkl ) can simply be found by solving a system of linear equations. Once c(fkl ) are obtained, we can regress the estimates on a vector of product attributes, to determine how the costs change as a function of energy efficiency. This two-step estimation procedure then allow to recover the marginal cost of providing energy efficiency. My data however allow me taking a more straightforward estimation strategy, given that I directly observe manufacturer prices, which are the costs faced by brand managers. Using that information, I propose two additional approaches to estimate the marginal cost of providing energy efficiency. I first revisit the matching estimator presented earlier. I also propose a simple hedonic model that regresses manufacturer prices on observed attributes. 21 5.1.1. Matching Estimator Revisited As discussed earlier, it is common for refrigerator manufacturers to offer product lines with several models that are quite similar and have few different attributes, such as their color, their finish (stainless or not) and their energy efficiency level. In particular, my sample contains 56 product lines with very similar refrigerator models that have different energy efficiency levels. I thus first estimate the cost of providing energy efficiency using variation in energy efficiency level within product line. For all refrigerator models matched (N=265), I regress the log of the weekly manufacturer price on a pair fixed effect, dummies for color, and on a proxy for energy efficiency: (9) ln(pricej,r,t ) = α + γj,j 0 + K X γ k Cjk + φfj + j,r,t , k where γj,j 0 is a product line fixed effect that is common to the matched refrigerator models j and j 0 , and Cjk are dummy variables that is equal to one if refrigerator j is of a given color. The proxy for energy efficiency, fj , is defined as the inverse of the annual electricity consumption. Note that the log-linear specification implies that the cost of providing energy efficiency is increasing and convex in fj if φ > 0, given that ∂c(fj ) ∂fj ) = φc(fj ) and ∂ 2 c(fj ) ∂ 2 fj ) = φ2 c(fj ). 5.1.2. Hedonic Regressions The second estimator uses information about observed attributes, but does not control for product line fixed effects. Note that I am observing one manufacturer price for each model in my sample. This approach allows me to use a large number of refrigerator models (N=3424). On the other hand, identification relies on cross-sectional variation, and is thus subject to an omitted variable bias. The extent of this problem is hard to assess, but it should be a source of concern. In my sample, not all models have the same information about attributes because manufacturers use different ways to describe their products. Keeping this caveat in mind, the second estimator that I propose is: (10) ln(pricej,r,t ) = α + βXj + φfj + j,r,t , 22 where Xj is a vector of attributes, other than energy efficiency.10 I first estimate equation 10 by OLS. I also use a more flexible specification, and estimate a generalized additive model (Ref), where the attribute size enters equation 10 with an unknown function that is estimated non-parametrically. For all estimators, I focus on estimating the average marginal cost of providing energy efficiency. Results. Table 7 presents the estimates for the various specifications. The first model presents the estimate of the marginal cost of providing energy efficiency obtained using the two-step estimation approach, where the costs were first estimated using the equilibrium conditions. The estimate of φ is positive, suggesting that costs are increasing and convex in energy efficiency. The matching estimator produces an estimate of similar magnitude, but he hedonic approach yields an estimate that is about four times smaller. Using the generalized additive model to control for size does not change the results substantially. The estimates from the hedonic models are also only marginally statistically significant. The matching estimator explains most of the variance (R2 = 96.2%). The hedonic model has a lower but still large, R2 of 87.4%. Figures 11(a) and 11 show respectively the predicted costs using the hedonic approach and the the two-step estimator. The hedonic approach does well on average, but the two-step estimator tends to over-estimate the costs, especially for large values. 5.2. Demand Estimation To account for heterogeneity in the way consumers value energy efficiency, I propose a framework that explicitly models consumers’ decision to collect and process energy information. In this framework, heterogeneity in the costs of collecting and processing energy information leads to heterogeneity in the heuristics consumers use to trade-off energy efficiency with other attributes. The proposed framework as a number of good properties. First, it is fully consistent with utility maximization, which allows me to conduct welfare analysis and quantify the change in consumer surplus. Second the structure of the model provides a theory of choice with and without Energy Star. The model thus allows me to simulate a counterfactual scenario without certification. Third, the model explicitly addresses the selection issue that arises in the estimation of consumers’ 10The following attributes are considered: dummies for refrigerator type (top-freezer, side-by-side or bottom-freezer), size interacted with dummies for refrigerator type, size interacted with dummies for overall quality (low: price <$1,000, medium: price >$1,000 & price <$2,800, high: price >$2,800), dummy for ice-maker, dummy for defrost technology (automatic vs. manual), dummy for water dispenser, dummy for an advance cooling technology, dummy for an air filtration technology, dummy for indoor lighting using LED, and dummies for brand. 23 valuation of energy efficiency. As shown by Bento et al. 2011, unobserved heterogeneity biases estimates of valuation of energy efficiency when left unaccounted for. My model is akin of the Heckman’s selection model, where the first step, the decision to get informed about energy efficiency, represents a latent decision to care or not about energy efficiency (the selection decision). The model proceeds as follows. Consider a discrete choice model where consumer i receives utility level Uij from purchasing an energy intensive durable j. The value of option j is a function of its quality (δj ), price (Pj ), expectation of annual energy cost (Cj ) and an idiosyncratic taste parameter (ij ): Uij = δj − ηPj − θCj + ij (11) In equation (11), the parameter η is the consumer’s sensitivity to prices and corresponds to the marginal utility of income, while θ is the sensitivity to energy cost. Cj is the product of energy price and energy consumption.11 I assume that with limited information a consumer will have few ways to make an accurate forecast of Cj for each option j in his choice set. Before collecting information, the consumer’s knowledge about Cj is then imperfect, and is modeled with a prior distribution F , where Cj ∼ F , ∀j. The consumer can collect energy information and learn the value of Cj for each j. In the context of purchasing a refrigerator, the process of information acquisition and learning may consist of taking the time to look at the EnergyGuide label, understand the various pieces of information, look up the electricity prices, and perform mental calculations to compute energy costs. Consumer’s purchasing decision is modeled as a two-step process. To begin, the consumer observes the quality δj and the price Pj for each refrigerator. Then, the consumer can collect and process energy information at a cost K. By doing so, the consumer observes a realization cj from F for each j. If the consumer does not collect/process information, no learning occurs and the consumer only knows the average energy cost for all refrigerator models, which is given by E[Cj ] = R C cf (c)dc = C, where f (c) is the probability density of F . Afterwards, the idiosyncratic taste parameters ij are realized12. Finally, the consumer decides which refrigerator to purchase. 11I assume that the consumer’s expectations about prices and utilization are constant, which is a reasonable assumption for my empirical application. 