Ecological Economics 68 (2009) 2336–2343

Contents lists available at ScienceDirect

Ecological Economics

j ourna l homepage: www.e lsev ie r.com/ locate /eco lecon

ANALYSIS

Joint environmental and cost efficiency analysis of electricity generation

Eric Welch a,⁎, Darold Barnum b

a Science, Technology and Environment Policy Lab, Department of Public Administration, College of Urban Planning and Public Affairs, University of Illinois, Chicagob Department of Managerial Studies, Department of Information & Decision Sciences, College of Business Administration, University of Illinois, Chicago

⁎ Corresponding author. 412 S. Peoria Street, CUPPA HChicago, IL, USA.

E-mail address: ewwelch@uic.edu (E. Welch).

0921-8009/$ – see front matter © 2009 Elsevier B.V. Adoi:10.1016/j.ecolecon.2009.03.004

a b s t r a c t

a r t i c l e i n f o

Article history:Received 17 September 2008Received in revised form 26 February 2009Accepted 6 March 2009

Keywords:Electricity generationCost and environmental efficiencyData envelopment analysisMaterial balance principleDEACarbonGreenhouse gas

Fossil-fuel based electricity generation produces the largest proportion of human-related carbon pollution in theUnited States. Hence, fuel choices by steam plants are key determinants of the industry's impact on national andglobal greenhouse gas emissions, and key foci for climate change policy. Yet, little research has been done toexamine the economic and environmental tradeoffs among thedifferent types of fuels that are usedby theseplants.This paper applies a Data Envelopment Analysis procedure that incorporates the materials balance principle toestimate the allocations of coal, gas and oil inputs that minimize carbon emissions and costs. Using EIA 906 andFERC 423 data, the paper estimates cost/carbon tradeoffs facing two sets of plants: those that use coal and gasinputs, and those that use coal, gas and oil inputs. Findings for our three-input sample show that there would be a79% increase in cost for moving from the cost-efficient point to the carbon efficient point, while there would be a38% increase in carbon for moving from the carbon efficient point to the cost-efficient point. These conclusionsindicate that, in general, the gap between efficient cost and efficient environmental production is wide, andwouldrequire substantial policy intervention, technological change or market adjustment before it could be narrowed.However, ourexaminationof individual plants shows thatwhat is true ingeneral is oftennot true for specific plants.Some plants that are currently less efficient than those on the production frontier could produce the same amountof electricitywith less carbon output and less fuel input. Additionally, many plants on the production frontier couldimprove both cost and carbon efficiency by changing their mixture of fossil-fuel inputs.

© 2009 Elsevier B.V. All rights reserved.

1. Introduction

During the past two decades, researchers, government officialsand even industry representatives have increasingly recognized thatgreenhouse gas emissions are contributing significantly to globalwarming. This is perhaps most obvious for the electricity generationindustry. The Environmental Protection Agency (EPA) calculates thatelectricity generation contributes approximately 39% of all human-related emissions of carbon dioxide in the United States (http://epa.gov/climatechange/emissions/co2_human.html#fossil). As govern-ment and industry begin to devise mechanisms for reducing green-house gases, analytical methods that provide better information topolicy making and industry decisions can result in improved eco-nomic and environmental outcomes.

One line of research has applied Data Envelopment Analysis (DEA),also called Activity Analysis (AA), to develop technical efficiencymodels for firms that include variables measuring pollution. Pollutionvariables usually have been included in the models as (1) inputs, (2)multiplicative inverse outputs, (3) outputs whose values are sums ofthe pollution values' additive inverses and a constant that will result in

all University of Illinois, 60304

ll rights reserved.

all sums being positive, or (4) weakly-disposable outputs (Färe et al.,1996; Scheel, 2001; Zaim, 2004; Kumar, 2006; Zhou et al., 2006; Färeet al., 2007; Chien et al., 2007; Burnett and Hansen, 2008; Zhou et al.,2008; Zhang et al., 2008; Camarero et al., 2008; Lozano and Gutiérrez,2008; Yang and Pollitt, 2008).

With respect to the electrical energy industry, Färe et al. (1996)were the first to include a pollution variable in their DEA methodol-ogy. A substantial number of DEA studies of electricity generation thatincorporate pollution variables have been published since that time.Zhou et al. (2008) include many of these through 2006 in their lite-rature survey of DEA studies related to energy and environment. Morerecent DEA studies of electricity generation that address pollutioninclude Chien et al. (2007), Färe et al. (2007), Burnett and Hansen(2008), and Yang and Pollitt (2008). None of these DEA studies haveincorporated carbon emissions into their models.

While these studies represent an impressive research base, theirimpact is limited by the fact that the findings offer no economicinterpretation: if a firm's technical efficiency declines (or increases)when a pollution variable is added to the model, the change providesno information about the economic cost (or benefit) of this outcome.For example, even though it is possible to show that an electricitygeneration plant using only natural gas will emit lower pollution buthave higher cost per unit of electricity produced than a plant usingonly low-grade coal, the aforementioned approaches cannot estimatethe economic consequences of the trade-off.

Fig. 1. Cost and environmental efficiency illustration.

2337E. Welch, D. Barnum / Ecological Economics 68 (2009) 2336–2343

Efficiency analysis in the electricity industry focuses on technicalefficiency alone, rather than on both technical and allocative efficien-cy, because validly estimating the costs of pollution from electricitygeneration is difficult (Färe et al., 1996; Söderholm and Sundqvist,2003). This situation may be improving. For example, the ExternEproject, one of the most recent and accurate cost estimation method-ologies, provides much-improved estimates of the marginal costs ofpollution emissions (Gully, 2006). Although not involving powergeneration, Munksgaard et al. (2007) use inputs of the amounts of sixpolutants needed to produce one Euro of output, with input weightsrestricted based on the range of estimated damage costs of eachpollutant, using costs based on ExternE and other recent estimates.

Kuosmanen and Kortelainen (2007) provide another recent at-tempt to use a DEA-type model to construct an Environmental Cost–Benefit Analysis in which the weights represent the shadow prices ofundesirable outputs. Thus, without directly estimating the prices, onlythose projects that have a positive value at the prices most favorableto the project will remain under consideration. The resultant shadowprices are refined invariousways inorder to choose among thoseprojectsstill in the running; although ExternE estimates are not mentioned,certainly these would assist in such refinement. For earlier examples ofthe use of shadow prices, see (Färe et al., 1993; Shaik et al., 2002).

