Changes in Electricity Demand in the United States from the 1970s to 2003 Marvin J. Horowitz* Gaining a better understanding of how electricity demand has changed over the past three decades, particularly in light of government involvement in influencing electricity demand, is an important step towards shaping energy policy in the U.S. and internationally. This study of U.S. electricity demand finds that those states that have moderate to strong commitment to energy efficiency programs reduce electricity intensity relative to what it would have been with weak program commitment; in the residential sector by 4.4 percent, in the commercial sector by 8.1 percent, and in the industrial sector by 11.8 percent. The findings are similar with respect to levels of electricity consumption in the commercial and industrial sectors, but not the residential sector. Moreover, the evidence in this paper indicates that energy efficiency program commitment in all three sectors of the U.S. economy has transformed electricity demand with respect to three key economic variables; electricity price, income as measured by per capita income or gross state product, and technological change. Also, this study finds that nationwide spillover from energy efficiency programs may be rapid and ubiquitous in the residential sector. Estimates of the impact of California’s energy efficiency programs confirm the speculation that these efforts have dramatically reduced state electricity intensity; to date, these impacts are likely the upper bounds of the impacts of strong state-level energy efficiency program commitment. 1. INTRODUCTION Growing constraints on the production and transmission of electricity make it critically important that decisionmakers better understand the factors correlated with, not to mention directly driving, its use. A better understanding is especially timely now that governments around the world have become increasingly involved The Energy Journal, Vol. 28, No. 3. Copyright ©2007 by the IAEE. All rights reserved. *
President, Demand Research, Fairfax, Virginia. Email:
[email protected]
This research was funded by the U.S. Environmental Protection Agency. The author would like to thank Caterina Hatcher, Jason A. Horowitz, Barry Brent, and the anonymous referees for their helpful comments and suggestions. The author takes sole responsibility for any errors.
93
94 / The Energy Journal in efforts to influence energy consumption, involvement that often results in new regulations, policies, and programs. For some of these, like building and appliance codes, compliance is mandatory; for others, like technical training, equipment rebates, and civic recognition, participation is voluntary. These programs, policies, and regulations are planned, funded, administered, and implemented in a variety of ways, sometimes through local governments, other times through local utilities, and still others times through regional or federal agencies. In this study, energy efficiency programs refer collectively to all such governmental efforts, irrespective of organizational origin or mode. For a detailed examination of the institutional foundations of many of these programs in the U.S., see Gillingham et al. (2004). This study is inspired, in part, by frequent comparisons of U.S. and California residential per capita electricity use, as shown in Graph 1.A. Versions of this graph have been presented in many different professional papers and forums, mainly by the staff members of public agencies and research institutions in California, one example of which is Rosenfeld (2005). It is exhibited as evidence that California’s aggressive energy efficiency program portfolio achieved the intended effect of damping residential electricity demand in California, weather and other state-specific conditions notwithstanding. Graph 1.A – 1.C. California and US47 Electricity Intensity, by Sector
The trends in Graph 1.A clearly show that from 1970 to 2003 residential per capita use in California grew at a much slower rate than that of the other contiguous 47 states (when combined into a single entity that excludes California, Hawaii, and Alaska, this union is referred to as US47). This is particularly true after 1976, about the time California launched its residential energy efficiency programs. These were among the first energy efficiency programs in the nation; not long after, wide-ranging commercial and industrial sector programs were adopted in California and elsewhere. Nevertheless, as is well-known, California leads all other states in promoting energy efficiency. According to a recent report co-published by the California Public Utilities Commission and the California Energy Commission (2006), since 1975 building and appliance standards have reduced energy costs for individuals and businesses in California by $56 billion, and these standards are expected to save another $23 billion by 2013. Further, according