12As it is customary in the literature (e.g. Rust (1986)), I assume that the idiosyncratic taste parameters, ij , are only realized at the time of making a purchase, but the consumer has a prior on their distribution. 24 Consumer i will search for energy information if the following inequality holds: (12) −K + E,C max{Uij (δj , ηPj , Cj , ij )} ≥ E max{EC [Uij (δj , ηPj , Cj , ij )]} j j When the consumer does not collect and process energy information, the expected energy costs are the same for all products and do not influence the purchase decision. So the model offers a simple way to capture consumer inattention to energy costs. I next discuss how the impact of the Energy Star certification can be modeled in this framework. Energy Star as a Heuristic. The Energy Star certification partly informs about energy costs and can serve as a heuristic to compare products in a binary manner along the energy dimension. I model its effect as follows. Prior to collecting information, the consumer has some beliefs about energy costs and Energy Star. As before, the consumer’s belief about the energy cost of the j-th refrigerator is given by the prior distribution F , ∀j. Uncertainty related to the Energy Star certification stems from the fact that the Energy Star label carries no information about its meaning. A priori, the consumer is uncertain about the meaning of the Energy Star certification, but can learn it by searching for additional information. Moreover, I also assume that the consumer is also uncertain about whether a product is certified Energy Star or not. With probability q, the consumer believes that product j is certified. If the consumer knows the meaning of the Energy Star certification, he will then know that the energy costs of Energy Star products are below a particular threshold corresponding to the Energy Star standard. Denote this threshold by S. If the consumer does not collect energy information, his belief about the meaning of Energy Star, S, is given by a prior distribution G. If the consumer learns the meaning of the certification, he observes a realization s from G. Define the indicator variable Dj that takes the value one if product j is certified Energy Star and zero otherwise. For a given realization s, the expected energy cost of an Energy Star product j is: Z s 1 cf (c)dc ≡ C0 (s), (13) E[Cj (s)|Dj = 1] = E[Cj |Cj ≤ s] = F (s) c where c is the lower bound of the support of F . Similarly, the expected energy cost of a non-Energy Star product j 0 is given by: (14) 1 E[Cj 0 (s)|Dj = 0] = E[Cj 0 |Cj 0 > s] = 1 − F (s) Z c cf (c)dc ≡ C1 (s), s where c is the higher bound of the support of F . The timing is now as follows. The consumer first chooses his effort level for collecting and processing information, defined by the variable e, which now takes three values. A consumer that 25 selects e = l, does not collect energy related information. At e = m, the consumer only collects information related to Energy Star, and learns the meaning of the Energy Star and which products are certified. Finally, for e = h, the consumer collects and processes enough information to form expectations about the energy costs associated with each option. After choosing e, the idiosyncratic taste parameters ij are realized and the consumer decides which refrigerator to purchase. The optimal level of effort is given by the following optimization problem: (15) max V(e) e∈{l,m,h} V(e) = −K(e) + E,D,S,C max{Uij (δj , Pj , Cj (S), Dj , ij )}|I(e) , j where K(e) is the information acquisition cost that varies with e and I(e) represents consumer’s knowledge about energy costs at the time of purchase. To estimate the model, I assume that the idiosyncratic taste parameters are extreme value distributed. In addition, I assume that the information acquisition costs have an unobservable idiosyncratic component that is also Type I extreme value distributed. For a level of effort, e, the cost for consumer i is given by: Ki (e) = K e + β e Xi + νie , (16) where, Xi is a vector of demographics, and the constant K e and the vector of coefficients β e parameterize the average information acquisition cost. νi,e is a mean zero stochastic component of the cost and gives rise to closed form solutions for the probability of choosing effort e: (17) exp (−K e − β e Xi + E,C,S [maxj {Uijrt }|I(e)]) Qi (e) = P k k k exp (−K − β Xi + E,C,S [maxj {Uijrt }|I(k)]) At the moment of choosing the effort level, consumer i is uncertain about the electricity costs (Cj ), the meaning of the Energy Star label (S), the certification (Dj ), and the idiosyncratic taste parameters. Because is extreme value distributed, the expectation in (17) simplifies to: (18) E,D,S,C max{Uijrt }|I(e) = ED,S,C log j J X Uijrt } |I(e) j To evaluate the expectation in 18, I specify beliefs about electricity costs and Energy Star such that consumers have rational expectations. I will assume that consumers take the average electricity price in their county as given and have a prior on the annual amount of kWh consumed by 26 refrigerators. The prior consists of a distribution F normally distributed with a mean that matches the empirical distribution of electricity consumption of all the refrigerator models available on the US market, for the trimester the purchase was made. Note that because the choice set changes between trimesters, notably due to the revision in Energy Star standards, I effectively allow the prior on electricity consumption to vary across time. The variance of the normal distribution F will be estimated. Beliefs about the meaning of Energy Star, G, is modeled with a flat prior. Specifically, the Energy Star threshold (previously noted S) is uniformly distributed, with a support centered at the mean of F , the prior for electricity consumption. Ex ante, consumers are also uncertain about whether a product is certified or not. I set the prior for the probability that product j is certified Energy Star, q, equal to the share of products that are certified Energy Star in their choice set, i.e., their local store. Therefore, the probability that consumers believe that N products are Energy Star certified among the J products available is given by a binomial distribution with mean Jq. I thus effectively assume that consumers have rational expectations with respect to the number of Energy Star models in their choice set. Finally, I need to specify learning. I will assume that learning is unbiased and realized before the purchase decision. Upon deploying effort e = h, consumer i living in region r learns a realization cr,j from F that corresponds to the true value of the electricity consumption (kWh/year) of refrigerator j. Consumers will multiply this number by the average electricity price in region r, which is assumed to be known. If e = m, consumers learn the true the meaning of Energy Star, i.e., the realization s from G will be such that the expected electricity cost for all Energy Star refrigerator models (Cr,1 (s)) in a region r will correspond to the true average electricity cost of Energy Star models in this region. The same will be true for non-Energy Star models. I assume that ex ante consumers are completely unaware of the rebate programs available in their regions and learn the existence of the rebate program only if they collect and process energy information, i.e., if e = m or e = h. Therefore, under this assumption the existence of a rebate program in a particular region does not influence the level of effort to collect and process information, but does influence the purchase decision. I introduce a constant τ that corresponds to the Energy Star perceived added quality if a product j is certified Energy Star (Dj = 1). If consumers value Energy Star refrigerators beyond their electricity cost savings, this will be captured by the coefficient τ . The model thus allows me distinguishing whether consumers adopt Energy Star products purely based on financial or 27 non-financial motives. I will assume that τ is only realized at the time of making a purchase for consumers that search for energy information (i.e., e = m or e = h), is unexpected, and differ for e = m and e = h. For the estimation, I approximate the expectation in (18) with respect to the distributions of D, S and C with Monte Carlo integration. I then form the simulated likelihood and