In sum, there is a substantial history of using DEA to assess tech-nical efficiency in the electricity production industry. Researchershave incorporated pollution variables in various ways; however, themethod has generally been limited by its inability to provide an eco-nomic interpretation. While recent efforts have begun to develop newmethods that represent promising new avenues for DEA, the papernow turns to the explanation and application of one particular inno-vation, the DEA-materials balance principal model (DEA–MBP), whichmay be effectively employed to analyze both economic and carboninputs and outputs.

2. The materials balance principle

Lauwers et al. (1999) introduced use of the materials balanceprincipal (MBP) in efficiency measurement, and recently Lauwers(2009) published an article justifying its application as a “basiccondition for model adaptation in a more general environmental eco-nomics” context. Theoretical and methodological aspects of applyingtheMBP in a DEAmodelwere addressed in detail by Coelli et al. (2005)in a working paper, and the key aspects of this working paper haverecently been published (Coelli et al., 2007). All four of these studieswere illustrated with applications to Belgium pig-finishing operations,with one desirable output (pig meat) and two inputs (piglets andfeed). The pollutant material of interest, phosphorous, is found in thedesirable output and in both of the inputs, with the balance of thematerial being found in the manure output.

The MBP method is more closely tied to economic methodologythan is true of most other DEA approaches that deal with undesirableoutputs, thereby increasing its value when physical productivity, eco-nomic prices, and pollution costs all are of concern. Also, this methoddirectly addresses the material balance issue; in the pig-finishing casephosphorus in the inputs are balanced by the phosphorus accountedfor in the outputs. Prior efficiency research has often treated pollutantsas byproducts, which obviates the fundamental material connectionbetween inputs and outputs, and therefore does not consider pollutionabatement by allocation of inputs based on their contents. The DEA–MBP approach provides a promising approach to understanding botheconomic and environmental trade-offs inherent in energy produc-tion, as well as in other applications.

3. Contributions and organization of this paper

In this paper, we apply the DEA–MBP methodology to analyze theallocation of fossil-fuels in electricity generation, considering both the

carbon emissions and economic costs. For the case of carbonemissions in the electricity generation industry, an efficiency analysisthat considers the carbon content of different fuel inputs can helpidentify appropriate environmental tradeoffs.

The paper's primary contribution is that it demonstrates a meth-odology that can help inform the public policy debate over how tobalance the costs and benefits of carbon abatement by electricityutilities. Because of the predominate role of carbonpollution in climatechange, and the predominate role of fossil-fuel electricity generatingplants in carbon discharges, this methodology provides a worthwhilevehicle for setting out the benefits and costs of substituting one fuelfor another.

In addition, the paper also extends the application of the DEA–MBP methodology by illustrating how the procedure can be used toidentify the consequences of specific decisions for individual plants.For example, consider Plant 6085 in our second sample, which istechnically efficient. By changing its input mix to the cost-efficientpoint, it can lower both its costs and its carbon emissions by 12% each.However, if the plant changes its mix further in order to furtherdecrease carbon output, its costs will begin to increase. If the plantadjusts its fuel mix to attain the carbon efficient point, its carbonoutputwill decrease by 36% from the initial level but costswill increaseby 64%. As this example illustrates, this paper provides a methodologyfor detailed plant analysis that offers enhanced analytical decision-making at the individual plant level for regulators and managers.

The rest of this paper is organized as follows. First we briefly describethe DEA–MBP method. Then we apply the method to two samples of U.S. electrical plants for the 2002–2005 period. The first sample considersonly two types of fuel inputs in order to clearly demonstrate themethod.The second sample employs all three of the major fossil-fuel types inorder to more realistically examine the real-world decisions that can bemade. Finally we discuss our findings and draw conclusions.

4. A brief explanation of the DEA–MBP model

In this section we briefly describe the materials balance principle, asapplied to electric power generation. Comprehensive explanations of theMBP and its general applications can be found in the original MBP work(Lauwers et al.,1999; Coelli et al., 2005; Coelli et al., 2007; Lauwers, 2009).

Suppose an electricity generation plant wishes to produce a givenamount of electricity with two types of input — coal and gas. Thesetwo fuels are substitutes, of course, but they are not perfect substitutesbecause of boiler configurations and many other factors (Powell et al.,1977; Woodruff et al., 2005; Mazer, 2007). Fig. 1 presents a simpleillustration of the amounts of coal and gas inputs needed to produce

2338 E. Welch, D. Barnum / Ecological Economics 68 (2009) 2336–2343

a given amount of electricity, with the results of imperfect factorsubstitutability shown by the dotted line (a piece-wise isoquant). Forexample, a technically-efficient utility could generate a fixed amountof electricity with about 4.25 Btu of gas and 0.75 Btu of coal, withabout 1.75 Btu of each, or with about 0.75 Btu of gas and 4.25 Btuof coal.

The isoquant, that is the efficient frontier, is defined by those plantsusing the lowest amount of one input for a given amount of the other.Any plant on the line is technically efficient, and any plant using moreinput is technically inefficient. In the illustration, plants A and D aretechnically efficient and plants B and C are technically inefficient.

Although a plant will be technically efficient with any input com-bination on the isoquant, the place it should be on the isoquantdepends on input prices if it wishes to minimize total cost. To deter-mine this point, we need to draw an isocost line, with each point onthe line representing the combination of inputs available for a givensum. Specific to the isocost line as it is drawn in Fig. 1, for a givenamount of money a buyer has several purchase options: 5 Btu equi-valents of coal and no gas, no coal and 2.5 Btu equivalents of gas, orabout 2.5 Btu equivalents of coal and 1.25 Btu equivalents of gas. If theentire line is moved downward, the amount of money required for thereduced quantities will drop, and if it is moved upward the amount ofmoney required for the increased quantities will rise. Because atechnically-efficient plant must purchase enough inputs to be on theisoquant, theminimum cost will occur where the isoquant and isocostlines are tangent, which in the example is the input usage of plant D.Note that although plant A is technically efficient, it is not cost efficientbecause the line (and therefore total cost) has to move further up inorder to pay for A's input combination.

