American Economic Journal: Economic Policy 2014, 6(4): 207–238 92936 http://dx.doi.org/10.1257/pol.6.4.207 Cash for Coolers: Evaluating a Large-Scale † Appliance Replacement Program in Mexico  * By Lucas W. Davis, Alan Fuchs, and Paul Gertler  This paper evaluates a large-scale appliance replacement program in Mexico that from 2009 to 2012 helped 1.9 million households replace their old refrigerators and air conditioners with energy- efficient models. Using household-level billing records from  the universe of Mexican residential customers, we find that refrigerator replacement reduces electricity consumption by 8 percent, about one-quarter of what was predicted by ex ante analyses. Moreover, we find that air conditioning replacement actually increases electricity consumption. Overall, we find that the program is an expensive way to reduce externalities from energy use, reducing carbon dioxide emissions at a program cost of over $500 per ton. (JEL L68, L94, O12, O13, Q41, Q54) E nergy consumption is forecast to increase dramatically worldwide over the next several decades, raising important concerns about energy prices, geopolitics, and greenhouse gas emissions. Much of the recent energy research has focused on transportation and the demand for gasoline (Knittel 2011; Mian and Sufi 2012; Busse, Knittel, and Zettelmeyer 2013; Allcott and Wozny forthcoming). However, an equally important area is residential energy consumption. This category makes up 14 percent of total energy use worldwide, and is expected to grow by 57 percent through 2040 (EIA 2013a). Meeting this increased demand represents a severe challenge from both an eco- nomic and environmental perspective. To curtail demand use and the associated nega- tive externalities, policymakers are increasingly turning to energy-efficiency programs as a politically palatable alternative to first-best approaches. Supporters of energy-effi- ciency policies argue that they represent a win-win, reducing e ­ xternalities while also *  Davis: Haas School of Business, University of California, Berkeley, CA 94270, Energy Institute at Haas, and National Bureau of Economic Research (e-mail: ldavis@haas.berkeley.edu); Fuchs: The World Bank, Poverty Global Practice, 1818 H Street, NW I 4-405, Washington, DC 20433 (e-mail: afuchs@worldbank.org); Gertler: Haas School of Business, University of California, Berkeley, CA 94270, and National Bureau of Economic Research (e-mail: gertler@haas.berkeley.edu). We are thankful to Judson Boomhower for excellent research assistance and to two anonymous referees, Catie Hausman, Paul Joskow, Paulina Oliva, Catherine Wolfram, and numerous seminar participants for comments that substantially improved the paper. The research assistant for this project was sup- ported in part under a general research contract from the California Energy Commission to the Energy Institute at Haas. The authors have not received any financial compensation for this project nor do they have any financial relationships that relate to this research. The findings, interpretations, and conclusions expressed in this work do not necessarily reflect the views of the World Bank, its Board of Executive Directors, or the governments they represent. The World Bank does not guarantee the accuracy of the data included in this work. †  Go to http://dx.doi.org/10.1257/pol.6.4.207 to visit the article page for additional materials and author disclosure statement(s) or to comment in the online discussion forum. 207 208 American Economic Journal: economic policyNOVEMBER 2014 helping participants reduce energy expenditures. Much of the push for these programs is based on estimates from ex ante analyses that assume no behavioral response.1 In this paper, we evaluate the impact and cost-effectiveness of a large-scale appli- ance replacement program in Mexico. Between 2009 and 2012, Cash for Coolers— henceforth, C4C—provided subsidies to 1.9 million households to help them replace their old refrigerators and air conditioners with newer, more energy-efficient models. To participate in the program, a household’s old appliance had to be at least ten years old and the household had to purchase an energy-efficient appliance of the same type. These old appliances were then transported to recycling centers to be disassembled. We find that refrigerator replacement reduces electricity consumption by an aver- age of 11 kilowatt hours (kWh) per month, an 8 percent decrease. This is a substan- tial decrease, but is considerably less than what was predicted ex ante by the World Bank and McKinsey (Johnson et. al 2009; McKinsey and Company 2009b). The World Bank study, for example, predicted savings for refrigerators that were about four times larger than our estimates. And while these same studies predicted even larger savings from air conditioner replacement, we find that electricity consump- tion actually increases after households receive a new air conditioner. We then present ancillary evidence supporting several behavioral responses to the program which help explain why our estimated savings are so much smaller than the ex ante predictions. Part of the explanation is that the ex ante predictions were overly optimistic about the program being able to recruit households with very old, very inefficient appliances. In practice, we find that most of the retired appliances were less than 12 years old. Another important explanation, especially for air con- ditioners, is increased usage. More energy-efficient air conditioners cost less to use, which leads households to use them more. This pattern of usage is reflected in our estimates, with near zero changes in electricity consumption during winter months and substantial increases in the summer. Finally, we illustrate how modest increases in appliance size and added features like side-by-side doors and through-the-door ice can substantially offset improvements in energy-efficiency. This paper helps address an urgent need for credible empirical work in this area. Allcott and Greenstone (2012) explain that “much of the evidence on the energy cost savings from energy-efficiency comes from engineering analyses or observa- tional studies that can suffer from a set of well-known biases.” They then go on to say, “We believe that there is great potential for a new body of credible empirical work in this area, both because the questions are so important and because there are significant unexploited opportunities for randomized control trials and quasi- experimental designs that have advanced knowledge in other domains.” Our paper is one of the first studies of an energy-efficiency program in a low- or middle-income country.2 Many low and middle-income countries are now adopting energy-efficiency policies. For example, development of energy-efficient appliances 1  McKinsey and Company (2009a), for example, uses ex ante analyses to argue that energy-efficiency invest- ments are a “vast, low-cost energy resource” that could reduce energy expenditures by billions of dollars per year. 2  The small existing literature on energy-efficiency is focused mostly on the United States. See, for example, Dubin, Miedema, and Chandran (1986); Metcalf and Hasset (1999); and Davis (2008). There is also a related ­ literature which uses utility-level data to evaluate energy-efficiency programs, again mostly in the United States (Joskow and Marron 1992; Loughran and Kulick 2004; Auffhammer, Blumstein, and Fowlie 2008; Arimura et al. 2012). Vol. 6 No. 4 davis et al.: Cash for Coolers 209 is one of the major initiatives of the Clean Energy Ministerial, a partnership of more than 20 major economies, aimed at promoting clean energy.3 And China recently announced a new large-scale program that will provide subsidies for energy-efficient refrigerators and air conditioners. In part, these policies reflect a widely held view that there is an abundant supply of low-cost, high-return investments in energy-effi- ciency, particularly in low- and middle-income countries (Zhou, Levine, and Price 2010; Johnson et al. 2009; McKinsey and Company 2009b). Most global growth in energy consumption over the next several decades is expected to occur in low- and middle-income countries. Between 2010 and 2040, total energy consumption is predicted to increase by 90 percent in non-OECD countries, compared to only 18 percent in OECD countries (EIA 2013a, Table 1). Many policymakers believe that e­ nergy-efficiency programs can be an effective tool for curtailing this growth in demand. But without credible empirical estimates of program impacts it is impos- sible to know how large a role energy-efficiency can play. A key feature of our analysis is the use of high-quality microdata. For this analysis we were granted access to household-level electric billing records for the universe of more than 25 million Mexican residential customers. The large number of house- holds in our analysis allows us to estimate effects precisely even with highly flexible specifications. In contrast, the primary source of data used in most previous research on energy-efficiency programs in the United States comes from s ­elf-reported measures of energy savings from utilities. Economists have long argued that these ­ self-reported measures of energy savings are overstated (Joskow and Marron 1992). ­ The fact that our analysis is based on a large-scale national program gives our results an unusually high degree of intrinsic policy interest. Program evaluation, particularly with energy-efficiency policies, is typically based on small-scale interventions imple- mented in one particular location. In these settings a key question is external validity: i.e., how well do parameter estimates generalize across sites. Utilities that choose to participate in these programs tend to be considerably different from the population of utilities, raising important issues of selection bias (Allcott 2014). The format of the paper is as follows. Section I provides background informa- tion about the electricity market in Mexico and the C4C program. Sections II and III describe the data, empirical strategy, and main results. Section IV compares our estimates to the ex ante predictions, presenting ancillary evidence indicating several important explanations for the smaller than expected savings. Section V evaluates cost-effectiveness, calculating the implied cost of the program per unit of energy savings, and Section VI offers concluding comments. I. Background A. Context and Program Rationale The Mexican Federal Electricity Commission (Comisión Federal de Electricidad or CFE) is the exclusive supplier of electricity within Mexico. CFE is ­ responsible 3  See http://www.cleanenergyministerial.org/ and http://superefficient.org/ for details. 210 American Economic Journal: economic policyNOVEMBER 2014 Table 1—Demographics and Appliance Saturation in Mexico, Census 2000–2010 2000 census 2005 census 2010 census Demographics:   Total population (in millions) 97.0 102.8 112.0   Total number of households (in millions) 22.6 24.7 28.7   Household size (persons) 4.3 4.2 3.9   Household head completed high school 26.8% 29.6% 32.1%   Number of rooms in home 4.32 4.19 4.58   Improved flooring 86.0% 89.2% 93.9% Electricity and appliance saturation:   Electricity in the home 94.7% 96.4% 97.5%  Refrigerator 68.2% 79.1% 82.5%   Washing machine 51.6% 63.0% 67.0%  Television 85.6% 90.9% 92.6%  Computer 9.2% 19.9% 30.0% Notes: This table describes data from the Mexican National census Censo de Poblacion y Vivienda from the years indicated in the column headings. These statistics were compiled by the authors using microdata from the long-form survey which is completed by a 10 percent representative sample of all Mexican households. All statistics are calculated using sampling weights. We have cross-checked total population, number of households, and appliance satura- tion at the national and state level against published summary statistics and the measures cor- respond closely. Improved flooring includes any type of home flooring except for dirt floors. for most electricity generation and all electricity transmission and distribution. Electricity service in Mexico is highly reliable, with total service interruptions per household averaging just over one hour per year (CFE 2011, Table 5.14). Residential customers are billed every two months. The standard residential tar- iff in Mexico is an increasing block rate with no monthly fixed fee and three tiers. Residential electricity consumption is subsidized. As of August 2011, customers on the first-tier (tariff 1) paid 0.73 pesos (US$ 0.057) per kilowatt hour. The second and third tiers are more expensive: 1.21 pesos (US$ 0.096) and 2.56 pesos (US$ 0.202) per kilowatt hour, respectively. As a point of comparison, the average retail price paid by residential customers in the United States is $0.117 (EIA 2013b). The Mexican Energy Ministry