Changes in Electricity Demand in the United States / 95 to a retrospective analysis of California’s energy efficiency programs by Global Energy Partners (2003), in 2001 California had 281 energy efficiency programs administered by 35 different entities. Collectively, they are said to have produced annual energy savings of 4,760 GWh. To complete the picture painted by the residential graph, Graph 1.B and 1.C contains similar trend comparisons for the commercial and industrial sectors. The construction of these graphs follows a similar approach to the construction of the residential graph, which plots electricity intensity, defined as total residential electricity consumption, in MBTU, per state population. Commercial sector electricity intensity in California and US47 is characterized by MBTUs of delivered electricity per dollar of the commercial or service sector component of gross state product (GSP), and in the industrial sector electricity intensity is characterized by MBTUs of delivered electricity per dollar of the manufacturing sector component GSP. All dollar values in these graphs, and in the remainder of this study, are standardized to the year 2000 using the gross domestic product (GDP) implicit price deflator. As Graph 1.B and 1.C illustrate, the commercial and industrial sector trends in California appear to be more favorable than the US47 trends. In both sectors, electricity intensity appears to be declining at a faster pace in California than in US47. While the differential trends may not be as pronounced as in the residential sector, they seem to suggest that strong energy efficiency program commitment has consequential impacts on electricity intensity in these sectors, too. The goal of this study is to delve into three outstanding questions provoked by Graph 1.A – 1.C. First, these graphs raise the fundamental question of how electricity demand changed in the U.S. over the past three decades. This question is complex because electricity demand can be thought of in two ways; as it relates to electricity intensity and as it relates to electricity consumption. This distinction is highly relevant in today’s national and international energy policy discussions. As noted by Herring (2006), in the mid-19th century the English economist William Stanley Jevons identified a “paradox” suggesting that decreases in energy intensity can lead to increases, rather than decreases, in energy consumption. It is therefore of practical interest, as well as historical interest, to shed light on both facets of electricity demand. The second question is an extension of the first. If electricity demand changed, do these changes differ between states depending on their degrees of commitment to encouraging energy efficiency? Third and last, does program commitment have observable, long-term impacts on the behaviors that affect electricity demand? To answer these questions, a particular research design, array of data, and set of analytical tools have been brought together. Each of these components, described in the next section, is well-accepted, transparent, and replicable. This makes the entire approach adaptable to any collection of well-defined geographic area, such as discrete communities, utility service territories, counties, provinces, or nations, for which basic economic and energy data are available.
96 / The Energy Journal 2. Data and Research Design This effort extends a genre of energy economics research – cross section time series reduced-form demand analysis – that has yielded, over many decades, important insights into national or aggregate energy use. Listings of older studies and brief descriptions of their methods and findings can be found in Dahl (1992) as well as several earlier literature reviews. Like these studies, the methods and data used in this study preclude a detailed structural analysis of demand involving systems of equations, and instead focus on a few basic correlates of electricity demand. Following the more recent examples of Bernstein, et al. (2003), Loughran and Kulick (2004), and Horowitz (2004), this study begins by estimating models of energy intensity. Energy intensity is a subject about which there has been a rapidly increasing international literature, e.g., World Energy Council (2001) and Bosseboeuf, et al. (2005), particularly because of global efforts to address climate change. It is typically constructed by using a variable in the denominator that closely affects the level of energy consumption, but is not affected by it. This facilitates statistical analysis by controlling for a key source of unexplained variability, allowing for direct comparisons between different time periods and cross sections. However, as noted above, while energy intensity and energy consumption are closely related, they can move in different directions, which in turn will have different implications for energy policy. Therefore, key findings in this study are expressed in terms of energy consumption, too. To implement the study agenda, residential, commercial, and industrial sector electricity intensity models are given a uniform set of variables and a common functional form. The independent variables for each sector are, respectively, the average retail price of electricity (P); the average retail price of natural gas (N); state economic status as captured by per capita GSP or per capita income (G); a time trend related to technological change in each sector (T); and, climatic conditions, i.e., heating (H) and cooling (C) degree days. The general function is: EIt,s,R = f (Pt,s,R,Nt,s,R,Gt,s,R,Tt,R,Ht,s,R,Ct,s,R)
(1)
in which subscript t represents a given year; subscript s represents a given state; and, subscript R represents a discrete level of commitment to energy efficiency programs. This split-case function asserts that R influences each of the behavioral relationships associated with electricity intensity. In other words, R is a transformative agent. This point of view is not new; the possibility that government programs and advertising can alter price, income, and other elasticities by transforming the underlying utility function of consumers is firmly established in economic theory, e.g., Becker and Murphy (1993) and Johnson and Myatt (2006). For practical purposes, the independent variables are taken to be exogenous and free of feedback, notwithstanding the awareness that from a structural perspective this premise is never rigorously correct.
