Econometrc Methods for Estmatng ENERGY STAR Impacts n the Commercal Buldng Sector Marvn J. Horowtz, Demand Research Angela Coyle, U.S. Envronmental Protecton Agency ABSTRACT The early stages of developng a natonal mpact evaluaton of EPA s ENERGY STAR for the commercal buldng sector are descrbed n ths paper. Beng broad n geographc scope and tme-frame there are numerous research ssues that ths study rases, such as how to account for the mpacts of local publc programs and how to control for economy-wde trends that could affect energy use. The methodology descrbed n ths paper nvolves dentfyng and quantfyng the major economc and physcal factors that determne energy use, such as prncpal busness actvtes, energy prces, nvestments n specalzed stocks of captal equpment, and clmate. When completed, n addton to provdng program evaluaton fndngs, ths study wll yeld useful sde benefts, such as updated energy prce elastcty estmates needed by economsts, envronmentalsts and energy forecasters. Introducton EPA s ENERGY STAR s natonal n scope, comprehensve n desgn, ambtous n purpose, and bound to be around for a long tme. It s endorsed, supported and sustaned by the U.S. Congress. In the commercal buldngs sector, ENERGY STAR s actvtes nclude voluntary partnershp agreements and outreach actvtes for bg and small busnesses, tools for comparng energy use across buldngs and for lnkng best energy management practces to fnancal performance, and natonal awareness campagns. Every commercal buldng, n every state, could be nfluenced by these actvtes, be t an ENERGY STAR labeled buldng, a casual partcpant, or nether. How s t possble to measure the natonal envronmental protecton mpacts of ths mult-faceted program? Ths paper descrbes a methodology that s n ts development stages for estmatng the changes n commercal buldng energy use that can be attrbuted to ENERGY STAR. The methodology frst entals dentfyng the major economc and physcal factors that determne energy use, such as prncpal busness actvtes, energy prce elastctes, nvestments n specalzed stocks of captal equpment, and clmate. Separately and collectvely, the varables suggest a number of dfferent modelng optons. Complcatng matters further, many other local publc programs are currently helpng to mprove and expand the markets for commercal sector energy effcent products and servces. Many of these programs are, n dfferent forms, market transformaton programs, loosely defned as publcly-funded actvtes that promote endurng pro-energy effcency behavor. The challenge of ths study s to estmate ENERGY STAR accomplshments whle takng nto account the accomplshments of these other publc programs.
Data Issues To estmate the clmate protecton mpacts of ENERGY STAR for the commercal buldngs sector a dataset has been constructed consstng of buldng level mcrodata and regonal and natonal level data. The buldng data are derved from U.S. Energy Informaton Admnstraton s (EIA) Commercal Buldngs Energy Consumpton Survey (CBECS), a natonal survey undertaken n 1986, 1989, 1992, 1995 and 1999. As yet, the CBECS 1999 survey data are not publshed. When they are, they wll be ncorporated nto the exstng ENERGY STAR mpact evaluaton dataset. Regonal and natonal data n the dataset for the CBECS survey years are derved from varous federal agency data sources, ncludng the U.S. Census Bureau and the Federal Reserve Board. Over the past ffteen years, many trends n socety and the economy have affected energy use n the commercal sector. In addton to changng prces for energy and other energy-usng equpment, energy use has been nfluenced by such forces as ncreasng affluence, new technologes and publc programs. As such, the am of ths study s to dentfy and control for the major factors that have drven commercal buldng energy use over the past 15 years. By dong so, energy use trends related to publc programs such as ENERGY STAR may be dfferentated from energy use trends that are strctly the result of market forces. To accomplsh ths objectve a prelmnary pooled cross sectonal econometrc model of energy demand has been specfed and estmated. For understandng the ratonale behnd the modelng effort t s useful to brefly descrbe the theory underlyng energy demand. To