estimates that residential customers face a price that is, on average, about one-half the average cost of providing power (SENER 2008). Table 1 describes demographics, electricity, and appliance saturation in Mexico. In the 2010 census, 97.5 percent of households reported having electricity in their homes. Electricity consumption per capita in Mexico is 1,900 kilowatt hours annu- ally, compared to 14,000 for the United States (World Bank 2013). Over the next several decades, electricity consumption in Mexico is forecast to increase 3.7 percent per year, more than triple the increase in the United States (EIA 2013a, 98). One of the major drivers of this increase in demand is the continued increase in residen- tial appliance ownership, due to poverty reduction and economic growth. Figure 1 plots ownership rates for televisions, refrigerators, and vehicles by income level in Mexico. As incomes increase households first acquire televisions, then refrigerators and other appliances, and it is not until income reaches substantially higher levels that households acquire vehicles (Wolfram et al. 2012; Gertler et al. 2013). Meeting this increased energy demand will require an immense investment in generation and transmission infrastructure. The Mexican Energy Ministry has cal- culated that $80 billion will need to be invested in new electricity generation and Vol. 6 No. 4 davis et al.: Cash for Coolers 211 1 Share of households with asset 0.8 Television 0.6 Refrigerator 0.4 0.2 Car or truck Histogram for income 0 100 1,000 10,000 100,000 Annual household income (2010 US$) Figure 1. Durable Good Ownership Rates by Income Level in Mexico transmission infrastructure between 2012 and 2026 (SENER 2012, 157). Energy- efficiency programs are viewed by policymakers as one of the ways to potentially reduce these looming capital expenditures. Part of the broader goal of our analysis is to consider whether energy-efficiency programs like C4C could serve as a substitute for these capital-intensive investments. The program was implemented, in part, because ex ante analyses had predicted that appliance replacements would lead to substantial decreases in electricity con- sumption. In independent studies of available energy-related investments in Mexico, the World Bank and McKinsey concluded that replacing residential refrigerators and air conditioners would be extremely cost-effective (Johnson et al. 2009; McKinsey and Company 2009b). In fact, both reports calculated a negative net cost of carbon abatement for these investments. That is, these were found to be investments that would pay for themselves even without accounting for carbon dioxide emissions or other externalities. At the heart of these predictions are optimistic predictions about the amount of electricity saved per replacement. We revisit these predictions later in the paper, contrasting them with the results from our empirical analysis. B. Program Details The C4C program was in place between March 2009 and December 2012. Unlike the US Cash for Clunkers program, the program was never viewed as an economic stimulus program. The objective of the program was to reduce electricity consump- tion and thereby reduce carbon dioxide emissions and other negative externalities. This was a national program. The only geographic requirement was that participants in the air conditioner replacement program had to live in a warm climate zone. This excluded 75 percent of Mexican households, including all households living 212 American Economic Journal: economic policyNOVEMBER 2014 in Mexico City, Guadalajara, Puebla, and other high-elevation areas. There were no geographic restrictions for refrigerator replacement. To participate in the program, a household needed to have a working refrigera- tor or air conditioner that was at least ten years old and to agree to purchase a new appliance of the same type (i.e., refrigerator or air conditioner). The old appliances were transported to government-financed recycling facilities and disassembled. The new appliances were required to meet national minimum energy-efficiency stan- dards and, in the case of refrigerators, to exceed standards by at least 5 percent. In addition, the new appliances had to meet certain size requirements. For example, refrigerators were supposed to be between 9–13 cubic feet, and with a maximum size no more than 2 cubic feet larger than the refrigerator which was replaced. The program provided direct cash payments in three amounts, approximately cor- responding to $30, $110, and $170. Retailers could charge $30 for delivering the new appliance and taking away the old one, reducing the net subsidy amounts to $0, $80, and $140. Eligibility for these different payment levels depended on a house- hold’s average historical electricity consumption. Households with very low levels of historic consumption were ineligible for the program. This minimum requirement was implemented in an attempt to prevent participation by households with non- working appliances. Above this threshold, households qualified for the $170 pay- ment, while households with higher levels of historic consumption received smaller payment amounts. This structure of decreasing payments was implemented out of distributional concerns in an attempt to avoid large cash payments to high-income households. More than three-quarters of participants qualified for the most generous $170 payment. In addition to the cash payments, the program offered on-bill financ- ing at a 14 percent annual interest rate, repaid over four years. Households could accept the cash payment, the on-bill financing, or both. In practice, all participants choose to accept the cash payments, but many participants decided not to accept the on-bill financing. From the households’ perspective, the program represented a substantial incen- tive for appliance replacement. Program participants paid an average of $427 per refrigerator, and $406 per air conditioner, so the cash payments represented a large share. Another nice feature of the program from the households’ perspective is that they received these subsidies immediately, with virtually no paperwork required. In order to participate, a household was required to show a recent electricity bill. The retailer then determined which subsidy a household was eligible for by entering the household’s account number into a website designed for this purpose. This differs from many appliance subsidy programs elsewhere in the world which require par- ticipants to fill out and mail application forms and proofs of purchase, and then wait for a rebate check to arrive in the mail. From the perspective of appliance manufacturers and retailers, the program rep- resented a large increase in demand. Data is not available to directly examine the incidence of the subsidy, but several factors lead us to believe that the benefits to manufacturers and retailers would have come primarily in the form of increased sales rather than increased prices. Appliance manufacturing and retailing are highly com- petitive in Mexico. There are at least ten manufacturers with a nonnegligible mar- ket share and a similar number of large national retailers. Moreover, m ­ ultinational Vol. 6 No. 4 davis et al.: Cash for Coolers 213 appliance manufacturers like GE, LG, Samsung, and Daewoo have a significant presence in Mexico and the global manufacturing capacity to adjust supply quickly in response to demand increases. C. Participation Between 2009 and 2012, the program provided subsidies for 1.9 million appli- ance replacements. About 90 percent of all replacements were refrigerators. The lower level of participation in the air conditioner program reflects the geographic restrictions and the fact that air conditioning is relatively uncommon in Mexico. In the 2010 ENIGH survey (Encuesta Nacional de Ingresos y Gastos de los Hogares), only 13 percent of households nationwide reported having air conditioners. In part, this low saturation reflects that many Mexicans live in the highland central pla- teau. Mexico City, for example, is located at 7,300 feet and has a mild climate year round. But even in warmer areas of Mexico, households are much more likely to own refrigerators than air conditioners, meaning that there were many more eligible participants for refrigerator replacement. The program reached a substantial fraction of all eligible households nationwide. With refrigerators, for example, Arroyo-Cabañas et al. (2009) estimate that as of 2009 there were approximately 23 million total refrigerators owned nationwide. Of these, they calculate that about 10 million (43 percent) were ten or more years old. By the end of the program, therefore, about 17 percent of all eligible refrigerators had been replaced. The program appears to have had a substantial impact on refrig- erator sales. During 2009, 2010, and 2011 there were 6.8 million refrigerators sold in Mexico.4 Based on pre-2009 data from Arroyo-Cabañas et al. (2009) we would have predicted 5.4 million sales. This yields a difference of 1.4 million refrigera- tors, similar to the total number of refrigerators replaced through C4C in those three years. This back-of-the-envelope calculation is based on a linear extrapolation and does not control for macroeconomic conditions. If anything, however, one would have expected the recession post-2008 to decrease sales relative to the trend. II.  Data and Empirical Framework A. Data Description The central dataset used in the analysis is a two-year panel dataset of household- level electric billing records. These data describe bimonthly electricity consumption for the universe of Mexican residential customers from May 2009 to April 2011. The C4C program was in place during this entire period. Each record includes the customer account number, county and state of residence, climate zone, tariff type, and other information. For confidentiality reasons, these data were provided without 4  This number comes from personal correspondence with the Mexican National Association of Electric Manufacturers (Cámara Nacional de Manufacturas Eléctricas, CANAME). Based on their own internal analysis of national-level sales data, CANAME concludes that C4C has generated through March 2012 a total of 900,000 additional refrigerator sales and 160,000 additional sales of air conditioners (both about 60 percent of total C4C replacements). 214 American Economic Journal: economic policyNOVEMBER 2014 customer names. The complete set of billing records includes data from 26,278,397 households. We dropped 15,262 households (< 0.001 percent) for whom the records are improperly formatted and 1,113 households for whom no state was indicated. We also drop 491,788 observations (1.9 percent) with 0 reported usage in every month of the panel. Residential customers are billed every two months using overlapping billing cycles. One-half have their meters read during odd-numbered months (e.g., January, March, etc.) and one-half have their meters read during even-numbered months (e.g., February, April, etc.). So for most households there are six billing cycles per year, and twelve billing cycles over the two-year sample period. There are also a small number of households with irregular billing cycles. The average number of months per billing cycle is 1.98 months, with 93 percent of all cycles representing 2 months. An additional 5 percent represent one month, with the remaining 2 per- cent representing 3+ months. These irregular billing periods arise for a variety of reasons. For example, some households in extremely rural areas have their meters read less than six times per year. We assign billing cycles to calendar months based on the month in which the cycle ends. And we normalize consumption to reflect monthly consumption by dividing by the number of months in the billing cycle. Thus, for example, a typical July observation reflects average monthly consumption during June and July. Equally important for the analysis is a second dataset which describes C4C par- ticipants. These data describe all participants in the program between March 2009 and June 2011, a total of 1,162,775 participants. Thus our program data cover the first 28 months in which the program was in place, a period during which approxi- mately 60 percent of the total replacements occurred. We dropped 51,823 par- ticipants (4.5  percent) for whom no installation date for the new appliance was recorded. We merged the remaining data with the billing records using customer account numbers. We were able to match 86 percent of C4C participants with identi- cal account numbers in the billing records. Each record in the program data includes the exact date in which the appliance was replaced, whether the appliance replaced was a refrigerator or an air conditioner, the amount of direct cash subsidy and credit received by the participant, the reported age of the appliance that was