Finally, we come to the insight presented by Coelli et al. (2005,2007). The amountof pollutionper Btu can be considered the “price” ofthat pollution. So, just as we did with the prices we had to payfor inputs, we can just substitute the price of pollution, construct anisopollution line, and use it to find the technically-efficient combi-nation of inputs that will minimize pollution. For example, for theisocarbon line in Fig. 1, all of the preceding comments about costefficiency apply to carbon efficiency. With this new pollution indi-cator, we can compare plants based on their relative contribution topollution.

Additionally, we can compare the input ratios of a cost-efficientpoint on the isoquant and a pollution-efficient point on the isoquant.For our illustration in Fig. 1, the isopollution line for carbon emissionswould have a slope of−2.55/1.43≈−1.8. Coal produces 2.55 units ofcarbon per Btu while gas produces 1.43 units of carbon per Btu. Tomaintain a constant amount of carbon, therefore, it is necessary to giveup 1.8 Btu of gas for every additional 1 Btu of coal. This means thatthis isopollution line would obtain tangency with the isoquant wherethe ratio of total gas Btu being produced to total coal Btu being pro-duced is about 3 to 1, as represented by the coordinates of plant E.Knowing this informationwould allow us to estimate the cost per unitminimizing pollution with the current technology and input char-acteristics, which would be the cost of moving from the cost-efficientpoint represented by plant D to the pollution-efficient point repre-sented by plant E. It would be possible for policy makers to use thisinformation as a basis for setting the size of a pollution tax on theworst-polluting input or on the size of an anti-pollution subsidy forthe least-polluting input, in order to encourage plants to utilize inputcombinations that minimize total pollution.

5. Application to electricity generation

This analysis takes advantage of two publicly available EnergyInformation Agency (EIA) datasets that record the consumption andproduction of electric power plants in the United States. The FederalEnergy Regulatory Commission (FERC) Form 423 dataset containsmonthly cost and quality of fuels for regulated electric utility plants.

These data are collected on a monthly basis by FERC for all fuel typesused either for steam turbine or combined-cycle gas and steam tur-bine for plants with generating capacity of 50 or more megawatts(MW). These data have been collected for more than three decadessuch that the reporting procedures, data cleaning, and disclosureactivities are well understood and standardized (Energy InformationAdministration, 2009). Data included in the FERC Form 423 includefuel type, quantity, Btu content measured in millions of Btu (MBtu),sulfur content, ash content, cost, as well as contract information,origin and destination information. The EIA 906 dataset containsmonthly plant level data on fuel type, Btu consumption, electricitygeneration, and heat content collected from utilities and non-utilitiesthat generate at least 1 MW.

The resulting data include regulated electric power plants of 1 MWcapacity or larger for four years from 2002 through 2005. Fuel typeand cost data (cents per MBtu) from FERC 423 and fuel type, Btucontent of fuel consumed (MBtu and calculated metric tons of carbon)and electricity generation data (megawatt-hour — MWh) from EIA906 are used in the analysis. All data are accessible at the EIA website(Energy Information Administration, 2009).

We apply the preceding methodology to two samples of U.S. elec-tricity generation plants, using data from 2002 to 2005. In order tominimize random errors, we aggregate plant data over the four yearssuch that all of the variables for each plantmeasure the four-yearmeanvalue. Our first plant sample consisted of 40 plants and our secondsample consisted of 30 plants.We illustrate thefirst sample graphicallyand the second using DEA–MBP linear programming models. For ourfirst illustration we consider coal and gas inputs, and total electricityoutput from only these two fuels. For our second illustration we addoil inputs and use total electricity output from all three fuels.

We did not introduce labor and capital inputs for several reasons.First, we wanted to unambiguously apply the methodology to elec-tricity plant pollution in the simplest way possible. Second, the maindecision variables relate to fuel because labor input is a very small andrelatively unimportant part of input resources and capital stock isrelatively fixed. Third, in this industry, fuel, labor and capital are notsubstitutes but complements, and one valid way of dealing with thisfact when using traditional DEA models is to include only one of thecomplementary variables (Barnum and Gleason, 2006, 2008a).

In order to demonstrate that differences in capital stock do notaffect the efficiency scores in our samples, we performed two differentregression estimations to assess the effect of plant capacity onabsolute efficiency. Absolute efficiency in the electricity productionindustry is often measured using ratios such as (MWh of electricity)/(MBtu of input) as discussed byMazer (2007), where MBtu is millionsof Btu. This efficiency ratio avoids the bias and precision problems thatarise when DEA scores are used as response variables in regressions(Barnum and Gleason, 2008b).

Using all units from both samples for whichwe had necessary data,we (1) regressed (MWh/MBtu) on plant capacity measured by BoilerMaximum Steam Flow and Boiler Maximum Steam Flow Squared,and (2) regressed (MWh/MBtu) on plant capacity measured by MWand (MW)2. Neither was statistically significant: The first regressionyielded an R-square of 0.0009 [F=0.01, P(F(1, 44)N0.01)N0.9995].The second regression yielded an R-square of 0.003 [F=0.04; P(F(1,44)N0.04)=0.9595]. These results suggest that efficiency is not sig-nificantly related to capital stock, which means that differences incapital should not affect our efficiency scores. We did not test laborseparately because we did not have labor data, but the number ofemployees required is strongly correlated to plant capacity (Woodruffet al., 2005), so it is reasonable to assume that neither plant capac-ity nor employment level substantially influences efficiency for oursamples.

For the first sample of plants, we limited ourselves to those plantsthat produced at least 1% of their electricity from gas and at least 1%from coal, and, for the second sample, we included all plants that used

2339E. Welch, D. Barnum / Ecological Economics 68 (2009) 2336–2343

some of each of the three inputs, but with no more than 97% of theelectricity was generated from coal. Further, at least for those plantsin our samples for which applicable data were available, much of thegenerating capacity was supplied by boilers that could operate witheither two or all three types of fuel, and all of the plants had boilercapacity that allowed significant flexibility in fuel choice. For example,for those plants in the first sample, the mean plant percentage ofelectricity that could be produced by coal was 79% and that could beproduced by gas was 93%. For the second plant sample, themean plantpercentages of output that could be produced by coal, gas and oilrespectively were 72, 83 and 72%. These proportions may or may notbe typical for the industry as a whole, but they do demonstrate thatsignificant flexibility in fuel choice is not unusual.