estimate the model via maximum likelihood. In the likelihood, the probability that consumer i chooses product j is given by: Hijrt = (19) X H̃ijrt (k) · Qi (k), k where H̃i,j,r,t (e) are the simulated choice probabilities conditional on the information set determined by the level of effort, e. Under the above assumptions about beliefs and learning, the alternative specific utility for each level of effort is given by: (20) h Uijrt = −ηPjrt + ψRrt XDjt + τ h Djt − θCjr + γj + ijrt m = −ηPjrt + ψRrt XDjt + τ m Djt − θEC [CrDj (s)] + γj + ijrt Uijrt l Uijrt = −ηPjrt − θED,S,C [CrDj (S)] + γj + ijrt Note that if e = l, the expectation ED,S,C [Cr,Dj (S)] is the same for all j, and is not influencing the choices. The model therefore assumes that consumers that do not search for energy information make purchase decision as if they were not paying attention to electricity costs and Energy Star. For consumers that rely on Energy Star information (e = m), the expectation EC [Cr,Dj (s)] takes only two possible values, and the difference EC [Cr,1 (s)] − EC [Cr,0 (s)] corresponds to the expected electricity cost savings, in region r, associated with Energy Star. The present specification accounts for two mechanisms by which the Energy Star certification influences choices. First, there is the purely financial motive, i.e., the fact that Energy Star products lead to electricity cost savings. Second, there is the effect of the label on consumers, which might both capture consumers’ willingness to pay for green products and/or higher perceived quality for energy efficient products. If consumers do not value Energy Star products beyond their electricity cost savings, we should expect that the parameter τ will take a value close to zero. 28 To account for additional heterogeneity, the model is estimated for three different income groups: (income < $50, 000, income ≥ $50, 000, < $100, 000, and income ≥ $100, 000). Further details about the estimation procedure can be found in appendix. Results. To account for the role of income, I estimated the information acquisition model for three different income groups. For comparison purpose, I have also estimated a simple multinomial logit model on the same data (see appendix). Table 8 presents the results. 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, |η|, decreases with income. Meanwhile, lower income consumers are also less sensitive to electricity costs. The effect of the Energy Star label varies across income levels. Consumers in the upper income group have the highest willingness to pay for Energy Star products, irrespectively of their beliefs about electricity costs. Looking at the effort probabilities (Table 9, constrained model), the estimated share of consumers that does not value Energy Star or electricity costs (e = l) is 34.7%. I find that about 46.2% of the consumers behave as if they were discounting electricity costs (e = h) and 19.0% of consumers rely on Energy Star (e = m). These probabilities vary by demographics. Table 9 reports these probabilities for the three income groups. Lower income group tend to have a high probability to have a low effort (e = l), while it is the opposite for the higher income group. For the low income group, the probability that they rely on Energy Star (e = m) is close to zero. The above numbers correspond to averages taken over the whole population of consumers. Overall, the information acquisition model produces estimates that are consistent with the hypothesis that there are different types of consumers with different levels of sophistication with respect to how they account for the energy efficiency attribute. Moreover, I find evidence that consumers with different income levels value energy efficiency differently, controlling for their degree of sophistication. 6. Counterfactual Scenarios In this section, I first illustrate how the model can replicate some of the main stylistic facts. I then perform sensitivity tests with respect to key parameters of the model. Finally, I simulate a number of counterfactual scenarios that show what would happen if the Energy Star program were not in effect, and provide estimates of the welfare effects of Energy Star. 29 For all scenarios, the size of the choice set is set equal to 75.13 To locate products in the quality space (all non-energy attributes), I use the estimated product fixed effects, obtained from the demand estimation, as the measure of observed quality. To create a representative choice set, I randomly sample 75 products from the 1065 products used in the demand estimation such that the distribution of the products in terms of brand, style, size, and energy efficiency fits the observed distribution for the whole US market for the year 2010. This approach ensures that the proportion of refrigerator models of a certain brand, style, size, and energy efficiency level in the constructed choice sets matches the proportion observed in the whole US refrigerator market. Table 10 compares the distribution of the products across the main attributes for the whole US market and the constructed choice set. For all simulations, consumers’ purchase decisions is simulated with the information acquisition model estimated in section 5.14 Quality (i.e., the product fixed effects) and the number of products offered by each brand is held constant. To account for uncertainty and the possibility of multiple equilibria, Monte Carlos are performed for each scenario considered. In this context, a Monte Carlo experiment consists in drawing the demand and cost parameters from their estimated distributions, and simulating the model for these particular parameter values; 10 experiments are performed for each scenario. Each experiment solves for the Nash equilibrium using the Gauss Seidel algorithm. To account for the discontinuity created by Energy Star, a non-gradient based optimization algorithm is used (simulated annealing). 6.1. Bunching The first question that I ask is whether the model can replicate the strong bunching at the minimum and Energy Star standards observed in the US refrigerator market (Figures 3 and 4). To simulate this scenario, I set the minimum energy standard to the standard enacted in 2001 (Table 1), and the Energy Star standard is 20% more energy efficient (the Energy Star standard since April 2008). The results suggest that the model replicates the main pattern observed. For the year 2010, 61% of the models on the market met the Energy Star standard (Table 10). For my preferred estimate 13In my sample, the average number of refrigerator models offered by a store is 250, on average. I set the size of the choice set to 75 for computational reasons. Although, I was able to solve the model for larger choice sets (e.g., 252), the simulation results were more robust and converged faster with smaller choice sets. Qualitatively, the results were the same with a larger choice set. 14Specifically, the information acquisition model that allows for heterogeneity with respect to income (Table 8). 30 of the marginal cost of providing energy efficiency (φ = 181), the model predicts that, on average, 69% of the products offered should meet the Energy Star standard in equilibrium.15 Although the model slightly over-predicts the proportion of models that meet the Energy Star standard, it performs well, however, in predicting bunching. Next, I assess the sensitivity of the results. Figure 12 compares the predicted distribution of energy efficiency to the distribution observed in the constructed choice set (75 models), for different values of the parameter φ. We observe that for lower estimates of φ, the model predictions differ substantially from the data, and the model poorly predicts bunching. For low marginal costs, firms will tend to offer much more energy efficient models, well above the Energy Star standard. 