replaced, and other program information. We drop 93 households (< 0.0001 percent of partici- pants) who replaced more than one air conditioner, leaving us with 957,080 total treatment households. We do not have data on other forms of energy use. This would matter much more if this were an energy-efficiency program aimed at home heating or cooking. In those cases, households are able to substitute between electricity, natural gas, bottled gas, and other energy types. With refrigerators and, in particular, air condi- tioners, most of the available substitutes also use electricity, and our estimates will reflect the net change in electricity consumption from all end-uses. This is not to say, however, that we are able to describe the full range of possible energy impacts of the program. For example, it could be that better refrigerators and air conditioners lead households to spend more time at home, driving less, and eating fewer meals outside the home. The estimated change in electricity consumption will reflect changes in the amount of time spent at home, but not these other impacts. Vol. 6 No. 4 davis et al.: Cash for Coolers 215 B. Empirical Strategy This section describes the estimating equation used for our estimates of the effect of refrigerator and air conditioner replacement on household electricity consump- tion. The basic approach is difference-in-differences. In the preferred specification, impacts are measured by comparing electricity consumption before and after appli- ance replacement using a rich set of time effects that vary across locations. Our empirical approach is described by the following regression equation,   ​y​  = ​ it​ β1​​1[New Refrigerator​ ​​ ]it β2  + ​ ]it ​​1[New Air Conditioner​ ​​ + ​ γi​, moy​ ωt​​  + ​ εit  + ​ ​​ , where the dependent variable y ​​ it​is electricity consumption by household i in month t measured in kilowatt hours. The covariates of interest are 1[New Refrigerator​ ]​ it​and 1[New Air Conditioner​ ​​ , indicator variables equal to 1 for C4C participants after ]it they have replaced their refrigerator or air conditioner. For replacements that occur in the middle of a billing cycle, we assign a value between 0 and 1 equal to the proportion treated. Parameters β ​1​​and β​​ 2​measure the mean change in electricity consumption associated with appliance replacement. Our preferred specifications include household by month-of-year fixed effects, γi​,moy​ ​ . That is, for each household we include 12 separate fixed effects, one for each calendar month. This controls not only for time-invariant household characteris- tics such as the size of the home, but also household-specific seasonal variation in electricity demand. For example, some households have air conditioning and some do not, so electricity demand varies differentially across the year for different households. The billing data includes identifiers for both the household and the housing unit. Consequently, we can observe when a new household moves into an existing hous- ing unit. This is a nice feature because one might expect participation in the program to be correlated with the decision to move. In the empirical analysis we treat each household by housing unit pair as a separate household. Thus with household by month-of-year fixed effects we are identifying the effects of C4C using only house- holds which remain in a housing unit for at least one year. All estimates also include month-of-sample fixed effects ​ ωt​​ . This controls for month-to-month differences in weather as well as for population-wide trends in electricity consumption. Many specifications include, instead, month-of-sample by county fixed effects. This richer specification controls for county-specific variation in weather, as well as differential trends across counties. Finally, the error term ​ ε​ it​ captures unobserved differences in consumption across months. In all results we cluster standard errors at the county level to allow for arbitrary serial correlation and correlation across households within counties. A potential concern for this empirical strategy is the possibility that participating households might have experienced other changes in their household at the same time they replaced their refrigerator or air conditioner. Participation in the program might systematically tend to coincide with, for example, other events like the arrival of a new baby, a household member receiving a new job, or the decision to purchase additional appliances. We are able to construct an event study figure and to report 216 American Economic Journal: economic policyNOVEMBER 2014 estimates from specifications that control flexibly for time trends, so the real con- cern is about changes that occur exactly at the same time as appliance replacement. Although it is impossible to rule out this concern completely, another test we can perform is to compare estimates by calendar month. For households who replace air conditioners, we find little change in consumption during nonsummer months, sug- gesting that these households did not simultaneously purchase additional appliances or make other changes that affect baseload consumption. And for households who replace refrigerators, we find similar effects across months of the year, suggesting that households did not purchase air conditioners, fans, or other types of cooling equipment simultaneously. C. Comparison Groups We report regression estimates based on several different comparison groups. We first report results estimated using an equal-sized random sample of ­nonparticipating households. Next we report results estimated using a sample that includes ­participating households only. In this specification the participating households who have not yet replaced are the comparison group, and we can continue to include time effects in these regressions because households replaced appliances at different times. Finally, we report estimates from a set of regressions that are estimated using matching. We consider two different matched samples. The first matched sample is based purely on location. We perform this matching using account numbers. Account num- bers include codes for the state and county where each household lives, as well as an internal code indicating the specific route used by meter readers. We do not have access to the route maps, and thus cannot use these codes to identify where within a county each household lives. But in selecting a comparison group, we can take advantage of the fact that households with the same meter reading route tend to live in close geographic proximity. For each C4C participant, we select as a comparison household the closest consecutive nonparticipating account number. In almost all cases this is another household on the same meter reading route. Weather is a major determinant of electricity consumption so this matching ensures, for example, that comparison households are experiencing approximately the same weather as the treatment households. Our second matched sample is constructed based on both location and ­pre­treatment electricity consumption. We are somewhat limited in that we only have two years of data, and thus in many cases do not have a large number of pretreatment observa- tions for electricity consumption. To ensure the best possible matches given this lim- itation, we match on all available pretreatment months. For example, if a household replaces in November 2010, we match using all observations between May 2009 and October 2010. When matching on both location and pretreatment consumption level we adopt the following approach. We first select for each participating house- hold the ten nonparticipating households with the closest account numbers. Then among these ten we select the nonparticipating household whose average monthly pretreatment consumption is closest to that of the participating household. For a small number of households (< 2 percent) we have zero months of pretreatment consumption and for these households we match on location only. Vol. 6 No. 4 davis et al.: Cash for Coolers 217 Figures 2A and 2B plot electricity consumption by month of the year for house- holds who replaced refrigerators and air conditioners and for the three compari- son groups. Notice that the scale for the y-axis is not the same in both figures and that the overall level of consumption is considerably higher among households who replaced their air conditioners. For participants, consumption averages 153 kilowatt hours per month in Figure 2A and 395 kilowatt hours per month in Figure 2B. There is a great deal of variation across households and months; the standard deviation of monthly observations is 110 in Figure 2A and 300 in Figure 2B. These figures provide an opportunity to assess the different comparison groups. For households who replaced their refrigerators, all three comparison groups fol- low patterns that are reasonably similar to participating households. However, for households who replaced air conditioners, nonparticipants do not appear to be a particularly good comparison group, with electricity consumption levels that are much lower and less seasonal. The matched comparison groups perform better, and in particular, the pattern for the match based on both location and pretreatment consumption is very similar on average to the treatment group. These matched sam- ­ ples help address potential concerns that nonparticipating households, as a whole, may not be a good comparison group. Households are self-selecting into the C4C program, and thus are likely to be different from nonparticipating households. Most importantly, they may have fundamentally different tastes for durable goods, and thus different trajectories for electricity consumption. Although we do not explicitly observe durable good holdings, matching on pretreatment electricity consumption is likely to be a good proxy.5 This is particularly true because we are matching also by location, and thus the matched households experience the same climate and are living in the same neighborhoods. Nonetheless we are acutely aware that this is nonexperimental data and thus pay great attention in the section which follows to possible differential trends in electricity consumption. These figures also provide an opportunity to perform an informal inventory of the key drivers of residential electricity consumption in Mexico. For participants in the air conditioner program, electricity consumption triples during the summer, implying that about two-thirds of summer consumption (and one-half of annual consumption) come from air conditioners and other cooling equipment. It seems clear that most of these households indeed had operating air conditioners prior to participation; otherwise you would not expect to see such a pronounced seasonal pattern. Winter consumption averages 140 kilowatt hours per month for participants in the refrigerator program and 200 kilowatt hours per month for participants in the air conditioner program. A typical 15-year-old refrigerator uses about 60 kilowatt hours per month (see Section IVA), so refrigerators represent between one-third and one-half of winter consumption. Other important sources of nonsummer e ­ ­ lectricity consumption include lighting, televisions, washing machines, microwaves, and elec- tric stoves, though none of these end-uses is as important as refrigerators (Gertler et al. 2013). The relative importance of both refrigerators and air conditioners helps explain why the program targeted these appliances. 5  Reiss and White (2005), for example, show that electricity consumption is determined to a large degree by durable good holdings. 218 American Economic Journal: economic policyNOVEMBER 2014 220 200 220 (kWh) (kWh) 180 200 consumption 160 180 consumption electricity 140 160 electricity 120 140 Participants Average Nonparticipants 100 120 Nonparticipants matched on location Participants Average and pretreatment consumption Nonparticipants Nonparticipants matched on location 80 100 Nonparticipants matched on location and pretreatment consumption January March May July September November Nonparticipants matched on location 80 Figure 2A. Comparing Participants to Nonparticipants: Refrigerators 800January March May July September November Participants 800 (kWh) (kWh) Nonparticipants Nonparticipants matched on location Participants 600 and pretreatment consumption consumption Nonparticipants Nonparticipants matched on location Nonparticipants matched on location 600 and pretreatment consumption consumption Nonparticipants matched on location 400 electricity 400 electricity Average 200 Average 200 0 January March May July September November 0 January March May July September November Figure 2B. Comparing Participants to Nonparticipants: Air Conditioners Notes: These figures plot average electricity consumption by calendar month for house- holds who replaced their refrigerators and air conditioners through the C4C program (“par- ticipants”), households who didn’t participate in the program (“nonparticipants”), and for two matched samples of nonparticipants. For all households the sample is restricted to observa- tions from the first year of the program (May 2009–April 2010). Additionally, for participants the sample is limited to those who participated during the second year of the program (May 2010–April 2011). This restriction ensures that the means for participating households are from before replacement. Vol. 6 No. 4 davis et al.: Cash for Coolers 219 III.  