Changes in Electricity Demand in the United States / 97 Although there are literally hundreds of energy efficiency program reports that come mainly from utility and state-level program evaluations and policy analysis, accurate quantification of R is problematic. This is because most of these program reports and evaluations are specific to individual programs, service territories, and years. To overcome this data limitation, the 48 states (with the District of Columbia merged into Maryland) are ranked and placed in quartiles – one strong (S) , two moderate (M), and one weak (W) – based on a general indicator of state support and encouragement of energy efficiency, referred to as program commitment. To create an indicator of the strength of program commitment, the Energy Information Administration (EIA) Form EIA-861 electric utility demand side management (DSM) annual total sector-level energy savings data from 1992 to 2003 were aggregated by state. Annual DSM expenditures, another indicator of program commitment reported on Form EIA-861, is not germane to the focus of this study because it is reported as total utility expenditures on DSM, not as separate expenditures for residential, commercial, and industrial sector programs. Without such differentiation, the sector-level analyses in this study would be misleading since there are states that clearly have strong or weak program commitment in one or two sectors, but not all three. Once aggregated, each state’s annual total sector-level energy savings, i.e., the energy savings achieved from the year’s activities as well as those from all previous years’ activities, were corrected for misreporting to make it consistent with EIA’s concept of cumulative. However, since additional information does not exist on energy savings degradation, no further adjustment was made to the corrected cumulative totals. Data correction efforts notwithstanding, it should be noted that the data carry other misreporting and measurement problems, too. For example, for several years the Tennessee Valley Authority (TVA) reported all of its DSM program savings as state of Tennessee residential sector savings only, despite the fact that TVA’s services spanned seven southern states and all three sectors of the economy. As a result, all of Tennessee’s DSM savings, and the TVA-motivated energy savings in the six other states, had to be treated as missing. However, this was the only instance where savings was dropped. Given the Form EIA-861 DSM data shortcomings, many of which have been called attention to over the years, the energy savings data are not used literally, but rather as indicators of energy efficiency program commitment. This ordinal ranking scheme is one that most energy service industry professionals would accept as largely representing relative energy efficiency program commitment. For example, the majority of states in the S and W quartiles match the state rankings of the annual state scorecard for energy efficiency programs produced by the American Council for an Energy Efficient Economy, e.g., Kushler and Witte (2001) and York and Kushler (2006). A complete alphabetical listing of states within each quartile, by sector, can be found in Table A5 of Appendix A. Within the split-case research design, two distinct time periods are defined. In the early or base period, meaning the years prior to 1992, many states ex-
98 / The Energy Journal perienced either no programs at all or the very beginnings of programs. In the latter period, 1992 and beyond, a small number of states had aggressive, mature energy efficiency programs, a larger number had newer programs, and a small number continued to have little or no involvement with energy efficiency programs. Although the cutoff between periods is obviously inaccurate for individual states (for example, it is well-established that California, Wisconsin, and Massachusetts, had strong program commitments well prior to 1992), this cutoff year is useful for a variety of important reasons. For one, it was the first year in which a major, new national energy policy, the Energy Policy Act of 1992, took effect. For another, it was the first year in which large scale national programs like the Environmental Protection Agency’s Green Lights and ENERGY STAR became operational. Also, as noted above, it is the first year that EIA’s DSM savings data are available for all three sectors of the economy. Nevertheless, it is essential to acknowledge the tradeoff in imposing uniform program periods for each of the three sectors. As will be addressed later in this study, this restriction motivates additional investigation into how energy efficiency programs may have transformed electricity demand in the residential sector. In addition to DSM data, EIA collects and maintains the energy sales and price data employed in this study. These data, which due to deregulation went through major revisions in 2005 for the years leading up to 2003, extend back as far as 1970. Another key dataset, GSP, extends as far back as 1977; it, too, went through major revision in 2005 for the years leading up to 2003, such that NAICS classifications now replace SIC classifications for 1997 and beyond. The remaining data used in this study are derived from other federal agencies; state personal income, population, and GSP data, as well as the implicit price deflator, are taken from the Bureau of Economic Analysis. Weather data are from the National Climatic Data Center, which weights each state’s heating and cooling degree day data by intra-state populations. Technology trend data are from the Federal Reserve Board. For this study, different technology trends are employed in the sector-specific analyses so as to best match the economic and energy-using characteristics of each sector. For the residential sector, market group index B51121 is employed; it represents U.S. production of a subset of electronic equipment, namely, computers, video, and audio equipment. On the other hand, for the commercial sector, market group index B52120 is employed; it represents the production of products that are more closely associated with the business world such as information processing and related equipment. Finally, for the industrial sector, market group index G334 is employed. It represents computers and related equipment and is the broadest category among the three trends. With the divisions of years and states, program contrast is available via four grouping of observations; (1) states with strong or moderate program commitment in the base period; (2) states with strong or moderate program commitment in the treatment period; and, (3) and (4) states with weak program commitment in both periods. Using this research design, counterfactual and decomposition analy-