begn, energy s not a good whch s demanded for ts value n and of tself. Rather, t s demanded as an nput to equpment and applances to produce end uses or servces such as lght and warmth. Snce energy consumpton occurs n conjuncton wth a stock of captal, ts demand s derved from the demand for the servces whch the captal provdes. Energy consumpton can thus be represented by the dentty: Q = M k= 1 R k A k where total consumpton of a gven fuel s the sum of fuel consumed by each type of captal equpment, A, n conjuncton wth utlzaton rate R k for the k th type of equpment and th type of fuel. The behavoral relatonshps between the demand for captal equpment and the derved demand for energy can be dentfed separately such that: A R = f ( P, P, P, Y, X ); = g( P j a, Y, Z ) where the demand for equpment usng fuel depends on the prce of fuel, the prce of alternatve fuel j, the prce of equpment P a, consumer ncome or total buldng expendtures Y, and lastly, a vector of other varables X. In theory, the utlzaton rate of fuel for a gven unt of equpment depends on ts own prce, ncome, and a vector of other varables Z. In the CBECS databases, nformaton on the buldng s equpment stock s lmted. Mergng the equpment demand model wth the utlzaton rate model results n an energy demand model for fuel : j
Q = h( P, P, Y, X, Z, A) j where A s a vector of equpment stock whch, unlke n the ntal energy demand dentty above, s not ted to end use consumpton. The vectors of other varables, X and Z, play mportant roles n ths model. For the present study, they nclude two knds of factors that nfluence trends n buldng energy use. These are buldng-specfc factors such as a buldngs prncpal buldng actvty and general economc factors both at the census dvson and natonal levels, such as per capta ncome. Energy Impact Models At present, n the absence of the 1999 CBECS survey, a prelmnary analyss has been completed of the four earler CBECS wth the ntenton of updatng the analyss once the most recent CBECS data are avalable. Among the many varetes of energy utlzaton functonal forms models that can be estmated, perhaps the most useful one for the ENERGY STAR evaluaton focuses on electrcty use, electrcty producton beng the sngle largest source of buldng-related greenhouse gas emssons. Therefore, the prelmnary analyss of the CBECS data entals estmaton of an energy demand model takng electrcty energy use ntensty ndex, or KWHEUI, as the dependent varable. Ths ndex s calculated as total annual buldng electrcty use dvded by buldng square feet. One of the key advantages of ths ndex s that t normalzes buldng consumpton, thereby reducng, but not elmnatng, buldng scale effects. Suppressng subscrpts related to regonal locaton and year of survey, the model of a buldng s energy consumpton takes the general form: KWHEUI = b + b P' + b WEATHER ' + b BLDG ' + b EQUIPMENT 4 0 1 2 ' + b USAGE ' + b PBA ' + b TIME ' + e 5 3 6 7 where each term represents a vector of varables and assocated coeffcents. For example, WEATHER conssts of two varables,.e. heatng degree days (HDD) and coolng degree days (CDD) calculated at base 65, and PBA represents thrteen dfferent prncpal busness actvtes, or buldng types, as dentfed n CBECS. Snce every CBECS sample was ndependent of every other one, there s only one data pont per surveyed buldng n the sample. It s mportant to note that the composton of the buldng sample vs-à-vs buldng fuel systems must be explctly taken nto consderaton n estmatng unbased model parameters. The CBECS survey documents the use of four major fuels that are commonly consumed n commercal buldng. These are electrcty, natural gas, fuel ol and steam. To control for the fact that energy use behavor wll dffer n buldngs wth dfferent combnatons of fuels, f for no other reason than dfferences n relatve fuel prces, separate models must be developed for each buldng/fuel system combnaton. The model presented n Table 1 s for buldngs wth electrcty and natural gas systems, only. Ths frst group of buldngs makes up approxmately ffty-fve percent of the CBECS sample. The second group largest group of buldngs s one wth electrcty use, only. Ths group makes up another 24 percent of the CBECS sample. The model for these