The unit being analyzed in this paper, usually called the DecisionMaking Unit (DMU), is the electric power plant. However, the unit ofanalysis could be entire utilities, with each utility DMU containingmultiple plants. Assuming a utility's plants use different types of fuel,then it could shift production among them to achieve the optimal mix.Indeed, for those utilities capable of buying electricity from otherutilities, even more fuel flexibility would be possible. Likewise, shiftsfrom one type of fuel to another would be even more feasible if statesor even nations were used as the unit of analysis.

The methodology applies even if the rate of technical substitutionis completely linear over its entire range, with the only consequencebeing that the isocost and isopollution lines will intersect with the

Table 1First plant sample data, means of annual values, 2002–2005.

Plant name Plant code Utility name

A B Brown 6137 Southern Indiana Gas & Elec CoApache Station 160 Arizona Electric Pwr Coop IncAsheville 2706 Carolina Power and Light CompanyB C Cobb 1695 Cosumers Energy CoBay Front 3982 Northern States Power CoBlack Dog 1904 Northern States Power CoBlount Street 3992 Madison Gas & Electric CoBlue Valley 2132 Independence (City of)Chesterfield 3797 Virginia Electric and Power CoDan E Karn 1702 Consumers Energy CoDeerhaven 663 Gainesville Regional UtilGreene County 10 Alabama Power CoHamilton 2917 Hamilton (City of)Hawthorn 2079 Kansas City Power & Light CoIrvington 126 Tucson Electric Power CompanyJack Watson 2049 Mississippi Power CoKraft 733 Savannah Electric & Power CoLansing Smith 643 Gulf Power CompanyLon Wright 2240 Fremont City ofMcIntosh 6124 Savannah Electric & Power CoMuskogee 2952 Oklahoma Gas & Electric CoNeil Simpson II 7504 Black Hills Power & Light CoNortheastern 2963 Public Service Co of OklahomaNorthside 667 JEAO H Hutchings 2848 Dayton Power & Light CoPolk 7242 Tampa Electric CoQuindaro 1295 Kansas City (City of)R S Nelson 1393 Gulf States Utilities CoRawhide 6761 Platte River Power AuthorityRiver Rouge 1740 Detroit Edison CompanyRiverside 1081 MidAmerican Energy CompanyRodemacher 6190 Central Louisiana Electric CoS A Carlson 2682 Jamestown (City of)Silver Lake 2008 Rochester (City of)Sutherland 1077 IES Utilities IncTrimble County 6071 Louisville Gas & Electric CoUrquhart 3295 South Carolina Elec & Gas CoWabash River 1010 PSI Energy, IncWeston 4078 Wisconsin Public Service CorpYates 728 Georgia Power Co

Note: For this sample, mean 2002–2005 costs per MBtuwere $1.61 for coal and $4.99 for gas.and MBtu is 1 million Btu.

isoquant at one or the other of its end points rather than tangentscloser to its middle.

5.1. Graphical analysis: the first plant sample

Plant data from the first sample are shown in Table 1. We can showthe relationship between inputs, holding the output constant, bydividing each input by the output. These relationships are graphed inFig. 2, which also illustrates the isoquant, isocarbon and the isocostlines, with all based on four-year averages.

As can be observed in Fig. 2, the isoquant is linear over much ofits range, although the rate of technical substitution does changebefore reaching the endpoints. (Regressing the Gas Btu/kWh variableon the Coal Btu/kWh and (Coal Btu/kWh)-squared reports an R-square of 76%. Coal's regression coefficient is −2.4 and coal-square'sis +0.10 [t-statistics of 7.3 and 6.0 respectively], which indicates astrong curvilinear relationship convex toward the origin, which is thepattern of a conventional isoquant.)

Also of interest, some plants are substantially more efficient thanothers, and most of the plants do not appear to be mixing inputs suchthat theywill be cost efficient. Also, it is notable but not surprising thatthe points of tangency with the isoquant are different for the isocostand isocarbon lines, with the two points occurring at the oppositeends of the intermediate-rate-of-substitution section of the isoquant.This would mean that taxes on coal and or subsidies for gas would

State Electricity generated(MWh)

Coal consumed(MBtu)

Gas consumed(MBtu)

IN 3,113,275 30,294,398 790,320AZ 2,867,964 27,926,967 2,550,636NC 2,327,664 21,986,752 1,410,674MI 2,137,471 20,974,899 508,244WI 187,144 2,502,618 244,302MN 1,941,594 15,856,662 3,725,422WI 452,211 5,325,384 951,687MO 266,080 3,536,185 163,373VA 8,819,872 78,253,149 8,748,091MI 3,558,851 33,231,640 2,909,784FL 1,463,766 13,606,948 2,770,980AL 3,793,263 35,122,303 2,509,806OH 296,085 3,966,695 93,407MO 4,189,024 40,586,506 2,569,911AZ 1,068,182 6,854,478 4,753,060MS 3,953,298 37,629,418 1,666,149GA 1,206,994 12,728,829 987,059FL 3,988,193 20,492,696 14,586,950NE 472,943 5,487,513 132,764GA 1,050,721 10,851,116 1,101,141OK 10,068,396 105,792,226 2,044,087WY 755,527 8,194,842 600,839OK 9,387,075 68,427,176 23,656,034FL 1,268,612 7,672,278 5,699,362OH 730,010 8,669,126 272,008FL 1,518,371 13,640,274 1,885,875KS 844,955 9,429,930 218,790LA 5,438,674 37,642,747 18,584,254CO 2,148,128 21,833,538 540,271MI 2,687,304 25,605,350 749,231IA 690,258 8,354,412 349,435LA 3,849,802 33,651,611 8,964,868NY 197,010 2,498,037 367,034MN 285,503 3,297,863 150,411IA 854,383 10,158,074 306,901KY 4,044,295 39,376,529 1,396,389SC 1,130,380 5,719,220 4,435,170IN 5,153,723 45,426,084 6,839,605WI 3,399,125 35,898,099 648,762GA 6,076,761 61,873,256 2,203,385

Mean carbon per Btu is 2.55 units for coal and 1.43 units for gas. MWh is Megawatt-hour,

Fig. 2. First sample plant.