Main Results This section presents estimates of the effect of appliance replacement on elec- tricity consumption. We present estimates from a range of different specifications. We start in Section IIIA with a graphical event study approach. Section IIIB then presents the baseline results, estimated with and without comparison households. And Section IIIC presents alternative specifications including matching estimates using our two matched samples and estimates that include polynomial time trends. Overall, the results are very similar across approaches. A. Graphical Results This subsection presents graphical results intended to motivate the regression analyses that follow. We focus in this section on refrigerators rather than air condi- tioners because they make up 90 percent of all replacements and because refrigerators lend themselves better to an event study analysis. Whereas the effect of refrigerator replacement is expected to be relatively similar across months of the year, the effect of air conditioner replacement is not. You would not expect to see, for example, much impact of air conditioner replacement on winter electricity consumption. This seasonal pattern, combined with the fact that air conditioner replacements tended to occur during warm months, makes evaluating air conditioner replacement better suited for a regression context. Figure 3 describes graphically the effect of refrigerator replacement on household electricity consumption. The x-axis is the time in months before and after refrigera- tor replacement, normalized so that the month prior to replacement is equal to zero. The figure plots estimated coefficients and ninety-fifth percentile confidence inter- vals corresponding to the effect of appliance replacement by month, controlling for household and county by month-of-sample fixed effects. In particular, we plot the estimates of α from the following regression, 12   ​yi​t​ ∑ ​  =  ​     ​​ ​α​1[​ k​ ]it  = k​ τi​t​ ​​ γ​  + ​  + ​ i​ ωc​ εi​  + ​ t​ t​ , k=−12 where ​τ​it​denotes the event month defined so that τ = 0 for the exact month in which the refrigerator is delivered, τ = −12 for 12 months before replacement, τ = 12 for 12 months after replacement, and so on. The coefficients are measured relative to the excluded category (τ = −1). Both sets of fixed effects play an important role here. Without the county by month-of-sample fixed effects (​ω​ct​ ), for example, the effect of replacement could be confounded with seasonal effects or slow-moving county-specific changes in residential electricity consumption. The sample used to ­ estimate this regression includes the complete set of households who replaced their refrigerators and an equal number of nonparticipating households matched to the treatment households using location and pretreatment consumption. During the months leading up to replacement, electricity consumption is flat, suggesting that the fixed effects are adequately controlling for seasonal effects and underlying trends. Beginning with replacement, electricity consumption falls 220 American Economic Journal: economic policyNOVEMBER 2014 10 5 Kilowatt hours per month 0 −5 −10 −12 −10 −8 −6 −4 −2 0 2 4 6 8 10 12 Months before and after replacement Figure 3. The Effect of Refrigerator Replacement on Household Electricity Consumption Notes: This figure plots estimated coefficients and ninety-fifth percentile confidence intervals describing monthly electricity consumption before and after refrigerator replacement. Time is normalized relative to the delivery month of the appliance (t = 0) and the excluded category is t = −1. Observations from before t = −12 and after t = 12 are dropped. The sample includes 858,962 households who received new refrigerators through C4C between March 2009 and May 2011 and an equal number of nonparticipating comparison households matched to treat- ment households using location and pretreatment consumption. The regression includes house- hold and county by month-of-sample fixed effects. Standard errors are clustered by county. sharply by approximately 10 kilowatt hours per month. Consumption then continues to fall very gradually over the following year. We attribute the fact that the decrease appears to take a couple of months to the fact that the underlying billing cycles upon which this is based are bimonthly, and to a modest amount of measurement error in the replacement dates. Moreover, the gradual decline between months +2 and +12 likely reflects a modest differential time trend between the treatment and compari- son households. In all periods the coefficients are estimated with enough precision to rule out small changes in consumption in either direction. With Figure 4 we perform the same exercise but assigning event study indica- tors to the comparison group, rather than the treatment group. For this figure, we assigned hypothetical replacement dates equal to the replacement date of the par- ticipating household to which each comparison household is matched. The figure exhibits no change in consumption at time zero, indicating that the sharp change observed in the previous figure is indeed driven by changes to the treatment group. The figure exhibits a slight upward trend, consistent with modest differential time trends between the treatment and comparison groups. To address potential concerns about modest trends of this type, later in the paper we will report estimates which include parametric time trends. Overall, results are similar in those specifications indicating that our estimates are not being unduly affected. Vol. 6 No. 4 davis et al.: Cash for Coolers 221 10 5 Kilowatt hours per month 0 −5 −10 −12 −10 −8 −6 −4 −2 0 2 4 6 8 10 12 Months before and after imputed replacement Figure 4. Assessing the Validity of the Comparison Group Notes: This figure is constructed in the same way as Figure 3 but for the comparison group rather than the treatment group. Nonparticipating households are assigned hypothetical replacement dates equal to the replacement dates of the participating household to which they are matched. B. Baseline Estimates Table 2 presents baseline estimates. Least squares coefficients and standard errors are reported from five separate regressions. The regressions in columns 1–3 are estimated using the complete set of participating households and an equal-sized ran- dom sample of nonparticipating households. The specification in column 1 includes household by calendar month and month-of-sample fixed effects. In this specifica- tion, refrigerator replacement decreases electricity consumption by 12.4  kilowatt hours per month. This is similar in magnitude to the difference observed in the event study figure. Mean pretreatment electricity consumption among households who replaced their refrigerators is 153 kilowatt hours per month so this is an 8 per- cent decrease. Whereas refrigerator replacement decreases electricity consumption, the estimates indicate that air conditioning replacement increases consumption by 6.6 kilowatt hours per month. Mean electricity consumption among households who replaced their air conditioners is 395 kilowatt hours per month, so this is less than a 2 percent increase. Column 2 adds month-of-sample by county fixed effects to control better for differences in weather and other time-varying factors. The point estimate for refrigerator replacement decreases to −10.3 and the point estimate for air condi- ­ replacement increases slightly and becomes statistically significant. In col- tioner ­ umn 3 we expand the specification to include an additional regressor corresponding to an interaction between air conditioning replacement and the six summer months 222 American Economic Journal: economic policyNOVEMBER 2014 Table 2—The Effect of Appliance Replacement on Household Electricity Consumption: Main Results (1) (2) (3) (4) (5) 1[New refrigerator]it −12.4** −10.3** −10.3** −11.4** −11.9** (1.4) (0.6) (0.6) (0.7) (0.75) 1[New air conditioner]it 6.6 7.2* 1.4 1.4 1.2 (5.6) (3.2) (1.1) (1.2) (1.3) 1[New air conditioner]it × 1[Summer months]it 14.3* 12.1* 13.6* (6.0) (5.9) (6.2) Household by calendar month fixed effects Yes Yes Yes Yes Yes Month-of-sample fixed effects Yes Yes Yes Yes Yes Month-of-sample by county fixed effects No Yes Yes Yes Yes Including treatment households only No No No Yes Yes Dropping month of replacement No No No No Yes Number of households 1,914,160 1,914,160 1,914,160 957,080 957,080 R2 0.91 0.91 0.91 0.93 0.93 Notes: This table reports coefficient estimates and standard errors from five separate regressions. In all regressions the dependent variable is monthly electricity consumption in kilowatt hours and the coefficients of interest corre- spond to indicator variables for households who have replaced their refrigerator or air conditioner through C4C. The sample includes billing records from May 2009 to April 2011 from the complete set of households that participated in the program and an equal-sized random sample of nonparticipating households. Mean pretreatment electricity use is 153 and 395 kilowatt hours per month for households who replaced refrigerators and air conditioners, respec- tively. Standard errors are clustered by county. ** Significant at the 1 percent level.  * Significant at the 5 percent level. (May–October). We would expect air conditioning replacement to have little effect on electricity consumption during cool months, and most meaningfully impact electricity consumption during warm months. The coefficient estimates appear to bear this out. While new air conditioners appear to have little impact during winter months, the estimates indicate an increase in summer electricity consumption of 14.3 kilowatt hours per month. Columns 4 and 5 of Table 2 present results from specifications in which we drop the comparison group entirely and estimate regressions using only participat- ing households. These regressions continue to include month-of-sample by county fixed effects and thus are identified by exploiting differential timing of replace- ment across households. The estimates in column 4 change little compared to the previous columns, suggesting that what matters most in these regressions is the ­ within-household comparison. Column 5, in addition, drops the month during which replacement occurred and results are again similar. Each column in Table 2 represents a single regression in which we estimate effects for both refrigerators and air conditioners. Estimates are essentially identi- cal when we, alternatively, estimate these effects with separate regressions in each case keeping only households who replaced a certain type of appliance and the ­ comparison households to which those households are matched. This is reassur- ing because it suggests that the time effects are adequately controlling for seasonal effects and underlying trends even though households who replaced air conditioners have considerably higher baseline consumption levels. Vol. 6 No. 4 davis et al.: Cash for Coolers 223 Table 3—The Effect of Appliance Replacement on Household Electricity Consumption: Matching Estimates   Matching on location and Matching on location  pretreatment consumption (1) (2) (3) (4) (5) (6) 1[New refrigerator]it −11.0** −10.9** −10.9** −9.5** −9.2** −9.2**   (0.7) (0.5) (0.5) (0.7) (0.5) (0.5) 1[New air conditioner]it 8.0 6.5* 0.1 9.5 8.3** 2.1*   (5.3) (3.2) (1.2) (5.2) (3.0) (1.0) 1[New air conditioner]it ×     15.5* 15.2*  1[Summer months]it     (6.3) (6.1)               Household by calendar Yes Yes Yes Yes Yes Yes   month fixed effects Month-of-sample fixed effects Yes Yes Yes Yes Yes Yes Month-of-sample by county No Yes Yes No Yes Yes   fixed effects               Number of households 1,914,160 1,914,160 1,914,160 1,914,160 1,914,160 1,914,160 R2 0.93 0.93 0.93 0.92 0.92 0.92 Notes: This table reports coefficient estimates and standard errors from six separate regressions. In all regressions the dependent variable is monthly electricity consumption in kilowatt hours and the coefficients of interest corre- spond to indicator variables for households who have replaced their refrigerator or air conditioner through C4C. The sample includes billing records from May 2009 to April 2011 from the complete set of households that partici- pated in the program and an equal-sized matched sample of nonparticipating households. Matching is performed using location only in columns 1–3 and using both location and pretreatment electricity consumption levels in col- umns 4–6. Standard errors are clustered by county. ** Significant at the 1 percent level.  * Significant at the 5 percent level. C. Additional Specifications Table 3 reports estimates using our matched comparison groups. The estimating equations and sample of participating households are identical to columns  1–3 of Table 2. But instead of a random sample of nonparticipants, these results are based on our matched comparison groups. Overall, the results are very similar to the previous table. When matching on location and pretreatment consumption, the point estimates for the effect of refrigerator replacement are somewhat smaller, ranging from −9.2 to −9.5 kilowatt hours per month. For air conditioner replacement we continue to ­ onsumption in see a distinct seasonal pattern, with near-zero changes in electricity c the winter, and an average increase of 15+ kilowatt hours per month in the summer. These results rely on the comparison group being a reasonable counterfactual for what would have happened to participating households had they not replaced their appliances. We find it reassuring that results are similar across comparison groups, and similar even when no comparison group is used at all in columns 4 and 5 of Table 2. Moreover, the sharp drop observed in electricity consumption among participating households, together with no sharp change in the comparison group, lends support to the interpretation of these changes as being caused by the program. Nonetheless, one could continue to be concerned about differential trends biasing our estimates. Our estimates assume that the change in electricity c ­onsumption 224 American Economic Journal: economic policyNOVEMBER 2014 Table 4—The Effect of Appliance Replacement on Household Electricity Consumption: Including Time Trends No time Linear Quadratic Cubic time trend time trend time trend trend (1) (2) (3) (4)           1[New refrigerator]it −9.2** −11.2** −11.2** −11.2**   (0.5) (0.7) (0.7) (0.7)         1[New air conditioner]it 2.1* 0.1 0.3 0.2   (1.0) (1.0) (1.0) (1.0)         1[New air conditioner]it × 15.2*  15.3* 15.0* 15.0*   1[Summer months]it (6.1)  (6.1) (6.1) (6.1)           Household by calendar month Yes Yes Yes Yes   fixed effects Month-of-sample by county Yes Yes Yes Yes   fixed effects           Number of households 1,914,160 1,914,160 1,914,160 1,914,160 R2 0.92 0.92 0.92 0.92 Notes: This table reports coefficient estimates and standard errors from four separate regressions aimed at assessing the robustness of the results with regard to including a parametric time trend for participants. In all regressions the dependent variable is monthly electricity consumption in kilowatt hours and the coefficients of interest correspond to indicator variables for households who have replaced their refrigerator or air conditioner through C4C. The sam- ple includes billing records from May 2009 to April 2011 from the complete set of households that participated in the program and an equal-sized matched sample of nonparticipating households selected using location and pre- treatment electricity consumption. Standard errors are clustered by county to allow for arbitrary serial correlation and correlation across households within municipalities. ** Significant at the 1 percent level.  * Significant at the 5 percent level. in the comparison group is an unbiased estimate of the counterfactual. This is not testable. However, we can test whether the changes over time in the treatment group ­ are the same as those in the comparison group in the pre-intervention period. Table 4 reports results including time trends. Specifically, we construct a time trend variable which, for participating households, is equal to the number of months since May 2009, and for nonparticipating households is equal to zero for all months. And we consider specifications which include this time trend variable linearly, as well as quadratic and cubic functions of this variable. Thus in these specifications we allow average consumption by participating households to evolve according to a polynomial time trend. For these estimates the comparison group is nonparticipants matched on location and pretreatment consumption. We find that the results are rela- tively insensitive to including a time trend. The coefficient on refrigerator replace- ment increases modestly from −9.2 to −11.2 once a time trend has been included and results are very similar with linear, quadratic, and cubic time trends. IV. Mechanisms Our estimates of savings are considerably smaller than the ex ante predictions that were used to motivate the program. The World Bank study, for example, ­ considers Vol. 6 No. 4 davis et al.: Cash for Coolers 225 an intervention essentially identical to C4C, in which refrigerators ten years or older are replaced with refrigerators meeting current standards. The World Bank predicted that these refrigerator replacements would save 481 kilowatt hours per year, with larger savings for very old refrigerators.6 The same study predicts that replacing air conditioners would save 1,200 kilowatt hours per year. We find that the actual sav- ings from refrigerator replacement averaged only 135 kilowatt hours per year, about one-quarter of the savings predicted by the World Bank. And for air conditioning, we find that electricity consumption increases after replacement by an average of 91 kilowatt hours per year. This section considers the key mechanisms that led actual savings to fall short of the ex ante predictions. We begin in Section IVA by examining the age of the appli- ances that were replaced. We show that while the World Bank predictions hinged on the program effectively targeting very old appliances, that most of the appliances that were replaced were close to the ten-year cutoff. Section IVB examines the seasonal pattern of treatment effects, finding that it points toward increases in air condition- ing usage during summer months. In Section IVC we discuss increases in appliance size and features, showing, for example, that side-by-side doors and through-­ the- door ice increase electricity consumption substantially. Then in Section  IVD we consider the possibility that some of the appliances may have been nonworking at the time of replacement. Finally, Section IVE presents complementary evidence from comparing estimated savings across different subsets of households. We find that the mechanisms explored in this section, taken together, can easily reconcile our results with the ex ante predictions. A. Appliance Age Figure 5 plots sales-weighted average electricity consumption for refrigerators and room air conditioners sold in the United States between 1980 and 2009. Similar data are not available for Mexico but the US experience is informative because the two countries have had identical energy efficiency standards since the mid-1990s for both appliances. US minimum energy-efficiency standards for refrigerators were first enacted in 1990, and then updated in 1993 and 2001. The second two changes are clearly visible in the figure with large, discontinuous decreases in consumption in 1993 and 2001. Mexico adopted the same standards in 1994 (NOM-072-SCFI-1994) and 2002 (NOM-015-ENER-2002). US minimum standards for room air condition- ers started in 1990, and were updated in 2000. Neither change resulted in an imme- diate, visible, change in average energy consumption. Mexico adopted the same standards in 1994 (NOM-073-SCFI-1994) and 2000 (NOM-021-ENER-2000). Over these three decades there was a dramatic decrease in electricity consump- tion for both appliances. Refrigerator electricity consumption decreased 67 per- cent while air conditioner consumption decreased 30 percent. For both appliances, however, most of this decrease occurred during the 1980s and early 1990s. These 6  See Johnson et al. (2009), Appendix C “Intervention Assumptions,” pages 123–124 (air conditioners) and page 125 (refrigerators). Another point of comparison is Arroyo-Cabañas et al. (2009) which predicted that replacing a pre-2001 refrigerator in Mexico would reduce electricity consumption by an average of 315 kilowatt hours per year. 226 American Economic Journal: economic policyNOVEMBER 2014 1,400 Average electricity consumption (kWhs per year) 1,200 1,000 800 600 400 200 Refrigerators Room air conditioners 0 1980 1990 2000 2010 Figure 5. Improvements in Appliance Energy-Efficiency over Time Notes: This figure was constructed by the authors using data from AHAM (2010). See Nadel (2002) and Rosenfeld and Poskanzer (2009) for similar figures. These series have been nor- malized to reflect average 2009 appliance sizes. Refrigerators experienced a modest increase in average size over this time period so the nonnormalized series shows a somewhat smaller change in electricity consumption. Room air conditioners, meanwhile, have experienced a modest decrease in average capacity so the nonnormalized series shows a somewhat larger change in electricity consumption. Data from 1998 are not available. data imply that, on average, replacing a 20-year-old refrigerator would save about 530 kilowatt hours per year, while replacing a 10-year-old refrigerator would save only about 250 kilowatt hours per year. Although the World Bank is not explicit about where its estimate came from, implicitly in predicting savings of 481 kilowatt hours per year, the analysts seem to have been assuming that the program was going to tend to draw a large fraction of refrigerators that were 20+ years old. For air conditioners it is harder to make sense of the World Bank estimate. In constructing Figure 5 we assumed 750 hours of annual usage. This is the number of hours used by the US Federal Trade Commission in reporting estimated yearly oper- ating costs in the yellow EnergyGuide labels, and is the baseline level of usage for statistics reported by the Association of Home Appliance Manufacturers (AHAM 2010). Although a reasonable starting point, this is probably too low of a level of usage for Mexico. In Figure 2B, households with air conditioners have about 2,200 kilowatt hours of excess consumption during summer months. Before replacement a typical air conditioner used about 1,000 watts, so assuming this entire excess is air conditioning this is 2,200 hours of annual usage. With this level of usage the implied savings of replacing a 25+ year-old air conditioner is about 900 kilowatt hours per year. To reach the World Bank’s prediction of 1,200 kilowatt hours one would need to assume a somewhat higher level of usage and to continue to assume that the pro- gram was effective at targeting very old units. Vol. 6 No. 4 davis et al.: Cash for Coolers 227 Thus, the World Bank predictions hinged on the program being successful at recruit- ing households with very old, very inefficient appliances. There is an ­economic argu- ment for this. After all, these households do have the most to gain from replacement. However, it also depends on the number of old appliances in circulation. According to Arroyo-Cabañas et al. (2009), when the program started there were approximately ten million refrigerators in Mexico over 10 years old, but only about 15 percent of which were 20+ years old. Similar analysis of room air conditioners is not available but most analysts assume that room air conditioners have a shorter average lifetime than refrig- erators. See, for example, the US Department of Energy’s Modeling System (NEMS). In practice, the program does not appear to have been particularly effective at tar- geting households with very old appliances. The average reported age of the refrig- erators that were replaced is 13.2 years. Almost 70 percent were reported to be 10–14 years old, 20 percent were 15–19, and only 10 percent were 20 years or older. The average reported age for air conditioners is 10.9 years and only 5 percent were reported to be more than 15 years old. There is likely to be significant measurement error in these self-reported ages. It can be difficult to determine an appliance’s age just by looking at it, and there was no particular incentive for participants to report this age correctly (aside from reporting it was 10+ years old). Nevertheless, this apparent lack of success at targeting very old appliances is striking, and can provide part of the explanation as to why our results differ from the ex ante predictions. B. Appliance Usage Another explanation for the differences is that the ex ante analyses did not account for possible increases in appliance usage. Although changes in usage are likely to be modest or even nonexistent for refrigerators, one would expect the new air condition- ers to be used more because they cost less to operate. Increases in usage can mean leaving the unit on more hours per day