buldngs s presented n Table 2. In both tables, PBA-specfc fxed coeffcents are suppressed. Table 1. Energy Demand Model: Electrcty and Natural Gas Systems Dependent Varable = LOG(KWHEUI) Varable Coeffcent Std. Error t-statstc Prob. C b 0-10.31284 1.126739-9.152824 0.0000 LOG(EL$R) b 1a -0.913169 0.032177-28.37994 0.0000 LOG(NG$R) b 1b -0.074616 0.032141-2.321512 0.0203 LOG(HDD) b 2a -0.102725 0.016140-6.364467 0.0000 LOG(CDD) b 2b 0.085253 0.012628 6.751123 0.0000 LOG(SQFT) b 3a -0.415111 0.009890-41.97354 0.0000 LOG(AGE) b 3b -0.198179 0.013064-15.16975 0.0000 EMS b 4a 0.150238 0.020850 7.205633 0.0000 SOLAR b 4b 0.197655 0.166147 1.189642 0.2342 HPMP b 4c 0.073678 0.026108 2.822022 0.0048 LOG(WEEKHR) b 5a 0.441736 0.021149 20.88647 0.0000 LOG(WORKRS) b 5b 0.372321 0.009391 39.64670 0.0000 MAINT b 5c 0.164147 0.018358 8.941294 0.0000 VCNCY3 b 5d -0.104991 0.019447-5.398812 0.0000 NGINTRP b 5e 0.060618 0.034398 1.762240 0.0781 LOG(EARNJOBR) b 7a 1.227444 0.102269 12.00216 0.0000 Y95 b 7b -0.108344 0.024925-4.346755 0.0000 Y92 b 7c -0.034820 0.023341-1.491784 0.1358 Y89 b 7d 0.025357 0.022151 1.144755 0.2523 Adjusted R-squared 0.486 n of observatons 10,414 The energy demand models are estmated n double-log form usng ordnary least squares and the Whte heteroskedastcty correcton. Due to the partcular qualty and features of the CBECS surveys, several data screens were mposed on the mcrodata. Among the most mportant of these were that buldngs reportng ther sze as larger than one mllon square feet were dropped, as were buldngs wth unusually small or large electrc energy ntensty. Buldng types are categorzed n conformance wth the CBECS defntons, wth the excepton beng that for ths study warehouses are dfferentated by whether or not they are refrgerated. The default buldng type s what CBECS desgnates as other and buldngs desgnated as vacant are excluded from the dataset. Nomnal dollar values are adjusted to 1992 dollars usng the consumer prce ndex.
Table 2. Energy Demand Model: Electrcty System, Only Dependent Varable = LOG(KWHEUI) Varable Coeffcent Std. Error t-statstc Prob. C c 0-17.20313 2.162896-7.953747 0.0000 LOG(EL$R) c 1a -1.222380 0.054084-22.60156 0.0000 LOG(HDD) c 2a -0.001829 0.018194-0.100511 0.9199 LOG(CDD) c 2b 0.114655 0.023393 4.901323 0.0000 LOG(SQFT) c 3a -0.388476 0.014709-26.41122 0.0000 LOG(AGE) c 3b -0.103517 0.023620-4.382579 0.0000 EMS c 4a 0.218905 0.037002 5.916080 0.0000 SOLAR c 4b -0.349178 0.238365-1.464887 0.1430 HPMP c 4c -0.018267 0.027005-0.676425 0.4988 LOG(WEEKHR) c 5a 0.459851 0.031895 14.41746 0.0000 LOG(WORKRS) c 5b 0.302413 0.013431 22.51671 0.0000 MAINT c 5c 0.245688 0.029458 8.340175 0.0000 VCNCY3 c 5d -0.124733 0.029708-4.198619 0.0000 LOG(EARNJOBR) c 7a 1.686530 0.194579 8.667589 0.0000 Y95 c 7b -0.281056 0.037777-7.439954 0.0000 Y92 c 7c -0.148113 0.036863-4.017983 0.0001 Y89 c 7d -0.055456 0.036096-1.536348 0.1245 Adjusted R-squared 0.473 n of observatons 4,419 In bref, the energy demand model for electrc and natural gas consumers n Table 1 ndcates the followng: the long-term demand for electrcty s approxmately unt elastc and the prce of natural gas has lttle effect on the demand for electrcty. the coeffcent of the HDD varable s negatve, ndcatng that electrc energy ntensty decreases wth weather severty. buldng sze s negatvely related to electrcty ntensty, ndcatng that buldng scale effects reman mportant to consder even after whole buldng energy use has been normalzed by square footage. buldng age s negatvely related to electrcty ntensty, and of the three types of energy effcent equpment n the model, the coeffcents of energy management controls systems and heat pumps are statstcally sgnfcant. four of the fve buldng use-related varables are statstcally sgnfcant; longer hours and more workers are assocated wth hgher buldng electrcty ntensty and a vacancy of more than three months s assocated wth lower electrcty ntensty. Lke the energy effcency features, a buldng mantenance program s assocated wth hgher electrcty ntensty. Accordng to the model, electrcty ntensty s not statstcally sgnfcantly dfferent for those buldngs that are on a natural gas nterruptble rate schedule compared to those that are not on such a schedule.