2340 E. Welch, D. Barnum / Ecological Economics 68 (2009) 2336–2343

have to be relatively high in order to induce power plants to changetheir fuel proportions for economic reasons.

More observations can bemade from Fig. 2. First, most facilities arelocated near the isocost–isoquant tangent point, a result that providessome face validity to the analysis. Second, the figure shows that mostof the plants are not technically efficient, and they could reduce boththeir costs and carbon outputs by becomingmore technically efficient.Third, there are a substantial number of plants – including some of thetechnically-efficient plants – that could reduce both their costs and

Table 2Second plant sample data, means of annual values, 2002–2005.

Plant name Plant code Utility name State

Allen S King 1915 Northern States Power Co MNApache Station 160 Arizona Electric Pwr Coop Inc AZAsheville 2706 Progress Energy Carolinas Inc NCC D McIntosh Jr 676 Lakeland City of FLChesterfield 3797 Dominion Virginia Power VADan E Karn 1702 Consumers Energy Co MIDeerhaven 663 Gainesville Regional Utilities FLE W Brown 1355 Kentucky Utilities Co KYGreene County 10 Alabama Power Co ALJohnsonville 3406 Tennessee Valley Authority TNKraft 733 Savannah Electric & Power Co GALansing Smith 643 Gulf Power Co FLManitowoc 4125 Manitowoc Public Utilities WIMcIntosh 6134 Savannah Electric & Power Co GANeil Simpson II 7504 Back Hills Corp WYNorthside 667 JEA FLPolk 7242 Tampa Electric Co FLQuindaro 1295 Kansas City City of KSR M Schahfer 6085 Northern Indiana Pub Serv Co INR S Nelson 1393 Entergy Gulf States Inc LARawhide 6761 Platte River Power Authority CORiverside 1927 Northern States Power Co MNRodemacher 6190 CLECO Power LLC LAS A Carlson 2682 Jamestown City of NYTrimble County 6071 Louisville Gas & Ekectric Co KYUrquhart 3295 South Carolina Electric & Gas Co SCVictor J Daniel Jr 6073 Mississippi Power Co MSWabash River 1010 PSI Energy Inc INYates 728 Georgia Power Co GAYorktown 3809 Dominion Virginia Power VA

Note: For this sample, mean 2002–2005 costs per MBtu were $1.53 for coal, $6.19 for gas andfor oil. MWh is Megawatt-hour, and MBtu is 1 million Btu.

carbon outputs by substituting gas for coal. These include the fourtechnically-efficient plants to the right of the isocost tangency pointwith the isoquant as well as the inefficient plants to their right thatare using about the same amount of gas but far more coal. Finally,the technically-efficient plant using the most gas could decrease bothcosts and carbon output by moving down the isoquant, thus usingless gas andmore coal, until it reaches the isopollution tangency point.In short, there are a substantial number of plants that could lowerboth fuel input costs and carbon output, a no-lose situation for allconcerned.

5.2. DEA–MBP models and procedures

We do not attempt to illustrate the second sample graphically be-cause it has three inputs, which is difficult to depict. Data are presentedin Table 2. We therefore use formal DEA–MBP analysis for the sample.

To determine whether a model with constant or variable returnsto scale should be used, we regressedMWhof electricity output on (1)plant size measured by Boiler Maximum Steam Flow and Boiler Max-imum Steam Flow Squared, (2) plant capacity measured by Mega-watts (MW) capacity and (MW)2, and (3) MBtu of energy input andits square. We included all plants from both samples for which wehad capacity data. All three equations were statistically significant, ofcourse, but in none of the three cases was the squared term statis-tically significant. Here are the results for the hypothesis that thesquared terms equal 0. (1) Fb0.005; P(F(1,43)N0.005)=0.9663. (2)F=0.29; P(F(1,43)N0.29)=0.5911. (3) F=1.26; P(F(1,43)N1.26)=0.2672. In sum, the hypotheses that there are either increasing ordecreasing returns to scale are rejected, so we selected a constantreturns to scale model.

Our DEA model to measure technical efficiency (1–4) is input orien-ted and, reflects constant returns to scale. All DEAs were conductedwith Tone's DEA-Solver software (2008). For each observation j=1,…, J

Electricity generated(MWh)

Coal consumed(MBtu)

Gas consumed(MBtu)

Oil consumed(MBtu)

3,156,562 26,639,437 19,409 4,571,2292,868,320 27,926,967 2,550,636 12,4062,355,908 21,986,752 1,410,674 348,2293,785,264 21,420,881 11,046,634 3,002,5608,975,334 78,253,149 8,748,091 1,271,8763,868,868 33,231,640 2,909,784 3,628,6251,527,861 13,606,948 2,770,980 776,4983,731,332 36,973,782 2,337,653 185,0223,831,755 35,122,303 2,509,806 484,0787,799,228 81,026,517 806,807 819,9181,226,153 12.738.829 987.059 220.2413,990,778 20,492,696 14,586,950 33,735309,165 3,106,869 17,315 1,524,2751,075,550 10,851,116 1,101,141 360,414756,477 8,194,842 600,839 14,7354,513,885 7,672,278 5,699,362 33,189,7411,613,025 13,640,274 1,885,875 871,625850,345 9,429,930 218,790 110,4709,676,052 106,723,273 466,263 2,456,6176,290,859 37,642,747 18,584,254 9,684,2392,149,848 21,833,538 540,271 18,0052,322,636 22,693,946 21,287 1,935,9803,956,719 33,651,611 8,964,868 1,165,716197,611 2,498,037 367,034 95524,047,756 39,376,529 1,396,389 36,5641,183,941 5,719,220 4,435,170 444,4599,430,611 63,615,040 22,985,123 76,6715,165,534 45,426,084 6,839,605 134,3406,084,117 61,873,256 2,203,385 74,9074,628,663 20,889,535 318,190 25,077,777

$5.48 for oil. Mean carbon per Btu is 2.55 units for coal, 1.43 units for gas, and 2.06 units

Table 3DEA cost results for second plant sample.