or adjusting the settings to achieve additional thermal comfort. Changes in air conditioner usage also reflect substitution between alternative cooling technologies (electric fans, evaporative coolers, natural ventila- tion, etc.). Air conditioners use much more electricity than these alternative cool- ing technologies. For example, a typical room air conditioner uses 500–1,000 watts while a fan uses less than 50 watts. So just about any form of substitution would have led to increased electricity consumption. Figures 6A and 6B plot the effect of appliance replacement by month of year. To create these graphs we estimate 12 separate regressions, one for each calendar month. In each regression we keep only observations from a single calendar month. For example, for May we keep only electricity consumption that was billed in May 2009 or May 2010. Thus the estimated coefficient reflects the changes in electric- ity consumption from May to May, identified using households who replaced their appliances during any of the months between. All regressions include h ­ ousehold fixed effects so the estimates should be interpreted as the change in consumption before and after appliance replacement. For refrigerators, the estimates are similar across calendar months. The estimates are precisely estimated so we reject the null hypothesis that all 12 estimates are equal, but the range is fairly narrow. The air conditioner estimates, however, follow a ­distinct 228 American Economic Journal: economic policyNOVEMBER 2014 50 50 40 40 30 per month 30 20 month 20 10 hours hours per 10 0 Kilowatt −10 0 Kilowatt −20 10 −30 20 May July September November January March −30 Figure 6A. The Effect of Refrigerator Replacement by Month of Year May July September November January March 50 50 40 40 30 per month 30 20 month hours 20 10 hours per Kilowatt 10 0 −10 Kilowatt 0 −20 10 −30 20 May July September November January March −30 May July September November January March Figure 6B. The Effect of Air Conditioner Replacement by Month of Year Notes: Each figure plots estimated coefficients and ninety-fifth percentile confidence intervals corresponding to an indicator variable for households that have replaced their appliance from 12 separate regressions, one for each calendar month. The dependent variable in all regressions is monthly electricity consumption in kilowatt hours and the regressions include, in addition household by calendar month fixed effects and month-of-sample by county fixed effects. The sample includes billing records from May 2009 to April 2011. The 1,914,160 households in the complete sample include 957,080 households who participated in C4C and an equal number of nonparticipating households matched on location and pretreatment consumption. Standard errors are clustered by county. Vol. 6 No. 4 davis et al.: Cash for Coolers 229 seasonal pattern. The effect of replacement on electricity c ­ onsumption is close to zero during winter months, but large and positive during summer months. The larg- est coefficient corresponds to September. Because the billing data is bimonthly, this reflects change in consumption during August and September, two of the warmest months in Mexico. The value of air conditioning is highest during hot months, and the evidence is consistent with an increase in usage during these months. For households that replaced air conditioners, the estimates imply a total increase of about 90 kilowatt hours annually. This could be explained by a modest increase in usage. Before replacement, households with air conditioners use on average about 400 kilowatt hours more per month during the summer than the winter (see Figure 2B). This is mostly air conditioning. Based on the analysis in Section IVB, replacing a 10–15-year-old air conditioner would be expected to reduce consumption from air conditioning by about 10 percent: i.e., 40 kilowatt hours per month. Instead, we are finding an increase of 20–30 kilowatt hours per month during the warmest months. This would have required only about a 20 percent increase in usage. One would expect air conditioner usage in Mexico to be particularly price elastic. In high-income countries, many households choose to maintain near ideal levels of thermal comfort at most hours of the day regardless of energy costs. In middle- income countries, however, most households operate their air conditioners only on hot days, or during particular hours of the day, so there is more scope for changes in usage. Still, the implied increase in usage is higher than one would have expected based on the pure price response. Estimates in the literature of the short-run price elasticity of air conditioner usage tend to be considerably smaller than 1 (see, e.g., Rapson 2014). Thus it seems likely that the increase in consumption is a result of not only the lower cost of operation, but also increased capacity and features. C. Appliance Size and Features Another reason the ex ante predictions were too optimistic is that they failed to incorporate increases in appliance size and features. Under the program’s rules, refrigerators and air conditioners were supposed to meet specific size requirements. New refrigerators were supposed to be between 9 and 13 cubic feet, and have a max- imum size no more than two cubic feet larger than the refrigerator which is replaced. Similar requirements were imposed for air conditioners. Many of the appliances for sale in Mexico during this period exceeded these requirements. For example, in a July 2009 report, the Mexican Consumer Protection Office tested 27 refrigerators for sale in Mexico (PROFECO 2009). The average size among refrigerators that were tested was 13.5 cubic feet, and 17 out of 27 were larger than 13 cubic feet. Each additional cubic foot of refrigerator capacity adds about ten kilowatt hours of electricity consumption per year.7 7  Current energy-efficiency standards in the United States and Mexico specify that refrigerators with ­ op-mounted freezers and automatic defrost without through-the-door ice have a maximum annual electricity use of t 9.80AV + 276.0 where AV is the total adjusted volume in cubic feet. Under C4C new refrigerators were supposed to be between 9 and 13 cubic feet, implying a range of minimum consumption from 364 to 403 kilowatt hours per year, with each cubic foot adding 9.8 kilowatt hours per year. 230 American Economic Journal: economic policyNOVEMBER 2014 Perhaps more important than the size increases is the fact that new appliances tend to have more advanced features that increase electricity consumption. Most new refrigerators have ice-makers, and many also have side-by-side doors and through- the-door ice and water. These features are valued by households but they are also energy-intensive. Side-by-side doors, for example, increase electricity consumption by 100+ kilowatt hours per year.8 And through-the-door ice increases electricity con- sumption by about 80 kilowatt hours per year.9 Air conditioners have also added fea- tures. They have become much quieter, and many new models have lower cycle speeds for operating at night, thermostats, and remote control operation. These features make air conditioners easier and more convenient to use, contributing to increased usage. D. Possible Nonworking Appliances Another potential mechanism that has been raised is nonworking appliances. Appliances were supposed to be in working order to be eligible for replacement. But if households were somehow able to replace nonworking appliances (or appli- ances that did not work well), this would provide an additional explanation for the gap between our estimates and the ex ante predictions. Although we think this may have occurred in some cases, we do not think this was widespread. First, the retailer was supposed to verify that the old appliance was in working order. Typically this was performed at the same time the old appliance was picked up. While it is true that the retailer had an incentive to see the transaction completed, it also would have been risky for a retailer to violate the program requirements grossly. Appliances were tested again upon arrival at the recycling centers, and although occasionally one might expect an appliance to be damaged in transit, it would have been suspicious if a large fraction of appliances from a particular retailer showed up defective. Second, as we mentioned in Section IB, households with very low levels of his- toric average electricity consumption were ineligible for the program. This require- ment was implemented explicitly to prevent households from replacing nonworking appliances. The minimum consumption level was 75 kilowatt hours per month for refrigerator replacement, and 250 kilowatt hours per month for air conditioner replacement. Although of course no simple rule like this is going to work perfectly, these minimums were set at reasonable levels such that households without working appliances in these categories would have likely been below the cutoffs. Finally, the pretreatment pattern of consumption (Figure 2) provides additional evidence that most appliances were working at the time of replacement. Households who replaced their refrigerators have winter consumption of 130–140 kilowatt hours 8  Current energy-efficiency standards in the United States and Mexico specify that refrigerators with top- mounted freezers and automatic defrost without through-the-door ice have a maximum annual electricity use of 9.80AV + 276.0 where AV is the total adjusted volume in cubic feet. Refrigerators with side-mounted freezers without through-the-door ice have a maximum annual electricity use of 4.91AV + 507.5. Side-by-side doors are typically only available at larger sizes. For a 20 cubic foot refrigerator, for example, the difference in maximum electricity consumption is 133.7 kilowatt hours per year. 9  Current energy-efficiency standards in the United States and Mexico provide separate requirements for refrig- erators with and without through-the-door ice. Refrigerators without through-the-door ice have a maximum energy use of 9.80AV+276.0 where AV is the total adjusted volume in cubic feet. The equivalent formula for refrigerators with through-the-door ice is 10.20AV + 356.0. Vol. 6 No. 4 davis et al.: Cash for Coolers 231 per month. It would be unusual to reach this level of baseload consumption without a working refrigerator. And households who replaced their air conditioners exhibit a pronounced seasonal pattern. This is not to say that every single air conditioner that was turned in was in perfect working condition, but you would not expect to see this threefold increase between winter and summer months if a large fraction of participants were replacing nonworking air conditioners. E. Heterogeneous Effects Table 5 reports estimates from three separate regressions, one per panel. We report estimates corresponding to interactions between indicator variables for appli- ance replacement and indicator variables for whether a participant belongs to a particular subset as indicated in the row headings. The sample used in these regres- sions includes all participants, along with our matched sample of nonparticipating households in which matching is performed using both location and pretreatment consumption. All regressions include household by calendar month and county by month-of-sample fixed effects and thus can be compared to the estimates in col- umn 5 of Table 3. Panel A describes how the effect of appliance replacement varies by the mean household income in the county where the participant lives. For refrigerators, the estimates are negative and statistically significant for all three income terciles. The largest decreases are observed in high-income counties. This could reflect that house- holds in these counties already tended to have larger and more feature-rich refrigera- tors pre-substitution, so there was less scope for increases along these dimensions to offset the efficiency gains. It might also be that in higher-income municipalities there was more of a tendency for households to turn in well-­ functioning ­ refrigerators. For air conditioners, the estimates are positive for all three income terciles, but not sta- tistically different from one another. Panel B presents estimates by the self-reported age of the old appliance. For both appliance types the estimates are very similar across age groups. Somewhat surpris- ingly, there is no evidence of larger savings for households who replace older appli- ances. We have already mentioned that these self-reported ages are likely observed with considerable measurement error, and this could explain the lack of a consistent pattern. It could also be that there are systematic differences in appliance size and features that tend to work in the other direction. For example, older appliances tend to be smaller with less features, tending to offset the pure age