6.2. A World Without Energy Star I now turn to the main counterfactual scenario at the core of the welfare analysis. I remove the Energy Star standard and simulate the market. I simulate this scenario for three values of the estimate of the marginal costs of providing energy efficiency. The important result is that in a world without Energy Star, firms will offer products that bunch almost exclusively at the minimum standard (Figure 13) for the largest value of φ. For lower estimates, the results from the simple monopoly case extend to a more complex setting–in a world without Energy Star, we observe differentiation along the energy efficiency dimension. Clearly, we observe that in a world without Energy Star, firms may still offer highly energy efficient products. Moreover, without Energy Star, product differentiation in the energy efficiency dimension increases. The above findings have important implications for the design and the evaluation of Energy Star, especially to estimate the energy savings associated with the program. Previous analyses of Energy Star have assumed that in a world without Energy Star, all products would bunch at the minimum energy efficiency standard. The present results suggest that this scenario seems likely; without a voluntary certification, firms have few incentives to offer highly efficient refrigerator models. 6.3. Welfare Analysis I propose a welfare measure that accounts for the change in consumer surplus, producer surplus, and externality costs. Note that the EPA deploys substantial efforts to manage and market the Energy Star program. In the present analysis, I will not consider the public funds required to run 15I assume that a product bunches at a standard if it is within +/- 2.5% of the standard. 31 marketing campaigns, pay salaries of public employees and other operating expenses related to the program, with the caveat that this could be an important cost associated with the program. Consumer Welfare. According to the information acquisition model, it is important to note that for consumers that do not fully process energy information or rely on Energy Star, there is a discrepancy between the electricity costs consumers believe they would pay and the electricity costs they effectively pay. That is, the utility they experience differs from the utility they thought they would experience. Because the demand model estimated in Chapter 2 is a model that rationalizes observed choices, it should be thought as a model of decision utility. Using the demand model as is to conduct the welfare analysis will thus not reflect what consumers will truly experience ex ante. As analysts, this raises the question of what concept of utility should be used to measure consumer welfare. In the context of energy intensive durables, this problem has been previously pointed out by several researchers. Similarly to Allcott and Wozny (2011), I propose a measure of consumer surplus based on the notion of experience utility. I first make the following assumption. Assumption 1. If e = h, decision utility equals experience utility. Assumption (1) simply says that under perfect information consumers experience what they believed they will experience. Under this assumption and using the estimates obtained from the information acquisition model, the observed component of experience utility, net of the (observed) costs of collecting and processing information is thus: (21) OXUi,j,r,t = γ̂j + τ̂ h Dj,t − η̂Pj,r,t + ψ̂Rr,t XDj,t − θ̂Cj,r − K k − β k Xi , Note that whether the Energy Star label effect, τ h is truly experienced can be debated. If a consumer believes that a product is of higher quality because of the Energy Star label and this belief is never updated, the consumer may experience the perceived quality. It can also be argued that the costs of collecting and processing information should not be treated as part of the consumer surplus, because their identification relies on our assumption about consumers’ priors. In the present analysis, I will thus perform sensitivity analysis and provide estimates of the welfare effects with and without the effects of τ h and Ki (e). To deal with the unobserved idiosyncratic taste parameters i,j,r,t and unobserved component of the costs, νie , I will compute the expected consumer surplus (Train 2009). Note that the probability that consumer i chooses and thus experiences product j is given by the choice probabilities in equation (19). Following the estimation, these choice probabilities can be simulated and are our 32 best representation of how each consumer selects his favorite refrigerator. I then derive the expected consumer surplus (ESC) by computing the expected experience utility, where the expectation is taken with respect to the observed choice probabilities. For consumer i that lives in region r and makes a purchase at time t, the expected consumer surplus (ESCi,r,t ) is given by: (22) J X X 1 ECSi,r,t = E,ν Ĥi,j,r,t (k) · Q̂i (k) · (OXUi,j,r,t + i,j,r,t − νie ) η = j 1 η k J X X j Ĥi,j,r,t (k) · Q̂i (k) · (OXUi,j,r,t ) , k where I have used the facts that E [i,j,r,t ] = 0, Eν [Ki (e) = K e + β e Xi + νi,e ] = K e + β e Xi , and that the unobservable components of utility do not enter the choice probabilities. I also multiply by the inverse of the marginal utility of income to convert utils in dollars (Train 2009). Finally, note that the measure of the consumer surplus should be interpreted as the expected consumer surplus over the lifetime of the refrigerator. To obtain an annual measure, I compute the immediate-annuity assuming a lifetime of 18 years, and the discount rate implied by the estimates of marginal utility of income (η) and sensitivity to electricity costs (θ). Externality Costs. To quantify the externality costs associated with the electricity generated to operate refrigerators, I focus on the emissions of carbon dioxide (CO2 ), sulphur dioxide (SO2 ), and nitrous oxide (N Ox ), and use emissions factors recommended by the EPA. I compute the dollar damages associated with carbon dioxide using the recent estimates of the social cost of carbon recommended to assess federal regulations (Greenstone, Kopits, and Wolverton 2011). For sulphur dioxide and nitrous oxide, I rely on two sources, I consider the estimates used by the Department of Energy in the cost-benefit analysis of the 2014 minimum energy efficiency standards for refrigerators (DOE 2011), and the average estimates provided by Muller and Mendhelson (2012). Table 11 presents the emission factors and the damage costs used. I compute the externality cost associated with each scenario by taking the product of the corresponding average electricity consumption purchased, the market size, the emission factors, and the damage costs of electricity generation. The average electricity consumption purchased is the average of the electricity consumption of the refrigerators sold, weighted by market shares. For 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). 33 Producer Surplus. For each refrigerator model in the choice set, I compute the brand manager’s markup using the estimated cost functions. I compute the average profits by multiplying markups with the simulated market shares. Note that I do not consider the change in manufacturers’ profits in the producer surplus. Rebates. I assume that when Energy Star is in effect, consumers can claim a $50 rebate for purchasing an Energy Star refrigerator. The rebate program is offered by a government agency. When Energy Star is not in effect, I assume that this agency will distribute the amount reserved for the rebate program to all consumers via a lump-sum payment. The amount of the lump-sum payment is set equal to the rebate amount ($50) times the probability that a consumer takes advantage of the rebate amount, previously referred as π, implied by the estimates of the marginal utility of income (η) and sensitivity to rebates (ψ). 