Changes in Electricity Demand in the United States / 99 ses were performed. The former permits estimation of the impact of energy efficiency program commitment on electricity demand. It does so by simulation, that is, by combining the covariates of one group of observations with the behavior of a parallel group of observations. In the present context, the counterfactual analysis reveals what energy use would have been for S and M states combined (hereafter referred to as SM states or quartiles), given their existing characteristics, had they behaved like W states. Afterwards, decomposition of the differences in group means of the dependent variables, based on Blinder (1973), is used to better understand the changes in electricity demand between quartiles and years. 3. Impacts of Program Commitment Using the framework just described, the impacts of program commitment can be measured using difference in differences (DID) estimators such as those described by Meyer (1995). The simplest and most widely-used assumes that the SM states and the W states are identical to each other in their general characteristics in each period, and further, are identical in their behaviors relative to key non-program related determinants of energy use in each period. Under this highly restrictive assumption, all of the differences in the changes in mean energy intensity can be ascribed to the treatment, that is, energy efficiency program commitment. This estimator, referred to here as DIDMeans, is: – – –W – DIDMEANS = ( EITSM – EIBSM ) – ( EITreat – EI WBase). reat ase
(2)
– where EI represents mean electricity intensity; the superscript SM represents the group of 36 states in the strong and moderate program commitment quartiles; the superscript W represents the group of 12 states in the weak program commitment quartile; and, the subscripts represent base and treatment periods. A second commonly-used difference in differences estimator, referred to as DIDCoeff, involves the use of regression models to estimate the coefficients of an indicator or dummy variable representing the mean difference in the behavior of the SM group of states versus the W group in the base and treatment periods. In the context of this study, panel models would be estimated that contain the SM and W quartiles combined, i.e., all 48 states: W W EIBSM = α0,Base + αjXSM + αkDSBaMse + eSBaMsWe ase j,B ase
(3)
W W M W EITSM = α0,Treat + αjXSM + αkDSTreat + eTSM . rea t j,Treat rea t
(4)
In these models, the Xj’s represent the various independent variables, the αj’s are the behavioral coefficients associated with the independent variables, the α0’s are the single model intercepts for each equation, and the e’s are the model error terms. Most importantly, the αk’s are coefficients associated with D, a vari-
100 / The Energy Journal able that takes a value of 1 if a state is in the W quartile, and 0 otherwise. The difference in differences estimator is then: DIDCOEFF = αkDTSM – αkDBSM reat ase
(5)
where the expressions αkDTSM and αkDBSM on the right side of the equation are reat ase taken from the estimated equations (3) and (4). Of critical note regarding the DIDCoeff estimator, the introduction of D into the electricity intensity models requires that fixed state effects be excluded from the specification. Hence, although this approach relaxes some of the restrictions of the simpler approach, in the context of this study it contains two major flaws. First it overlooks the pre-existing, fixed, energy-related differences between states that are due to regulatory regimes and other state-specific conditions. Second, it ignores the likelihood that the groups of states with different levels of commitment to energy efficiency programs may have different behavioral responses to key non-program related determinants of energy use. To overcome the limitations of these two conventional difference in differences estimators, this study offers an alternative approach to estimating the impact of program commitment on energy use, DIDCF, representing the difference in differences approach using counterfactuals. It entails the specification of a state fixed effects panel model of electricity intensity through which a simulation is performed. The simulation estimates what might have occurred to one group of observations had they experienced the conditions of the other group of observations. Because the condition in question is program commitment, and because contrasting program commitment across states is nearly impossible to identify with precision, the most viable groups of observations that can be used for estimating the electricity intensity model will be those on the opposing ends of program commitment. Hence, one of two models may be estimated such that: n
EI W = bW0 + Σ bWj X Wj + uW, or j=1
n
EIS = bS0 + Σ bSj XSj + uS j=1
(6) (7)