effect. Lastly, panel C reports estimates by the year of replacement. The program was launched in 2009 and we have in our analysis replacements made during each of the first three years. Savings tend to decrease over time. Refrigerators replaced during 2011 are associated with savings of only 3.2 kilowatt hours per month. And although conditioners the differences are not statistically significant, the point estimates for air ­ have the same pattern, showing larger increases in later years. One might expect to see this pattern if households who participated early in the program had the most to gain. For example, households with very old or very energy-­ inefficient appliances would have likely wanted to participate in C4C as soon as possible. As time goes on, how- ever, an increasing proportion of the participating households are close to indifferent 232 American Economic Journal: economic policyNOVEMBER 2014 Table 5—Heterogeneous Effects Refrigerators Air conditioners Panel A. By mean household income in county (2010 census) First tercile (less than $5,000/year) −6.7** (0.3) 5.4* (2.9) N = 305,669 N = 13,202 Second tercile ($5,000–$7,637/year) −10.0** (1.1) 7.6** (1.8) N = 275,941 N = 42,176 Third tercile (more than $7,637/year) −11.0** (0.9) 9.5 (6.5) N = 277,352 N = 43,226 Panel B. By age of old appliance (self-reported) Old appliance exactly 10 years old −9.2** (0.6) 8.9* (3.5) N = 380,803 N = 66,964 Old appliance 11–14 years old −9.1** (0.7) 6.8** (2.7) N = 214,940 N = 23,753 Old appliance 15+ years old −9.3** (0.5) 7.3* (3.1) N = 263,219 N = 7,887 Panel C. By year of replacement Appliance replaced in 2009 −9.7** (0.7) 6.4 (5.0) N = 180,507 N = 15,267 Appliance replaced in 2010 −9.5** (0.6) 8.3** (3.1) N = 497,148 N = 59,499 Appliance replaced in 2011 −3.2** (0.4) 11.7** (2.5) N = 181,307 N = 23,838 Notes: This table reports coefficient estimates and standard errors from three separate regressions, one per panel. In all regressions the dependent variable is monthly electricity consumption in kilowatt hours. We report estimates cor- responding to interactions between indicator variables for appliance replacement and indicator variables for whether a participant belongs to a particular subset as indicated in the row headings. The sample used in these regressions includes all participants, along with a matched sample of nonparticipating households in which matching is performed using both location and pretreatment consumption. All regressions include household by calendar month and county by month-of-sample fixed effects. Standard errors are clustered by county. The sample sizes indicated above are the number of treatment households in each category. The implied total number of participants differs slightly from the sample size in other tables because 486 households replaced both a refrigerator and an air conditioner. ** Significant at the 1 percent level.  * Significant at the 5 percent level. between replacing and not replacing. These newly eligible households tend to have less to gain on average from replacement, and the estimates appear to bear this out. Overall, the estimates are remarkably similar across subsets. Across groups, we find modest savings for households who replaced refrigerators, and modest increases in consumption for households who replaced air conditioners. These estimates pro- vide further corroboration of our main findings, indicating that the results are not driven by the experience of any particular subgroup. V. Cost-Effectiveness A. Baseline Estimates Panel A of Table 6 reports the mean annual impacts implied by our estimates. Based on the estimates in column 4 of Table 4 refrigerator replacement reduces Vol. 6 No. 4 davis et al.: Cash for Coolers 233 Table 6—Electricity Expenditures, Carbon Dioxide Emissions, and Cost-Effectiveness Both Air appliances Refrigerators conditioners combined (1) (2) (3) Panel A. Mean per replacement Mean annual change in electricity consumption per replacement −135 91 —   (kilowatt hours) Mean annual change in household expenditure per replacement −$13 $9 —   (2010 US$) Panel B. Totals Total replacements nationwide (between May 2009 and April 2011) 858,962 98,604 957,566 Total annual change in electricity consumption (gigawatt hours) −115.7 9.0 −106.7 Total annual change in household expenditures −$11.1 $0.9 −$10.2   (in millions 2010 US$) Total annual change in carbon dioxide emissions (thousands of tons) −62.2 4.8 −57.4 Panel C. Cost-effectiveness Total Direct program cost (in millions 2010 US$) $129.4 $13.4 $142.7 Program cost per kilowatt hour (2010 US$) $0.25 — $0.29 Program cost per ton of carbon dioxide (2010 US$) $457 — $547 Notes: Mean annual change in electricity consumption per replacement comes from column 4 of Table 4. Change in expenditures is calculated using an average price of $0.096 per kilowatt hour. Carbon dioxide emissions are cal- culated using 0.538 tons of carbon dioxide per megawatt hour (538 tons per gigawatt hour) following Johnson et al. (2009). Direct program cost is the dollar value of the cash subsidies and excludes administrative costs. In calcu- lating the program cost per kilowatt hour and program cost per ton of carbon dioxide we assumed that the program accelerated replacement by five years and used a 5 percent annual discount rate. e ­lectricity consumption by 135 kilowatt hours annually, while air conditioner replacement increases electricity consumption by 91 kilowatt hours per year. At average ­ residential electricity prices, refrigerator replacement saves households $13 annually, while air conditioner replacement costs households an additional $9 annually. Panel B describes the total impact of C4C between May 2009 and April 2011. In our sample there are close to 850,000 refrigerator replacements and 100,000 air conditioner replacements so our estimates imply a total reduction in electricity consumption of 106.7 gigawatt hours annually (858,962 × 135 + 98,604 × −91  = 106,700,000 kilowatt hours). At average residential electricity prices this is a reduction in household expenditures of $10 million annually. This panel also reports estimates of the total change in carbon dioxide emissions. One of the central goals of C4C was to reduce carbon dioxide emissions so these estimates are an important measure of the effectiveness of the program. Multiplying the change in electricity consumption by the average carbon intensity of electricity generation in Mexico yields a decrease of 57,400 tons of carbon dioxide emissions annually. Using an estimate for the social cost of carbon dioxide of $34 per ton these emissions reductions provide $2.0 million in benefits annually.10 Greenstone, Kopits, and Wolverton (2013) present a range of values for the social cost of carbon dioxide 10  according to different discount rates and for different time periods that is intended to capture changes in net agri- cultural productivity, human health, property damages from increased flood risk, and other factors. These estimates were then updated by US IAWG (2013). With a 3 percent discount rate (their central value) for 2010 they find a social cost of carbon dioxide of $34 per ton. 234 American Economic Journal: economic policyNOVEMBER 2014 Electricity generation also emits sulfur dioxide and other criteria pollutants. According to CEC (2011), Mexican plants emit 2.4 times as much sulfur dioxide, 1.7 times as much nitrogen oxide, and 2.2 times as much particulates (PM10) per kilowatt hour as US plants. Muller, Mendelsohn, and Nordhaus (2011) estimate the external damages from these pollutants for different forms of US power generation. Coal-fired power plants are the most damaging ($0.028 per kilowatt hour), while oil ($0.02) and, in particular, natural gas ($0.002) are less damaging. Using the mix of electricity generation in Mexico and scaling damages by 2.4 to reflect higher emis- sions levels yields additional benefits of $2.9 million annually. These calculations reflect the changes in energy consumption from appliance operation but not changes in energy consumption from other parts of the appli- ance life cycle. The program accelerated appliance production and recycling, both of which are energy-intensive. Incorporating these sources of energy consumption would offset the estimated reductions, but only modestly. Taking into account mate- rials production and processing, assembly, transportation, dismantling, recycling, shredding, and recovery of refrigerant, Kim, Keoleian, and Horie (2006) find that energy usage during operation accounts for 90 percent of total refrigerator life-cycle energy use. We are not aware of a similar life-cycle analysis of air conditioners but their energy consumption is also heavily driven by operation. Panel C of Table 6 reports baseline estimates of cost-effectiveness. Based on the total number of participants and the subsidies that they received we calculate that direct program costs were $129 million for refrigerators, and $13 million for air condi- tioners. This includes the cash subsidies received by households, but not costs incurred in program design, administration, advertising, or other indirect costs. Dividing by the estimated change in electricity consumption provides a measure of the direct pro- gram cost per kilowatt hour reduction. The relevant change here is the total discounted lifetime change in electricity consumption. For this calculation we adopt a 5 percent annual discount rate and assume that the program accelerated appliance replacement by five years. Under these assumptions the program cost per kilowatt hour is $0.25 for refrigerators and $0.29 overall. We do not report program cost per kilowatt hour separately for air conditioners because the program led to an increase in consumption. The program cost per ton of carbon dioxide emissions can be calculated similarly. For both refrigerators only and for the entire program, this exceeds $450 per ton. These estimates of program cost per kilowatt hour are high compared to most available estimates from energy-efficiency programs in the United States. For exam- ple, US electric utilities reported in 2011 spending $4.0 billion in energy-efficiency programs leading to 121 terawatt hours of energy savings, implying an average direct program cost per kilowatt hour of $0.033.11 Economists have long argued that these self-reported measures likely overstate the cost-effectiveness of these pro- grams (Joskow and Marron 1992). Nonetheless, it is striking that our estimate for C4C is about nine times larger. With regard to carbon dioxide abatement, Knittel (2009) finds that the direct program cost for Cash for Clunkers exceeded $450 per ton, similar in magnitude to our estimates. EIA (2013b, Tables 10.1 and 10.5). As another point of comparison, Allcott (2011) reports a program cost per 11  kilowatt hour for peer-comparison reports from OPOWER ranging from $0.02– $0.05. Vol. 6 No. 4 davis et al.: Cash for Coolers 235 Our estimates of program cost per kilowatt hour remain high under more generous assumptions. With a 0 percent discount rate, the program cost per cost per kilowatt hour is $0.27, and the program cost per ton of carbon dioxide is $497. If one assumes that the program accelerated appliance retirement program by ten years, then the program cost per kilowatt hour is $0.17, and the program cost per ton of carbon dioxide is $307. Alternative program designs might have modestly improved c ­ ost-effectiveness. Some have argued, for example, that C4C would have been more cost-effective if partici- pants had been required to purchase appliances that greatly exceed ­ energy-efficiency standards.12 Had the new refrigerators been forced to meet US 2014 standards, we calculate that the refrigerator program would have had a program cost per kilowatt hour of $0.20, and a program cost per ton of carbon dioxide of $363. B. Welfare These measures of cost-effectiveness provide some but not all of the pieces of information necessary to evaluate whether or not the program is welfare-improving. In considering welfare, it is important to distinguish between marginal households who are induced to replace their appliance because of the program and inframar- ginal households who are getting paid to do what they would have done otherwise. The cost-effectiveness measures above assume that all households are marginal, potentially substantially overstating the environmental benefits of the program. Distinguishing between marginal and inframarginal participants is also important for evaluating the economic costs of the program. Inframarginal participants value each $1 in subsidy at exactly $1, so for them the subsidy should be viewed as a pure transfer from taxpayers to program participants. Marginal households, how- ever, value each $1 in subsidy by at most $1. These households otherwise would have stayed with their old, energy-inefficient durable good, but are induced by the subsidy to replace. For these participants the program is shifting income away from taxpayers who value it 1:1, toward participants who value it at less than 1:1. If demand is linear, for example, then there is a welfare loss of $0.50 per $1 of subsidy. In addition to this welfare loss, collecting tax revenues distorts labor and other markets. This social cost of public funds is above and beyond the welfare loss from recipient households valuing the subsidies less than 1:1. That is, even for house- holds who value these subsidies at close to 1:1, there still is welfare loss because the subsidies must be financed. These distortions are particularly unfortunate when the funds go toward households who are inframarginal because welfare losses are being incurred to transfer income to households who would have purchased the energy- efficient durable good even in the absence of the subsidy. 