Plantcode

Original $ costper unit output

% Changes in cost per unit output

OrigToTE OrigToCE OrigToEE TEtoCE TEtoEE

Means 25.10 −6 −26 32 −21 4110 18.76 −0.2 −9 63 −9 63160 20.42 0 −16 49 −16 49643 30.52 0 −44 0 −44 0663 27.63 −16 −38 10 −27 31667 50.72 0 −66 −40 −66 −40676 31.06 −3 −45 −2 −43 2728 17.86 −4 −4 71 −0.1 79733 21.84 −13 −22 40 −10 601010 21.79 −5 −22 40 −18 471295 19.26 −10 −11 58 −1 771355 19.30 −6 −12 58 −6 681393 35.87 −13 −52 −15 −46 −31702 22.93 −5 −26 33 −21 411915 20.88 0 −18 46 −18 461927 19.57 0 −13 56 −13 562682 31.09 −33 −45 −2 −18 472706 18.79 −1 −9 62 −8 653295 32.63 0 −48 −6 −48 −63406 17.11 0 −0.3 78 −0.3 783797 20.14 −1 −15 52 −15 523809 37.03 0 −54 −18 −54 −184125 42.74 −34 −60 −29 −40 86071 17.06 0 0 79 0 796073 25.44 0 −33 20 −33 206085 18.56 0 −8 64 −8 646124 23.60 −14 −28 29 −15 516190 28.64 −15 −40 7 −30 266761 17.13 0 −0.4 78 −0.4 787242 23.13 −5 −26 32 −22 397504 21.59 −15 −21 41 −7 66

Notes: All first-row mean values are computed from column data. Plant 6071, in bold, isboth technically efficient and cost efficient. The other technically-efficient DMUs appearin italics. For all DMUs, cost per unit of electricity output is $17.06 for cost efficiency and$30.52 for environmental efficiency. Therefore, the percentage increase in cost formoving from the cost-efficient point to the carbon efficient point is (30.52−17.06)/17.06=78.9%.

2341E. Welch, D. Barnum / Ecological Economics 68 (2009) 2336–2343

there are data on n=1,…, N inputs and on m=1,…, M outputs,where xj = xj1; N ; xjN

� �aℝN

+ and yj = yj1; N ; yjM� �

aℝM+ . For both

samples, m=1; for the first sample, n=2 and for the second samplen=3. The DEA score θ estimates the technical efficiency of the targetDMU k.

minλ

θ ð1Þ

subject toXJ

j=1

xjnλj V θxkn n = 1; N ;N ð2Þ

XJ

j=1

yjmλj z ykm m = 1; N ;M ð3Þ

λj z 0 j = 1; N J: ð4Þ

Our DEA model to measure cost efficiency and environmentalefficiency (5–8) also is input oriented and reflects constant returns toscale. We show it in vector form.

cx⁎ = minxλ

cx ð5Þ

subject to x z Xλ ð6Þ

y0 V Yλ ð7Þ

λ z 0: ð8Þ

In this case, c is the vector of “prices,” which we assume arecommon for all observations, and x is the vector of fuel input amountsfor the target DMU. In the case of cost efficiency, the prices are therelative amounts paid for each of the three types of fuel input, andin the case of environmental efficiency the prices are the relativeamounts of carbon produced by each of the three fuel types. Thevector x⁎ contains the target DMUs' inputs that minimize cost (orcarbon), the matrix X contains the input values for all DMUs includedin the analysis, and the matrix Y contains the output values for allDMUs included in the analysis. The vector y0 contains the originaloutputs for the target DMU, and λ is the vector of intensity weights.

The procedure is as follows. First we estimate technical efficiency(TE), cost efficiency (CE) based on input prices, and environmentalefficiency (EE) based on the carbon content of the fuels. Using thesethree estimates, we project the Btu input from each fuel necessary toproduce 1 MWh of electricity (one unit of output) if a DMU is tech-nically efficient, if it is cost efficient, and if it is environmentallyefficient. Then, using the input price per Btu of each fuel, we canestimate the total cost per megawatt-hour of generated electricity foreach DMU based on its original inputs, its technically-efficient inputs,its cost-efficient inputs, and its environment-efficient inputs. Next, wecan perform the same estimates using input carbon content per Btuof each fuel, that is, total carbon per unit of electricity output for eachDMU based on its original inputs, its technically-efficient inputs, itscost-efficient inputs, and its environment-efficient inputs.

5.3. DEA analysis: the second plant sample

Cost and carbon outcomes for the second plant sample appearin Tables 3 and 4, respectively. Calculated means in the first row ofTable 3 show that on average plants would reduce costs by 6% if theywere technically efficient and increase costs by 32% if they wereenvironmentally efficient. Similarly, Table 4 means show that plantswould reduce carbon emissions by 6% if they attained technicalefficiency, and by 26% if they attained environmental efficiency.

The 6% ratio is particularly important. It indicates that if all plantswere to efficiently use currently-available technology, then both costs

and carbon output would decline by 6%. It would be worthwhile todeterminewhat changes would be necessary to increase the efficiencyof the most inefficient plants to the level of their efficient peers. Thismight be a relatively low-cost method of significantly reducing carbonpollution, especially because some of the expense would be offset byfuel cost savings.

In Table 3, note that plant 643 is environmentally efficient but doesnot have the highest cost per unit of electricity output. And, plant 6071is cost efficient, but, as shown in Table 4, does not have the highestcarbon output per unit of electricity output. Both plants are technicallyefficient. Thus, being the best on one objective does not necessarilymean that a plant will be the worst on the other.

The second DMU (plant 160) in Table 3 is a technically-efficientproducer; however, it is neither cost efficient nor environmentallyefficient. If it attained cost efficiency it would reduce its costs by 16%,and it would increase its costs by 49% if it became environmentallyefficient. The same DMU in Table 4 would reduce its carbon output by3% if it were to become cost efficient and by 30% if it were to becomeenvironmentally efficient.

This last example has an interesting and powerful implication: itis possible to identify technically-efficient power plants that couldsimultaneously improve cost and environmental efficiency. In fact,five of the 30 plants in Table 4 (160, 1927, 3406, 6085, and 6761), alltechnically efficient, could improve both cost and environmental ef-ficiency by moving to the cost-efficient point on the isoquant. Inaddition, two technically-efficient plants in Table 3 (667 and 3809),could decrease costs and carbon by moving to the environmentally-efficient point on the isoquant.

Technically-efficient DMUs can produce anywhere on the isoquant,so their costs per unit output will vary. However, cost-efficient DMUs

Table 4DEA carbon results for second plant sample.