6.3.1. Results When Energy Star is not in effect, and firms offer products that bunch exclusively at the minimum standard (Scenario 1), the Energy Star program leads to important energy savings. Without Energy Star, the electricity consumption of a refrigerator purchased increases by 49 kWh/year, on average. Assuming a market size of 9.01 million refrigerators, the upper bound on energy savings implied by this estimate corresponds to a reduction of 441 GWh/year. Webber, Brown, and Koomey (2000) estimated that the overall energy savings associated to the program for the refrigerator market were about 1400 GWh/year, for the period 1997-2000. This larger estimate can be in part attributed to the fact that refrigerators were much less energy efficient on average before 2001, the last time the minimum energy efficiency standards for refrigerators was revised. For the range of estimates of the marginal utility of income and sensitivity to electricity costs that I obtain, I find that consumers are worst-off in a market with less efficient refrigerators. For the three measures of consumers surplus reported, consumers surplus decreases when Energy Star is not in effect. In this scenario, removing Energy Star also makes firms worst-off. Offering products that bunch exclusively at the minimum standard leads to a large loss of profits ($159 million annually). Under Scenario 2, removing the Energy Star program thus reduces social welfare. Using lower estimates of the marginal cost of providing energy efficiency, the welfare effects have a different sign. I find that without Energy Star, choices might even become more energy efficient. This highlights an unintended consequence of the program, which I call the crowding-out effect. When Energy Star is not in effect, products may not bunch exclusively at the minimum 34 energy efficiency standard, but instead become more differentiated. This increase in differentiation combined with the fact that more consumers are perfectly informed induces market shares for the most energy efficient products to increase, enough to have more energy efficient choices. 7. Conclusions In this paper, I show that manufacturers, brand managers, and retailers, respond to the Energy Star program. I propose a model that rationalizes the observed market outcomes and estimate the model using data for the refrigerator market. The model can predict the observed market outcomes quite closely. Using the estimated model, I show that if the refrigerator market were not subject to Energy Star, products would bunch almost exclusively at the minimum standard. References Allcott, H., and N. Wozny (2011): “Gasoline Prices, Fuel Economy, and the Energy Paradox,” Working paper, Massachusetts Institute of Technology. Bento, A. M., L. H. Goulder, M. R. Jacobsen, and R. H. von Haefen (2009): “Distributional and Efficiency Impacts of Increased US Gasoline Taxes,” American Economic Review, 99(3). Blumstein, C., B. Krieg, L. Schipper, and C. York (1980): “Overcoming Social and Institutional Barriers to Energy Conservation,” Energy, 5, pp. 355–371. Borenstein, S. (2010): “The Redistributional Impact of Non-Linear Electricity Pricing,” Working Paper 204, Energy Institute at Haas. Datta, S., and S. Gulati (2009): “Utility Rebates for Energy Star Appliances: Are they Effective?,” Working paper, University of British Columbia. DOE (2011): “Energy Conservation Program: Energy Conservation Standards for Residential Refrigerators, Refrigerator-Freezers, and Freezers,” Federal Register, 76(179). 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. Houde, S. (2012): “How Consumers Respond to Product Certification: A Welfare Analysis of Energy Star,” Working paper. 35 Ito, K. (2010): “Do Consumers Respond to Marginal or Average Price? Evidence from Nonlinear Electricity Pricing,” Working Paper 210, Energy Institute at Haas. 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. McWhinney, M., A. Fanara, R. Clark, C. Hershberg, R. Schmeltz, and J. Roberson (2005): “ENERGY STAR Product Specification Development Framework: Using Data and Analysis to Make Program Decisions,” Energy Policy, 33(12), 1613 – 1625. Muller, N. Z., and R. Mendelsohn (2012): “Efficient Pollution Regulation: Getting the Prices Right: Corrigendum (Mortality Rate Update),” American Economic Review, 102(1). Rust, J. (1986): “Structural Estimation of Markov Decision Processes,” in Handbook of Econometrics, ed. by R. F. Engle, and D. McFadden, vol. 4 of Handbook of Econometrics, chap. 51, pp. 3081–3143. Elsevier. Sallee, J. M. (2011): “Rational Inattention and Energy Efficiency,” Working paper. Train, K. (2009): Discrete Choice Methods with Simulation. Cambridge University Press. Webber, C., R. Brown, and J. Koomey (2000): “Savings Estimates for the Energy Star Voluntary Labeling Program,” Energy Policy, 28(15), 1137 – 1149. 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. 36 8. Figures and Tables Figure 1. Energy Star and EnergyGuide labels 37 Figure 2. The Three Main Manufacturers and their Major Brands. Circle with dates correspond to mergers and acquisitions. Mabe and GE entered a joint venture in 1986. Doted squares are the major brands produced by each manufacturer. 0 .1 .2 Density .3 .4 .5 38 0 10 20 30 Energy Efficiency: Percentage Better than Minimum Standard Figure 3. Empirical Density of Energy Efficiency for Full-Size Refrigerators, Year 2010 Energy Star standard: 20% better than the federal minimum standard. Sources: US EPA, Federal Trade Commission and California Energy Commission. 39 15 20 25 30 35 400 450 500 550 600 15 20 25 30 35 20 25 30 35 20 25 30 35 30 35 30 35 20 400 450 500 550 600 15 20 25 25 30 35 Standards 400 450 500 550 600 25 15 2010 400 450 500 550 600 20 25 400 450 500 550 600 15 2009 15 20 2008 400 450 500 550 600 15 15 2007 400 450 500 550 600 2006 Electricity (kwh/year) 2005 400 450 500 550 600 2004 400 450 500 550 600 2003 30 35 Minimum Energy Efficiency Standard ENERGY STAR: 2001 ENERGY STAR: 2004 ENERGY STAR: 2008 15 20 25 30 Adjusted Volume (Cu.Ft.) Figure 4. Choice Set for Bottom-Freezer Refrigerator without Ice-Maker: 2003-2010. Each dot represents a refrigerator model on the market. Most refrigerator models bunch at the minimum standard or Energy Star standard. Following revisions in the Energy Star standard, decertified models exit the market. Sources: US EPA, Federal Trade Commission and California Energy Commission. 35 40 Mark−Up: Promo. Price vs. Manuf. Price % Mark−Up Energy Star % Mark−Up Energy Star Mark−Up: MSRP vs. Manuf. Price 45 45 % Mark−Up Non−Energy Star % Mark−Up Non−Energy Star Figure 5. Markups on Energy Star Models vs. Non Energy Star Models Percentage markups for 56 product lines are compared. Each product line has at least one Energy Star model and one non Energy Star model. Panel A uses manufacturers’ suggested retail prices and manufacturers’ prices (retailers’ costs) to compute markups. Panel B uses promotional prices and manufactures’ prices. Both figures shows that the percentage markups of Energy Star models tend to be larger than the percentage markups of non Energy Star models. Revision Energy Star 1 Normalized Price 1.05 1.1 41 Non Energy Star Models .95 Decertified Energy Star Models: 15% Better than Minimum 2008w1 Energy Star Models: 20% Better than Minimum 2008w27 2009w1 Weeks 2009w26 2010w1 (a) Manufacturers’ Suggested Retail Prices .95 Normalized Price 1 1.05 1.1 Revision Energy Star Non Energy Star Models .9 Decertified Energy Star Models: 15% Better than Minimum 2008w1 Energy Star Models: 20% Better than Minimum 2008w27 2009w1 Weeks 2009w26 2010w1 (b) Promotional Prices Figure 6. Prices Before and After the 2008 Revision of Energy Star Each panel displays average normalized weekly prices, with 5% confidence intervals, of refrigerators that belong to different efficiency classes. Three efficiency classes are considered: models that were not certified Energy Star before and after the revision (less than 15% more efficient than the minimum standard), models that lost the Energy Star 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.) The normalized price for each model is computed by divided its weekly price (MSRP or promotional) by its average weekly price for the pre-revision period (January 1st , 2008 to April 28th , 2008). The average normalized price and standard errors in each efficiency class are computed by fitting a cubic spline on normalized prices. Refrigerators with no observed sales after November 15th , 2009 are not considered. This criterion ensures that refrigerators that were decertified but exited the market are excluded from the analysis. 42 1.05 MSRP Energy Star Models .85 .9 Normalized Price .95 1 Decertified Energy Star Models 2009w13 2009w40 2010w13 2010w40 Weeks (a) Manufacturers’ Suggested Retail Prices 1.05 Promo. Price Energy Star Models .95 Normalized Price 1 Decertified Energy Star Models 2009w13 2009w40 2010w13 2010w40 Weeks (b) Promotional Prices Figure 7. Prices Before and After the 2010 Decertification Due to Problematic Test Procedures Each panel displays average normalized weekly prices with a 5% confidence interval. Average normalized prices are computed using the same methodology than for Figure 6. Prices of the 16 models that were decertified are compared to the prices of Energy Star models that did not loose their certification. 