12  The United States, for example, will have new energy-efficiency standards for refrigerators in 2014 that require a 25 percent decrease in consumption compared to previous standards. A typical refrigerator meeting these more stringent standards uses 63 fewer kilowatt hours annually. The old standard both in the United States and Mexico requires that refrigerators with top-mounted freezers and automatic defrost without through-the-door ice have a maximum annual electricity use of 9.80AV + 276.0 where AV is the total adjusted volume in cubic feet. The new US standard for this refrigerator type adopts a formula 8.07AV + 233.7 so a 12 cubic foot refrigerator uses 63 fewer kilowatt hours per year. 236 American Economic Journal: economic policyNOVEMBER 2014 These welfare losses must be compared to welfare gains from decreased exter- nalities. The total change in externalities depends on the total number of households induced to adopt the energy-efficient durable good, and the reduction in externali- ties per adoption. With this first component, it is important to avoid counting infra- marginal households. This is often challenging empirically because while one can observe the number of adoptions, it is difficult to construct a credible counterfactual to describe what would have occurred in the absence of the policy. Typically even more difficult to measure is this second component. Accordingly, this is where we focused our attention in the previous sections. We find that the program incurred direct costs of about $140 million in exchange for carbon dioxide abatement worth $2.0 million per year and criteria pollutant abate- ment worth $2.9 million per year. Whether or not this is a welfare-improving trade-off depends on how much the households value the subsidy per $1 and on the social cost of public funds. With linear demand, participants value the $140 million in subsidies at $70 million, with $70 million in welfare loss. Added to this, one would want to mul- tiply $140 million by the social cost of public funds. Even for low values of the social cost of public funds, this would add tens of millions in additional welfare loss. Thus, overall, it appears that the costs of the program exceeded the benefits. VI. Conclusion Meeting the increase in energy demand over the next several decades will be an immense challenge and in most countries it seems unlikely in the short term that there will be the political will to implement Pigouvian-style taxes on the externali- ties associated with the production and consumption of energy. Thus it is perhaps not surprising that policymakers are increasingly turning to energy-efficiency programs. Proponents argue that these programs represent a win-win, reducing energy expen- ditures while also decreasing greenhouse gas emissions and other externalities. In countries where energy prices are subsidized, there is even a potential third win as governments reduce the amount they spend on subsidies. Moreover, among avail- able energy-efficiency programs, appliance replacement subsidies would appear to have a great deal of potential. Residential appliances have experienced dramatic gains in energy efficiency, so there would seem to be scope for these programs to substantially decrease energy consumption. Thus it is hard not to be somewhat disappointed by the estimated savings. We found that households who replace their refrigerators with energy-efficient models indeed decrease their energy consumption, but by an amount considerably smaller than was predicted by ex ante analyses. Even larger decreases were predicted for air conditioners, but we find that households which replace their air condition- actually end up increasing their energy consumption. Overall, we find that the ers ­ ­ program is an expensive way to reduce energy use, reducing electricity consumption at a program cost of $0.29 per kilowatt hour, and reducing carbon dioxide emissions at a program cost of over $500 per ton. These results underscore the urgent need for careful modeling of household behavior in the evaluation of energy-efficiency programs. Households receive utility from using appliances, so they can and should increase usage in response to increases in energy Vol. 6 No. 4 davis et al.: Cash for Coolers 237 efficiency. This rebound is a good thing—it means that households are increasing their utility (Borenstein forthcoming). It does, however, complicate the design of energy- efficiency policy and ceteris paribus, in pursuing environmental goals it will make sense for policymakers to target technologies for which demand for usage is inelastic. Our results also point to several additional lessons for the design and evaluation of energy-efficiency programs. Over time cars, appliances, and houses have become more energy efficient, but also bigger and better. These size and quality increases are another form of the demand for increased usage, and it makes sense to take them into account when designing policy. There is also a tendency for energy-efficiency programs to lose effectiveness over time. While initially a program tends to attract participants with the most to gain, as time goes on the pool will be made up increas- ingly by participants who just barely meet the eligibility requirements. REFERENCES Allcott, Hunt. 2011. “Social Norms and Energy Conservation.” Journal of Public Economics 95 (9–10): 1082–95. Allcott, Hunt, and Michael Greenstone. 2012. “Is There an Energy Efficiency Gap?” Journal of Eco- nomic Perspectives 26 (1): 3–28. Allcott, Hunt. 2014. “Site Selection Bias in Program Evaluation.” Unpublished. Allcott, Hunt, and Nathan Wozny.  Forthcoming. “Gasoline Prices, Fuel Economy, and the Energy Par- adox.” Review of Economics and Statistics. Andrade Salaverria, Dora Patricia. 2010. Evaluación Ambiental y Plan de Manejo Ambiental del Pro- grama de Eficiencia Energética Coordinado por la Secretaria de Energía. World Bank. Mexico City, April. Arimura, Toshi H., Shanjun Li, Richard G. Newell, and Karen Palmer. 2012. “Cost-Effectiveness of Electricity Energy-Efficiency Programs.” Energy Journal 33 (2): 63–99. Arroyo-Cabañas, F. G., J. E. Aguillón-Martínez, J. J. Ambríz-García, and G. Canizal. 2009. “Electric Energy Saving Potential by Substitution of Domestic Refrigerators in Mexico.” Energy Policy 37 (11): 4737–42. Association of Home Appliance Manufacturers (AHAM). 2010. Fact Book 2010. Washington, DC. Auffhammer, Maximilian, Carl Blumstein, and Meredith Fowlie. 2008. “Demand Side Management and Energy Efficiency Revisited.” Energy Journal 29 (3): 91–104. Borenstein, Severin. Forthcoming. “A Microeconomic Framework for Evaluating Energy Efficiency Rebound and Some Implications.” Energy Journal. Busse, Meghan R., Christopher R. Knittel, and Florian Zettelmeyer. 2013. “Are Consumers Myopic? Evidence from New and Used Car Purchases.” American Economic Review 103 (1): 220–56. Comisión Federal de Electricidad (CFE). 2011. Programa De Obras E Inversiones Del Sector Eléc- trico 2011–2025. Subdirección de Programación. http://www.cmic.org/comisiones/Sectoriales/ energia/electricidad/POISE/POISE2011_2025%20WEB.pdf. Commission for Environmental Cooperation (CEC). 2011. North American Power Plant Emissions. Quebec: CEC. Davis, Lucas W. 2008. “Durable Goods and Residential Demand for Energy and Water: Evidence from a Field Trial.” RAND Journal of Economics 39 (2): 530–46. Davis, Lucas, Alan Fuchs, and Paul Gertler. 2014. “Cash for Coolers: Evaluating a Large-Scale Appli- ance Replacement Program in Mexico: Dataset.” American Economic Journal: Economic Policy. http://dx.doi.org/10.1257/pol.6.4.207. Dubin, Jeffrey A., Allen K. Miedema, and Ram V. Chandran. 1986. “Price Effects of Energy-Efficient Technologies: A Study of Residential Demand for Heating and Cooling.” RAND Journal of Eco- nomics 17 (3): 310–25. Gertler, Paul, Orie Shelef, Catherine Wolfram, and Alan Fuchs. 2013. “The Demand for Energy-Using Assests among the Word’s Rising Middle Classes.” Unpublished. Greenstone, Michael, Elizabeth Kopits, and Ann Wolverton. 2013. “Developing a Social Cost of Car- bon for Use in U.S. Regulatory Analysis: A Methodology and Interpretation.” Review of Environ- mental Economics and Policy 7 (1): 23–46. 238 American Economic Journal: economic policyNOVEMBER 2014 Johnson, Todd M., Claudio Alatorre, Zayra Romo, and Feng Lui. 2009. Low-Carbon Development for Mexico. Washington, DC: World Bank. Joskow, Paul L., and Donald B. Marron. 1992. “What Does a Negawatt Really Cost? Evidence from Utility Conservation Programs.” Energy Journal 13 (4): 41–74. Kim, Hyung Chul, Gregory A. Keoleian, and Yuhta A. Horie. 2006. “Optimal Household Refrigerator Replacement Policy for Life Cycle Energy, Greenhouse Gas Emissions, and Cost.” Energy Policy 34 (15): 2310–23. Knittel, Christopher R. 2009. “The Implied Cost of Carbon Dioxide under the Cash for Clunkers Pro- gram.” Energy Institute at Haas Working Paper 189. Knittel, Christopher R. 2011. “Automobiles on Steroids: Product Attribute Trade-Offs and Technolog- ical Progress in the Automobile Sector.” American Economic Review 101 (7): 3368–99. Loughran, David S., and Jonathan Kulick. 2004. “Demand-Side Management and Energy Efficiency in the United States.” Energy Journal 25 (1): 19–43. McKinsey and Company. 2009a. “Unlocking Energy Efficiency in the U.S. Economy.” McKinsey and Company. 2009b. “Low Carbon Growth—A Potential Path for Mexico.” Presentation in Mexico City, August 20, 2009. Metcalf, Gilbert E., and Kevin A. Hasset. 1999. “Measuring the Energy Savings from Home Improve- ment Investments: Evidence from Monthly Billing Data.” Review of Economics and Statistics 81 (3): 516–28 Mian, Atif, and Amir Sufi. 2012. “The Effects of Fiscal Stimulus: Evidence from the 2009 Cash for Clunkers Program.” Quarterly Journal of Economics 127 (3): 1107–42. Muller, Nicholas Z., Robert Mendelsohn, and William Nordhaus. 2011. “Environmental Accounting for Pollution in the United States Economy.” American Economic Review 101 (5): 1649–75. Nadel, Steven. 2002. “Appliance and Equipment Efficiency Standards.” Annual Review of Energy Envi- ronment 2002 (27): 159–92. Procuraduría Federal del Consumidor (PROFECO). 2009. “Frío del Bueno: Enfría Tus Alimentos Sin Calentar Tu Recibo de Luz.” Revista Del Consumidor, July 9, 58–64. Rapson, David. 2014. “Durable Goods and Long-Run Electricity Demand: Evidence from Air Condi- tioner Purchase Behavior.” Journal of Environmental Economics and Managment 68 (1): 141–60. Reiss, Peter C., and Matthew W. White. 2005. “Household Electricity Demand, Revisited.” Review of Economic Studies 72 (3): 853–84. Rosenfeld, Arthur H., and Deborah Poskanzer. 2009. “A Graph Is Worth a Thousand Gigawatt-Hours: How California Came to Lead the United States in Energy Efficiency (Innovations Case Narrative: The California Effect).” Innovations: Technology, Governance, Globalization 4 (4): 57–79. Secretaría de Energía (SENER). 2008. Estudio Sobre Tarifas Eléctricas y Costos de Suministro. El Presupuesto de Egresos de la Federación. Secretaría de Energía (SENER). 2012. Prospectiva del Sector Eléctrico: 2012–2026. http://sener.gob. mx/res/PE_y_DT/pub/2012/PSE_2012_2026.pdf. US Energy Information Administration (EIA). 2013a. International Energy Outlook 2013. U.S. Department of Energy. Washington, DC: EIA. US Energy Information Administration (EIA). 2013b. Electric Power Annual 2011. U.S. Department of Energy. Washington, DC: EIA. US Interagency Working Group on the Social Cost of Carbon (IAWG). 2013. Technical Support Doc- ument: Technical Update on the Social Cost of Carbon for Regulatory Impact Analysis. Washing- ton, DC: USGPO. Wolfram Catherine, Orie Shelef, and Paul Gertler. 2012. “How Will Energy Demand Develop in the Developing World?” Journal of Economic Perspectives 26 (1): 119–38. World Bank. 2013. World Development Indicators for 2010. Washington, DC: World Bank Publica- tions. Zhou, Nan, Mark D. Levine, Lynn Price. 2010. “Overview of Current Energy-Efficiency Policies in China.” Energy Policy 38 (11): 6439–52.