Plantcode

Original carbonper unit output

% Changes in carbon per unit output

OrigToTE OrigToCE OrigToEE TEtoCE TEtoEE

Means 25.51 −6 2 −26 9 −2110 24.57 −0.2 3 −25 3 −25160 26.11 0 −3 −30 −3 −30643 18.35 0 38 0 38 0663 26.36 −16 −4 −30 14 −18667 21.29 0 19 −14 19 −14676 20.25 −3 25 −9 29 −6728 26.48 −4 −4 −31 0.02 −28733 28.00 −13 −10 −34 4 −251010 24.38 −5 4 −25 9 −211295 28.92 −10 −12 −37 −2 −291355 26.27 −6 −4 −30 2 −261393 22.66 −13 12 −19 28 −71702 24.91 −5 2 −26 7 −221915 24.51 0 3 −25 3 −251927 26.65 0 −5 −31 −5 −312682 35.00 −33 −28 −48 9 −212706 24.96 −1 1 −26 3 −263295 18.46 0 37 −1 37 −13406 26.86 0 −6 −32 −6 −323797 23.92 −1 6 −23 6 −233809 22.77 0 11 −19 11 −194125 35.86 −34 −29 −49 7 −236071 25.32 0 0 −28 0 −286073 20.71 0 22 −11 22 −116085 28.72 0 −12 −36 −12 −366124 27.89 −14 −9 −34 6 −236190 25.54 −15 −1 −28 17 −156761 26.28 0 −4 −30 −4 −307242 24.35 −5 4 −25 10 −217504 28.80 −15 −12 −36 3 −25

Notes: All first-row mean values are computed from column data. Plant 643, in bold, isboth technically efficient and environmentally efficient. The other technically-efficientDMUs appear in italics. For all DMUs, carbon per unit of electricity output is 25.32 unitsfor cost efficiency and 18.35 units for environmental efficiency. Therefore, thepercentage increase in carbon for moving from the carbon efficient point to the cost-efficient point is (25.32−18.35)/18.35=38.0%.

2342 E. Welch, D. Barnum / Ecological Economics 68 (2009) 2336–2343

and environmentally-efficient DMUs must produce at specific cost-efficient and environmentally-efficient points, which are tangent tothe isocost and isocarbon lines, respectively. This means that the cost-efficient and environmentally-efficient points will have carbon-costtrade-offs that are identical for all of them.

Cost efficiency results (Table 3) show that cost efficiency is at-tained at a cost per output unit of $17.06; this is the cost at which thereis no difference between technical efficiency and cost efficiency. Plant6071 is both technically and cost efficient, but it is environmentallyinefficient. To attain environmental efficiency it must increase costs by79% and decrease carbon outputs by 28%. Alternatively, environmentalefficiency is attained at a cost per output unit of $30.52. Plant 643 istechnically and environmentally efficient, but operates at a cost in-efficient point. To attain cost efficiency, it would reduce costs by 44%and increase carbon by 38%. The percentage increase in costs formoving from the cost-efficient point to the carbon efficient point forall plants is (30.52−17.06)/17.06=78.9%.

Results for environmental efficiency (Table 4) show that carbonefficiency is attained with 18.35 units (plant 643) while cost effi-ciency occurs at 25.32 units (plant 6071) for all plants in the popu-lation. This means that movement from the carbon efficient point tothe cost-efficient point would result in a 38% increase in carbonoutput: (25.32−18.35)/18.35=38.0%.

6. Discussion and conclusions

This paper applies a DEA method that is combined with materialbalance principles to jointly analyze fuel and pollution efficiency fromelectricity production. Several important conclusions are evident.First, based on our samples, both fuel cost and carbon pollution could

be lowered simultaneously, using current technology, by increasingthe technical efficiency of inefficient plants to a level closer to that oftheir efficient peers. Because carbon pollution is directly related to thenumber of Btus burned, increasing technical efficiency by decreasingBtu input while holding output constant would also increaseenvironmental efficiency.

Second, plants that already are technically efficient often can sim-ultaneously lower both their costs and their pollution levels. Forexample, 5 of the 12 technically-efficient plants in our second samplecould improve both their cost and environmental efficiencies bymoving to the cost-efficient point on the isoquant.

Third, there is a substantial gap between the isoquant–isocarbonand isoquant–isocost tangent points, which means that any techni-cally-efficient plant that exists at or between one of these points canonly decrease carbon by increasing costs, or only decrease costs byincreasing carbon. It would take large subsidies for gas or large pol-lution taxes on coal in order to change fuel-choice preferences of thissubset of technically-efficient plants.

Of course, institutional aspects of the industry would need to beintegrated into any attempts to apply our findings in a true policysetting. Beyond the competitive dynamics of the energy market, fuelsupplies are often secured in long term contracts and proximity to thesource of the fuel is a primary consideration. As a result, selection offuels that enable the plant to move toward the isocarbon efficiencypoint may be constrained by the characteristics of the fuel supplymarket (Joskow, 1985). Also, boiler loading procedures may also limitthe ability of plants to approach cost or environmental efficiencies.Incremental loading is a procedure in which different boiler units areselected to operate to generate electricity such that marginal costs areminimized. There is evidence that dynamic demand for energy makesit difficult to calculate fuel costs in advance (Powell et al., 1977; Vardiand vi-Itzhak, 1981).

In sum, our findings clearly point to a significant opportunity forthe application of the DEA–MBP method to help inform greenhousegas policy. While identifying specific policy lies outside of the scope ofthis paper, it is reasonable to consider new incentive systems toencourage the selection of technologies, operational techniques, fuelsuppliers and other factors that simultaneously comply with thedesire for cost efficiency and need for carbon reduction.

Acknowledgements

We sincerely thank the anonymous referees, who contributedsubstantial acumen. It was clear that they spent a significant amountof time with the paper, and they made detailed observations and veryvaluable suggestions. This research was partially supported by theGreat Cities Institute at the University of Illinois at Chicago, whosemission it is to conduct and support engaged, interdisciplinary, high-impact research and practitioner partnerships that address key urbanissues on a local and global scale.

References

Barnum, D.T., Gleason, J.M., 2006. Biases in technical efficiency scores caused by intra-input aggregation: mathematical analysis and a DEA application using simulateddata. Applied Economics 38, 1593–1603.