43 p p H θ f c(f) {f*H,p*H} c(f) {f*L,p*L} H θ f πH L θ f L θ f {f*, p*} πL H θ =c ' ( f ) 2 π0 L θ =c ' ( f ) f f (a) Separating Equilibrium (b) Pooling Equilibrium Figure 8. Possible Equilibria without Energy Star; Monopoly Selling to Two Consumer Types p H θ f p H θ f {f*H,p*H} τ τ {f*H,p*H} c(f) o τ {f*L,p*L} c(f) o πH L θ f τ o L θ f o {f*L,p*L} πL πH πL fES (a) Separating Equilibrium f fES f (b) Special Case, Separating Equilibrium Figure 9. Possible Equilibria with Energy Star; Monopoly Selling to Two Consumer Types 44 p Decertified Model τ o H θ f {f*H,p*H} c(f) πH p H θ f Decertified Model L θ f τ {f*L,p*L} (a) High Markup Case c(f) L θ f o πL fES fES' o τ o {f*L,p*L} τ πL fES fES' f (b) Low Markup Case Figure 10. Effect of Decertification f 45 Estimates Marginal Cost ($) Data (Manufacturer Prices) 0 200 400 600 800 1000 Product Id (a) Estimated Marginal Costs w. Hedonic Approach vs. Observed Costs (Manufacturer Prices) Second Stage Predicted Marginal Cost ($) Data (Manufacturer Price) 0 200 400 600 800 1000 Product Id (b) Estimated Marginal Costs w. Two-Step Estimation vs. Observed Costs (Manufacturer Prices) Figure 11. Predicted Costs (ĉ(f )) 0 0 .1 .05 Density Density .1 .2 .3 .15 46 0 10 20 30 Energy Efficiency: Percentage Better than Minimum Standard 40 0 10 20 30 Energy Efficiency: Percentage Better than Minimum Standard 40 (b) Predicted, High Marginal Cost: φ = 181 Density .02 0 0 .01 .01 Density .02 .03 .03 .04 .04 (a) Data 0 10 20 30 40 50 Energy Efficiency: Percentage Better than Minimum Standard 60 (c) Predicted, Medium Marginal Cost: φ = 90 0 10 20 30 40 50 60 70 Energy Efficiency: Percentage Better than Minimum Standard 80 (d) Predicted, Low Marginal Cost: φ = 40 Figure 12. Distribution of Energy Efficiency with Energy Star, Observed and Predicted 0 .5 Density 1 1.5 2 47 0 10 20 30 Energy Efficiency: Percentage Better than Minimum Standard 40 0 .05 Density .1 .15 .2 (a) High Marginal Cost: φ = 181 0 10 20 30 Energy Efficiency: Percentage Better than Minimum Standard 40 0 .02 Density .04 .06 .08 (b) Medium Marginal Cost: φ = 90 0 10 20 30 40 50 60 70 Energy Efficiency: Percentage Better than Minimum Standard 80 (c) Low Marginal Cost: φ = 40 Figure 13. Distribution of Energy Efficiency without Energy Star, Predicted 48 49 Table 1. Minimum and Energy Star Standards for Full-Size Refrigerators Refrigerator Type Refrigerator-freezers and refrigerators other than allrefrigerators w manual defrost All-refrigerators manual defrost Refrigerator-freezers partial auto. defrost Refrigerator-freezers auto. defrost w top-mounted freezer w/o auto. icemaker Built-in refrigerator-freezer auto. defrost w topmounted freezer w/o auto. icemaker Refrigerator-freezers auto. defrost w top-mounted freezer w auto. icemaker w/o through-the-door ice Built-in refrigerator-freezers auto. defrost w topmounted freezer w auto. icemaker w/o through-thedoor ice All-refrigeratorsŮauto. defrost Built-in All-refrigerators auto. defrost Refrigerator-freezers auto. defrost w side-mounted freezer w/o auto. ice-maker Built-In Refrigerator-freezers auto. defrost w sidemounted freezer w/o auto. icemaker Refrigerator-freezers auto. defrost w side-mounted freezer w auto. icemaker w/o through-the-door ice Built-In Refrigerator-freezers auto. defrost w sidemounted freezer w auto. icemaker w/o through-thedoor ice Refrigerator-freezers auto. defrost w bottom-mounted freezer w/o auto. ice-maker Built-In Refrigerator-freezers auto. defrost w bottommounted freezer w/o auto. icemaker Refrigerator-freezers auto. defrost w bottom-mounted freezer w auto. ice-maker w/o through-the-door ice Built-In Refrigerator-freezers auto. defrost w bottommounted freezer w auto. icemaker w/o through-thedoor ice Refrigerator-freezer auto. defrost w bottom-mounted freezer w through-the-door ice Built-in refrigerator-freezer auto. defrost w bottommounted freezer w through-the-door ice Refrigerator-freezers auto. defrost w top-mounted freezer w through-the-door ice Refrigerator-freezers auto. defrost w side-mounted freezer w through-the-door ice Built-In Refrigerator-freezers auto. defrost w sidemounted freezer w through-the-door ice Minimum Energy Efficiency Standard Effective Effective Effective January 1993 January 2001 Starting 2014 13.5AV+299 8.82AV+248.4 7.99AV + 225.0 10.4AV+398 16.0AV+355 NA NA 8.82AV+248.4 9.80AV+276.0 6.79AV + 193.6 7.99AV + 225.0 8.07AV + 233.7 NA NA 9.15AV + 264.9 NA NA 8.07AV + 317.7 NA NA 9.15AV + 348.9 NA NA 11.8AV+501 NA NA 4.91AV+507.5 7.07AV + 201.6 8.02AV + 228.5 8.51AV + 297.8 NA NA 10.22AV + 357.4 NA NA 8.51AV + 381.8 NA NA 10.22AV + 441.4 16.5AV+367 4.60AV+459.0 8.85AV + 317.0 NA NA 9.40AV + 336.9 NA NA 8.85AV + 401.0 NA NA 9.40AV + 420.9 NA 5.0AV+539.0 9.25AV + 475.4 NA NA 9.83AV + 499.9 17.6AV+391 10.20AV+356.0 8.40AV + 385.4 16.3AV+527 10.10AV+406.0 8.54AV + 432.8 NA NA 10.25AV + 502.6 Note: AV: adjusted volume measured in cubic feet. For all-refrigerator, AV is calculated as follows: AV = 1.63 x total refrigerator volume. For refrigerator-freezer, AV is calculated as follows: AV= fresh volume + 1.63 x total freezer volume. Energy Star Effective April 2008 -20 % 50 Table 2. Market Share by Manufacturer Manufacturer GE Electrolux Whirlpool Maytag (Amana) Haier W.C. Wood Other 1995 Market Share 2000 2005 2008 35% 17% 27% 10% 10% 0% 0% 1% 34% 21% 24% 14% 5% 0% 0% 2% 29% 25% 25% 11% 0% 2% 1% 7% 27% 23% 33% 6% 1% 10% Source: Appliance Magazine; data compiled by the Department of Energy. Table 3. Full-Size Refrigerator Models Offered by Brand, Californian Market Brand Manufacturer(s) Kenmore GE, Electrolux, Whirlpool GE Whirlpool Whirlpool Whirlpool General Electric Kitchen Aid Amana Maytag Whirlpool Frigidaire White-Westing. LG Others Total Electrolux Electrolux 1995 Nb of Model Models Share 2000 Nb of Model Models Share 2005 Nb of Model Models Share 2008 Nb of Model Models Share 2010 Nb of Model Models Share 130 0.08 241 0.14 569 0.17 875 0.17 975 0.19 219 91 131 0.13 0.05 0.08 122 70 70 0.07 0.04 0.04 125 92 72 187 95 294 0.07 0.05 0.04 0.11 0.05 0.17 163 210 119 534 355 455 0.05 0.06 0.03 0.16 0.10 0.13 409 300 157 606 528 621 0.08 0.06 0.03 0.12 0.10 0.12 286 328 175 477 597 782 0.05 0.06 0.03 0.09 0.11 0.15 0.30 1 165 1369 5185 0.03 0.27 1 197 1267 5227 0.04 0.24 1 828 1798 0.50 1 632 1859 0.36 1 1028 3702 Source: California Energy Commission (CEC) Appliance Database. Only full-size refrigerators models on the Californian market for each year are considered. Model shares correspond to the number of models, non-sales weighted, offered by each brand. Manufacturer(s) for each brand as of January 2012. 51 Table 4. Summary Statistics: Product Lines Used for Matching Estimator Nb of Product Lines Nb of Products Percentage Energy Star MSRP Non Energy Star Models (mean, std) MSRP Energy Star Models (mean, std) Promo. Price Non Energy Star Models (mean, std) Promo. Price Energy Star Models (mean, std) Percentage Stainless Steel Percentage with Icemaker Nb Top Freezer Nb Bottom Freezer Nb Side-by-Side 56 265 65% $1209 $1315 $1122 $1238 20% 68% 164 49 52 $755 $744 $708 $719 Table 5. Matching Estimator: Difference in Markups Energy Star vs. Non Energy Star Dependent Variable EnergyStarj = 1 Product Line FE Week FE Markup Markup Markup vs. Manuf. Price) (Promo Price vs. Manuf. Price) (MSRP vs. Manuf. Price) (Promo Price vs. Manuf. Price) 0.019∗ (0.0082) Yes No 0.030∗∗ (0.0089) Yes No 0.019∗ (0.0082) Yes Yes Markup (MSRP 0.030∗∗ (0.0088) Yes Yes Nb of Product Lines 56 56 56 56 Nb of Observations Adjusted R2 17009 0.547 17009 0.496 17009 0.548 17009 0.505 Note: For all regressions, the explanatory variables are product line fixed effects and a dummy variable that takes the value 1 if product j is certified Energy Star. The first column shows that within a given product line, the markup between the manufacturer suggested retail price (MSRP) and manufacturer price paid by the retailer is 7.4 percentage point higher. The second column shows that the markup between the promotional price and the manufacturer price is 11 percent point higher. Adding week fixed effects has a marginal impact on the estimates. Clustered standard errors in parentheses. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 52 Table 6. Price Change After Decertification of Energy Star Models Dependent Variable: Pt P̄T af ter =0 j T af tert = 1× Decertif iedj = 1 T af tert = 1 Decertif iedj = 1 2008 Revision in Energy Star Standard MSRP Promo. MSRP Promo. (I) (II) (III) (IV) -0.025∗∗∗ (0.0015) 0.035∗∗∗ (0.0011) -0.00064 (0.0012) Week FE Nb of Models Nb of Decertified Models Nb of Observations Adjusted R2 -0.065∗∗∗ (0.0040) -0.018∗∗∗ (0.00058) -0.000055 (0.0029) -0.019 (0.044) 0.044∗ (0.017) -0.00027 (0.020) EnergyStarj = 1 2010 Revision in Certification MSRP Promo. Price MSRP Promo. (V) (VI) (VII) (VIII) -0.0031∗∗∗ -0.015 -0.025∗∗∗ (0.0057) 0.0052∗∗∗ (0.00082) 0.00053 (0.0041) 0.036∗∗∗ 0.012∗∗∗ No No (0.00062) Yes (0.0097) Yes No No (0.0020) Yes (0.0028) Yes 1,905 1,196 1,905 1,196 1,905 1,196 1,905 1,196 1,515 16 1,515 16 1,515 16 1,515 16 80,237 80,237 80,237 80,237 52,759 52,759 52,759 52,759 0.034 0.000 0.015 0.000 0.031 0.001 0.064 0.098 Note: For all regressions, the dependent variable is the weekly price divided by the average price before the decertification took place. For the first four specifications, only refrigerator models that were not certified Energy Star (control) and had lost their Energy Star certification (treatment) are considered. For the fifth to eighth specifications, only refrigerator models that were certified Energy Star as of January 1st are considered. For 2008, I assume that the decertification of Energy Star models occurred in the 17th week. For 2010, I assume that the decertification occurred in the 5th . The dummy variable T af tert takes the value one for all weeks after the decertification and zero otherwise. The dummy variable Decertif iedj takes the value one for refrigerator models that lost their Energy Star certification. 53 Table 7. Estimates: Marginal Cost of Providing Energy Efficiency Model 1 2nd Stage Eq. Conditions Model 2 Pair FE Model 3 Hedonic GAM Model 4 Hedonic w. 181.3 (85.3) 0.127 (.026) 171.4 38.4 - 41.8 (20.3) 0.057 (0.003) 40.1 26.8 - Product Line FE No Yes No No Product Attributes Yes No Yes Yes 999 0.877 265 0.962 3424 0.870 3424 0.874 Dependent Variable: log(Manufacturer Price) Efficiency (1/kWh): φ̂ Adjusted Volume (cu.ft.) Nb of Models R2 Note: All specification shows that the manufacturer prices are increasing with energy efficiency level. Energy efficiency is the inverse of the annual electricity consumption of a refrigerator (1/kwh). Model 1 uses the estimated manufacturers prices obtained from the first order conditions of the oligopoly model as a dependant variable. Only refrigerator models used for the demand estimation are considered. Models 2-4 uses the manufacturer prices observed in the data. Model 2 uses only products that belong to similar product lines and control for product line FE. Models 3 and 4 uses all refrigerator models observed in the sample suring the period 2008-2010. Models 1, 3, and 4 use the following regressors to control for product attributes: dummies for refrigerator type (top-freezer, side-by-side or bottom-freezer), size (adjusted volume) interacted with dummies for refrigerator type, size interacted with dummies for overall quality (low: price <$1,000, medium: price >$1,000 & price <$2,800, high: price >$2,800), dummy for ice-maker, dummy for defrost technology (automatic vs. manual), dummy for water dispenser, dummy for an advance cooling technology, dummy for an air filtration technology, dummy for indoor lighting using LED, and dummies for brand. Model 4 uses a generalized additive model to control for the effect of size (adjusted volume), and thus model the effect of size with a flexible non-parametric function. 54 Table 8. Heterogeneity w.r.t. Income All Income Income <$50,000 ≥$50,000 & <$100,000 Income ≥$100,000 Multinomial Logit Price (η̂) -0.459 (0.005) Energy Star (τ̂ ) .155 (0.029) Rebate (ψ̂) 0.060 (0.036) Electricity Costs (θ̂) -2.267 (0.146) -0.589 (0.008) .056 (0.032) 0.053 (0.048) -0.513 (0.160) -0.483 (0.007) .114 (0.038) 0.085 (0.043) -1.343 (0.197) -0.404 (0.006) .187 (0.038) -0.013 (0.048) -2.045 (0.191) -0.574 (0.001) -.036 (0.020) .261 (0.096)(0 0.023 (0.173) -3.441 (0.211) -0.480 (0.001) .085 (0.031) .286 .074) 0.110 (0.057) -3.090 (0.175) -0.400 (0.001) .125 (0.027) .497 (0.037) 0.010 (0.051) -3.410 (0.183) Information Acquisition Model Price (η̂) -0.460 (0.001) h Energy Star (τ̂ ) .079 (0.020) Energy Star (τ̂ m ) .778 (0.154) Rebate (ψ̂) 0.096 (0.045) Electricity Costs (θ̂) -3.870 (0.100) Note: The top panel corresponds to the estimates of the multinomial logit estimated for different income groups. The estimates in the first column are from the base model with all income groups, and are presented for comparison purpose. The bottom panel presents the estimates from the information acquisition model. 55 Table 9. Effort Choice Probabilities: Information Acquisition Model All e=h e=m Income <$50,000 e=l e=h e=m e=l Income ≥$50,000 & <$100,000 e=h e=m e=l Income ≥$100,000 e=h .614 - .431 17.1 7.3 .268 .301 21.8 -1.5 ¯ 2 ×kW h .564 5.8 1.0 .362 .073 0.0 -0.2 ¯ 2 ×kW h .473 .135 .392 -0.2 -27.9 5.4 -2.4 ¯h 1.8 ×kW .595 0.0 5.1 .229 .176 -0.1 1.2 ¯h 4.5 ×kW e=m e=l Unconstrained Model Q(e) K(e) ($) VoI ($) σbelief s .528 123.4 78.2 .241 .231 .384 .002 6.3 -311.7 -3824.3 -12.5 11.6 -35.4 ¯ ¯h 2 ×kW h 2 ×kW Constrained Model: Variance of Beliefs Estimated Q(e) K(e) ($) VoI ($) σbelief s .462 .190 .347 .218 .194 .588 -244.0 -184.7 -174.5 -170.5 331.7 65.9 1.2 -27.6 ¯h ¯h 4.2 ×kW 0.3 ×kW 2 Note: In the unconstrained model the variance of the prior beliefs (σbelief s ) for the distribution of electricity consumption (referred as F ) is set exogenously. The prior F is assumed to be normally distributed with a mean equals to the empirical mean of electricity consumption of all refrigerators in the sample, the standard deviation is set to be equal to two times the empirical mean. Information acquisition costs (K(e)) are unconstrained, i.e., they can take positive or negative values. Costs are positive for the model estimated on transactions with all households, and the two upper income groups, this violates the rational model. This suggests that the variance of the beliefs is too small to rationalize the choice probabilities. In the constrained model, the behavioral parameters that enter the purchase probabilities are set equal to the estimates obtained in the unconstrained model and treated as data, costs are estimated but are constrained to be negative, and the standard deviation of the prior beliefs is estimated. The standard deviation is set equal to a scaling factor times the empirical mean of electricity consumption of all refrigerators in the sample. The last row of the table presents the estimates of the scaling factor. For all models, the information costs vary as a function of demographic information: income, education, age of the head of the household and family size. The choice probability for each effort level are computed for all consumers in the subsample. Only the mean across the whole sub-sample is reported.The value of information (VoI) for e = h is given by E[maxj Uj |I(e = h)] − E[maxj Uj |I(e = l)], and for e = m by E[maxj Uj |I(e = m)] − E[maxj Uj |I(e = l)]. The value of information is positive for e = h, but tend to be negative for e = m. 56 Table 10. Summary Statistics: US and Constructed Choice Sets Attribute US Market 2010 (1,580 Models) Nb of Models Model Share (%) Constructed Choice Set (75 Models) Nb of Models Model Share (%) Brand Frigidaire GE Kenmore Whirlpool 268 313 237 292 16.96 19.81 15.00 18.48 6 13 31 7 10.00 17.34 41.34 9.33 Others 470 29.75 18 24.00 1,063 517 67.28 32.72 37 38 49.33 50.67 472 597 511 29.87 37.78 32.34 12 19 44 16.00 25.33 58.67 Energy Star Certification No 623 Decertified Model 21 Yes 957 39.43 1.30 60.64 22 53 29.23 70.67 Size AV< 23.5 cu. ft. AV≥ 23.5 cu. ft. Refrigerator Style Top Freezer Side-by-Side Bottom-Freezer For the US market, only full-size refrigerators that were on the market for the year 2010 are considered. Sources: California Energy Commission, Federal Trade Commission, and US EPA. 57 Table 11. Emission Factors and Externality Costs Non-baseload Output Emission Rates (U.S. Average) Pollutant Estimate Source CO2 1,583 lb/MWh CH4a 35.8 lb/GWh N 2Oa 19.9 lb/GWh USEPA, eGRID2007 SO2 6.13 lb/MWh N Ox 2.21 lb/MWh Damage Cost (2008 $) Pollutant Low Estimate High Estimate Source CO2 $21.8/t $67.1/t Greenstone, Kopits, and Wolverton (2011) SO2 $2,060/t $6,700/t low: Muller and Mendelsohn (2012), high: USEPAb N Ox $380/t $4,591/t low: Muller and Mendelsohn (2012), high: DOEc Note: (a) Externality costs associated to CH4 and N 2O are assumed to be the same than for CO2. CH4 and N 2O are converted in CO2 equivalent using estimates of global warming potential (GWP). The GWP used for CH4 is 25, and the GWP used for N 2O is 298. Source: IPCC Fourth Assessment Report: Climate Change 2007. (b) Estimate used in the illustrative analysis of the 2012 regulatory impact analysis for the proposed standards for electric utility generating units. (c) Higher value of the estimate used in the Federal Rule for new minimum energy-efficiency standards for refrigerators (1904-AB79). 