a margnal change n average earnngs per employee across census dvsons s assocated wth a 1.2 percent ncrease n energy ntensty. Ths varable ncorporates an aggregate ncome effect reflectng regonal economc condtons. the tme trend varable ndcates that electrcty ntensty dd not fall apprecably between 1986 and 1992, but dd fall notceably from 1986 to 1995. In the alternate verson of ths model for buldngs that use electrcty, only, presented n Table 2, many of the coeffcents of the ndependent varables n ths model are smlar n sgn and magntude to those of the pror model. Two notable dfferences n ths model are (a) moderately hgher estmates of the prce elastcty of electrcty demand and the earnng elastcty of demand and (b) a statstcally sgnfcant drop n electrcty ntensty between 1986 and 1992, as well as between 1986 and 1995. Although the explanatory powers of these models appears to be moderately good, several complcatons are present. One ssue s suggested by the theory underlyng energy demand, namely that the demand for energy-related equpment, such as energy management and control systems, heat pumps and solar energy s at least partally nfluenced by a buldng s energy ntensty. In other words, these explanatory varables may be endogenous or correlated wth the energy demand model s error term. Prelmnary exploraton of ths ssue has been undertaken through the constructon of varous nstrumental varables. However, tests of these nstrumental varables have thus far ndcated that they are relatvely weak. It s therefore best to assume, for the tme beng at least, that the demand for ths type of equpment s less affected by overall buldng energy ntensty than t s by other nonenergy related factors. Another mportant ssue that these models do not address s the effect of publc polces, such as electrc utlty demand sde management (DSM) programs wthn each census dvson, on the trend n energy use. To partally address ths ssue, data related to commercal program DSM savngs can be ncorporated nto the model. However, there reman other publc programs that must also be accounted for. Estmaton of Program Impacts Ths model framework provdes a bass for statstcally modelng annual buldng energy consumpton wth pooled data from the fve CBECS surveys. Once an approprate model s developed t may be used to forecast average buldng energy consumpton n future years. A comparson between actual energy consumpton and the forecasted energy consumpton may then provde as estmates of the publc program effect. The quantfed effect can then be translated nto natonal energy savngs, and ultmately, clmate protecton mpacts. For example, usng the econometrc model provded n Table 1, suppose that ths model has been estmated usng all but the last CBECS dataset. Then, takng the most recent CBECS dataset, the energy use of all of the buldngs that are represented by a partcular model,.e., all of the buldngs wth electrc and natural gas use, are forecasted usng the observed values of each buldng s ndependent varables. In Fgure 1, the average predcted energy use for these buldngs s the pont of market equlbrum occurrng at the ntersecton of the dashed supply and demand curves on the rght hand sde of the fgure. Ths estmate can be compared to the average actual energy
use of these buldngs, portrayed as the market equlbrum formed by the ntersecton of supply and demand at the left of the fgure. Note that the equlbrum energy prce s held constant to emphasze that the forecast s based on the actual values of the ndependent varables for a gven buldng. Snce the model s desgned to control for market forces and buldng-specfc features, the dfference between the average energy use that was predcted to be occur due to market forces and buldng specfc features, and the average energy use that actually occurred s, by nference, the total non-market or publc program effect. Fgure 1. Dervaton of Publc Program Impacts P S IMPACTS S P* Actual Predcted D D EUI actual EUI predcted Q As the full publc program effect ncorporates more than EPA s ENERGY STAR mpacts, the ENERGY STAR evaluaton must account for several other publc programs and polces at the federal, regonal, and statewde levels that may not be controlled for n the model tself. For the most part, the most effectve approach to accountng for the many smaller programs s va adjustment of the total publc program effect usng publshed sources of program nformaton. Many of these publshed estmates may be taken at face value, whle others may requre refnement based on addtonal analyses. Concluson Ths paper has provded an overvew, and llustratons, of the work that s gong nto developng the frst natonal-level program mpact evaluaton of EPA s ENERGY STAR for the commercal buldng sector. As the dscussons n ths paper reveal, there reman many techncally dffcult ssues to negotate before ths work can be completed. These ssues nclude the functonal form of the econometrc model(s), the level of data aggregaton, and ssues related to specfc endogenous and exogenous varables. Despte the techncal obstacles, the econometrc approach to estmatng the market transformaton effects of EPA s ENERGY STAR appears promsng. As EIA s CBECS remans a unque source for natonal buldng consumpton and buldng characterstcs data,
ths approach offers the possblty of conductng ongong and perodc commercal sector program evaluatons. Also, t suggests that there may be many other energy effcency programs, ncludng those sponsored at the state and regonal level that may be able to take advantage of the nsghts and nnovatons of the present study to conduct future program evaluatons wth mnmal data collecton costs. Fnally, n addton to provdng program evaluaton fndngs, ths study offers useful sde benefts, such as updated energy prce elastcty estmates needed by economsts, envronmentalsts and energy forecasters.