Barnum, D.T., Gleason, J.M., 2008a. Measuring the efficiency of electricity generatingplants with fixed proportion technology indicators. Accepted for presentation atthe 7th International Conference on Data Envelopment Analysis, July 2009,Philadelphia, http://ssrn.com/abstract=1319761 (Accessed 2/14/2009).

Barnum, D.T., Gleason, J.M., 2008b. Bias and precision in the DEA two-stage method.Applied Economics 40 (18), 2305–2311.

Burnett, R.D., Hansen, D.R., 2008. Ecoefficiency: defining a role for environmental costmanagement. Accounting, Organizations and Society 33 (6), 551–581.

Camarero, M., Picazo-Tadeo, A.J., Tamarit, C., 2008. Is the environmental performance ofindustrialized countries converging? A ‘SURE’ approach to testing for convergence.Ecological Economics 66 (4), 653–661.

Chien, C.F., Chen, W.C., Lin, Y.C., 2007. A case study to evaluate the productivity changesof the thermal power plants of the Taiwan Power Company. IEEE Transactions onEnergy Conversion 22 (3), 680–688.

2343E. Welch, D. Barnum / Ecological Economics 68 (2009) 2336–2343

Coelli, T.J., Lauwers, L., and Van Huylenbroeck, G., 2005. Formation of technical,economic and environmental efficiency measures that are consistent with thematerials balance condition. Centre for Efficiency andProductivity AnalysisWorkingPaper 06-2005, http://www.uq.edu.au/economics/cepa/docs/wp/wp062005.pdf(Accessed 12/17/2005).

Coelli, T.J., Lauwers, L., Van Huylenbroeck, G., 2007. Environmental efficiency measure-ment and thematerials balance condition. Journal of Productivity Analysis 28, 3–12.

Energy Information Administration, 2009. Official energy statistics from the U.S.government: Electricity database files, http://www.eia.doe.gov/cneaf/electricity/page/data.html (Accessed 2/14/2009).

Färe, R., Grosskopf, S., Lovell, C.A.K., Yaisawarng, S., 1993. Derivation of shadowprices forundesirable outputs: a distance function approach. The Review of Economics andStatistics 75 (2), 374–380.

Färe, R., Grosskopf, S., Tyteca, D., 1996. An activity analysis model of the environmentalperformance of firms—application to fossil-fuel-fired electric utilities. EcologicalEconomics 18 (2), 161–175.

Färe, R., Grosskopf, S., Pasurka, J., 2007. Pollution abatement activities and traditionalproductivity. Ecological Economics 62 (3–4), 673–682.

Gully, F., 2006. Social choice, uncertainty about external costs and trade-off betweenintergenerational environmental impacts: the emblematic case of gas-basedenergy supply decentralization. Ecological Economics 57 (2), 282–305.

Joskow, P.L., 1985. Vertical integration and long-term contracts: the case of coal-burningelectric generating plants. Journal of Law, Economics, & Organization 1 (1), 33–80.

Kumar, S., 2006. Environmentally sensitive productivity growth: a global analysis usingMalmquist–Luenberger index. Ecological Economics 56 (2), 280–293.

Kuosmanen, T., Kortelainen, M., 2007. Valuing environmental factors in cost–benefitanalysis using data envelopment analysis. Ecological Economics 62 (1), 56–65.

Lauwers, L., 2009. Justifying the incorporation of thematerials balance principle into frontier-based eco-efficiency models. Ecological Economics 68 (6), 1605–1614 (15 April).

Lauwers, L., Van Huylenbroeck, G., Rogiers, G., 1999. Technical, economic and environ-mental efficiency analysis of pig fattening farms. Ninth European Congress ofAgricultural Economists. 24 August 1999.

Lozano, S., Gutiérrez, E., 2008. Non-parametric frontier approach to modelling therelationships among population, GDP, energy consumption and CO2 emissions.Ecological Economics 66 (4), 687–699.

Mazer, A., 2007. Electric Power Planning for Regulated and Deregulated Markets. IEEEPress, Hoboken, N.J.

Munksgaard, J., Christoffersen, L.B., Keiding, H., Pedersen, O.G., Jensen, T.S., 2007. Anenvironmental performance index for products reflecting damage costs. EcologicalEconomics 64 (1), 119–130.

Powell, J., Gleason, J., Burton, J., 1977. Power generator scheduling by dynamicprogramming. European Journal of Operational Research 1 (3), 154–160.

Scheel, H., 2001. Undesirable outputs in efficiency valuations. European Journal ofOperational Research 132 (2), 400–410.

Shaik, S., Helmers, G.A., Langemeier, M.R., 2002. Direct and indirect shadow price andcost estimates of nitrogen pollution abatement. Journal of Agricultural ResourceEconomics 27, 420–432.

Söderholm, P., Sundqvist, T., 2003. Pricing environmental externalities in the powersector: ethical limits and implications for social choice. Ecological Economics46 (3), 333–350.

Tone, K., 2008. Saitech DEA-Solver, http://www.saitech-inc.com/products/prod-dsp.asp(Accessed 2/14/2009).

Vardi, J., vi-Itzhak, B., 1981. Electric Energy Generation: Economics, Reliability, andRates. MIT Press, Cambridge, Mass.

Woodruff, E.B., Lammers, H.B., Lammers, T.F., 2005. Steam Plant Operation. McGraw-Hill, New York.

Yang, H. and Pollitt, M., 2008. Incorporating both undesirable outputs and uncontrol-lable variables into DEA: The performance of Chinese coal-fired power plants.European Journal of Operational Research, In Press. doi:10.1016/j.ejor.2007.12.052.

Zaim, O., 2004. Measuring environmental performance of state manufacturing throughchanges in pollution intensities: a DEA framework. Ecological Economics 48 (1),37–47.

Zhang, B., Bi, J., Fan, Z., Yuan, Z., Ge, J., 2008. Eco-efficiency analysis of industrial systemin China: a data envelopment analysis approach. Ecological Economics 68 (1–2),306–316.

Zhou, P., Ang, B.W., Poh, K.L., 2006. Slacks-based efficiency measures for modelingenvironmental performance. Ecological Economics 60 (1), 111–118.

Zhou, P., Ang, B.W., Poh, K.L., 2008. A survey of data envelopment analysis in energy andenvironmental studies. European Journal of Operational Research 189 (1), 1–18.