58 Table 12. The Effects of Removing the Energy Star Standard Scenario 1 w Firms’ Responses Oligopoly Model Scenario 2 w Firms’ Responses Oligopoly Model Scenario 3 w Firms’ Responses Oligopoly Model Avg kWh Purchased (kWh/year) ∆withoutES−withES Mean Std 49.13 27.96 ∆withoutES−withES Mean Std -6.57 5.09 ∆withoutES−withES Mean Std -1.93 9.22 Externality Cost Low Damage (M$/year) 10.68 6.08 -1.43 1.11 -0.42 2.00 Externality Cost High Damage (M$/year) 35.16 20.01 -4.70 3.64 -1.38 6.60 Avg Price Paid ($) 14.24 30.94 3.74 5.09 10.07 3.81 Consumer Surplus (M$/year) -96.65 55.74 -69.79 12.17 -92.35 9.45 Consumer Surplus w/o Search Cost (M$/year) Consumer Surplus w/o Search Cost & w/o Label Effect (τ ) (M$/year) Profits (M$/year) -64.71 51.24 -23.80 5.35 -35.48 8.79 -59.91 48.88 -12.25 5.05 -22.84 8.68 -159.31 225.23 63.04 43.85 159.23 34.38 Welfare Low Damage (M$/year) -266.64 181.08 -5.32 36.87 67.30 32.68 Welfare High Damage (M$/year) -291.12 176.28 -2.05 37.46 68.26 31.83 Welfare Low Damage, w/o -234.69 183.63 40.67 42.39 124.17 29.48 Search Cost (M$/year) Welfare High Damage, w/o -259.17 178.58 43.94 43.32 125.13 28.80 Search Cost (M$/year) Welfare Low Damage, w/o Search Cost -229.89 184.54 52.22 42.72 136.81 29.42 & w/o Label Effect (τ ) (M$/year) Welfare High Damage, w/o Search Cost -254.37 179.31 55.50 43.63 137.77 28.71 & w/o Label Effect (τ ) (M$/year) Note: For each scenario, 10 simulations of the demand model were performed. Each simulation takes a random draw of parameter values from the estimated distributions. For each simulation, the differences between the metrics obtained for the state of the world without Energy Star and the state of the world with Energy Star are computed. For each scenario, the first and second columns report respectively the mean and the standard deviation of the difference in various metrics for the 10 simulations. The externality costs are computed for two estimates of the dollar value of the damage associated to electricity generation. The market size for refrigerators is assumed to be 9.01 millions. The consumer surplus is converted to an annual measure by assuming that consumers will own (and believe they will own) their refrigerators for 18 years. The total welfare is the sum of the total consumer surplus, externality costs, and produced surplus. The last four rows report different estimates of the welfare effects. The welfare estimates without search costs exclude the cost of collecting and processing information from the consumer surplus. The last two rows present welfare estimates where both the search costs and the label effect, captured by the parameter τ are excluded. All dollar figures are in 2008 dollars. 59 Appendix A. Proofs: Screening Problem forthcoming Appendix B. Data Cleaning and Manipulation Demand Estimation. The primary data for the demand estimation were provided by a large appliance retailer. The retailer offers a large selection of refrigerator models, has at least one brick-and-mortar store in each US state and a national online store. The main data consist of all transactions where a full-size refrigerator was bought. The data cover the period 2008-2010. For each transaction, I observe the date, the model of the refrigerator, attributes, the suggested retail price, the promotional price, the manufacturer price,16 taxes paid, and the zipcode of the store where the transaction was made. For a subset of transactions, I also observe consumer demographics, such as household size, income, education, homeownership, housing type and age of the head of the household. I restrict attention to transactions with demographic information and that can be attributable to households that have to pay for their electricity bills. In particular, I only consider transactions made by homeowners living in single family housing units that bought no more than one refrigerator in any given year. This rules out heterogeneity in the sensitivity to energy costs due to the split incentive problem (Blumstein, Krieg, Schipper, and York 1980), i.e., the fact that some consumers of energy intensive durables do not pay for energy costs. For the structural estimation, a random sub-sample of the transactions that fit the above criteria is used. I next discuss the construction of the key variables used in the estimation, and their main sources of variation. To perform the estimation of the demand model, a random sub-sample of the transactions is used. The sub-sample is constructed as follow. First, the sub-sample is drawn from the set of transactions that fit the following criteria (the restricted sample): • transactions made by consumers that are homeowners; • transactions made by consumers living in single family housing units; and • transactions made by consumers that made no more that one refrigerator purchase in any given year. Second, I employ the following stratified sampling method to create the sub-sample. For a given targeted sample size, I sample transactions in each state so that the state market shares in the sub-sample are equal to the state market shares in the restricted sample. Sampling is done at the store level. That is, I first randomly select a store in each state and then keep a number of transactions from this store, also randomly selected, to match state market shares. If needed, I sample additional stores until I match state market shares. The main motivation to sample at the store level is to restrict the number of choice sets that need to be imputed. Moreover, if one is concerned about endogeneity problems that could be eliminated with store fixed effects, the present approach allows me to limit the number of fixed effects to estimate. Average Electricity Prices. The use of average electricity prices is partly motivated by recent empirical evidence (Borenstein (2010), Ito (2010)) that suggests that electricity consumers may in fact respond to variation in average prices, more than marginal prices. In the present case, the use of average electricity 16 The manufacturer price does not vary across time, and corresponds to the manufacturer price the retailer paid when a given refrigerator model entered the market. 60 prices is also dictated by the fact that household’s location is not perfectly known. Therefore, it is impossible to match households with their exact electricity tariff and infer marginal price. Average electricity prices at the county level are computed as follow. Using form EIA-861 of the Energy Information Administration, I compute the average residential electric price for each electric utility operating in the US for the years 2008. I then match electric utility territories with each of the county where I sampled at least one store. For counties with only one electric utility, I use the average electricity price for this particular utility. For counties with several electric utilities, I take the arithmetic mean of each utility average price to construct the county level price. Ideally, we would like to weight prices by the number of consumers served by each utility. However, this information is not available at the county level. Appendix C. Demand Estimation forthcoming