where again, the superscripts W and S represent states that fall into one or the other levels of R, that is, either weak or strong program commitment. In these representations, time series and state subscripts are suppressed, the βj’s are the coefficients associated with each of the Xj independent variables, and the u’s are independent error terms. For any collection of years over which these models are estimated, the counterfactual is estimated by inserting the mean X’s from one group into the other group’s model. For example, using the W quartile’s model, the values for the SM states can be used for calculating the counterfactual such that: n
EICSMounterfactual = bW0 + Σ bWj XjSM. j=1
(8)
Changes in Electricity Demand in the United States / 101 In this way, the electricity intensity model is used to simulate the mean electricity intensity of SM states had their demand behavior been like that of the W states. Once the counterfactual is estimated for the SM states, DIDCF is calculated by subtracting the change in their factual mean values from the change in their counterfactual mean values: – – – – DIDCF = ( EITSM – EIBSM ) – ( EICSMounterfactual, Treat – EICSMounterfactual, Base) (9) reat ase This estimate is referred to as net impact and is reported throughout this study in percentage terms for both electricity intensity and the level of electricity use. Prior to reporting the results of these analyses, it is instructive to inspect the sector-level electricity intensity trends. All things being equal, differential trends in the treatment period are to be expected if program commitment has an effect, if the states are properly categorized, and if the dividing year between periods is relatively accurate. Graphs 2.A to 2.C illustrate these trends, which are constructed in the identical fashion to the earlier Graph 1.A to 2.C. In these graphs, the unshaded areas contain the trends for the base periods and the shaded areas contain the trends for the treatment periods. For each of the three sectors, it is apparent that the base period trends are similar, save for minor differences. In the industrial sector base period, the direction of change in intensity differs initially between the quartiles, but eventually track each other fairly closely. In the commercial sector, the trends track each other even as the relative magnitude of the changes appear to be stronger for the W quartile. In the residential sector the magnitude of the changes seems to be smaller for the W quartile than the SM quartiles, but otherwise the trends track closely. In contrast to the base period, the trends in the treatment period show substantial differences in the commercial and industrial sectors. In the commercial sector the SM quartile trend declines relative to the W quartile trend. In the industrial sector the W quartile trend is seen to rise relatively steadily, while the SM quartile is neutral, if volatile. In the residential sector the trends go in the same upward direction, but the W quartile’s trend appears to rising more slowly than that of the other quartiles. Graph 2.A – 2.C. Electricity Intensity in the Base and Treatment Periods
102 / The Energy Journal As the fulcrum on which this study’s findings turn is the contrast between program commitment levels, it is important to bear in mind that any of the following conditions will blur the contrast in the research design and produce bias: (a) strong or moderate program commitment in the base period for one or more of the SM states, (b) strong or moderate program commitment in the treatment period for one or more of the W states; (c) misplacement of one or more of the M states into either the S or W state quartiles; or, (d) spillover from one or more of the SM states to one or more of the W states in either the base or treatment period. Blurring of contrast caused by any of these conditions necessarily produces bias in favor of no effect or negative effect. Likewise, misplacement of the beginning year of the treatment period, or the end year of the base period, will also result in an underestimation of program commitment impacts. However, since there are no detailed data representing the exact timing and magnitude of program commitment in each of the 48 states over the entire study period, the cutoff year used to differentiate program commitment represents an objective, consistent, and practical compromise. From a modeling standpoint, the validity of the counterfactual findings relies on the assumption that in the absence of differing levels of program commitment the energy use trends of the quartiles would not greatly differ. In this light, the graphs of the base period trends offer evidence of quartile comparability and suggest that it is appropriate to use the electricity intensity model of the W quartile to simulate the counterfactual for the other quartiles. Even stronger evidence of their comparability is found in Table A1 in Appendix A, which contains descriptive statistics in their actual, as opposed to logarithmic form, for the SM quartiles and the W quartile for all the variables in the models. As can be seen, the distributions of the variables within sectors are generally similar in the base period. Despite the apparent similarities in the industrial sector in the SM and the W quartiles, the comparability of the structure, or industrial mix, of the quartiles might yet be called into question. Hence, to explore differential industrial mix further, EIA’s 2002 Manufacturing Energy Survey (MECS) was used to identify the the top ten electricity-consuming industries nationally. These top ten industries, of which the largest electricity users are chemicals, primary metals, and paper, accounted for the overwhelming majority of industrial sector electricity use in 2002, a total of 82.9 percent. In 1977, these ten industries accounted for 57 percent of manufacturing GSP in the W quartile, and 62 percent of manufacturing GSP in the SM quartiles; moreover, the proportions remain much the same through the years, such that by 2002 they account for 61 percent of manufacturing GSP for the W quartile and 66 percent for the SM quartiles. These general facts point to the overall mix of industries being comparable across quartiles, lending credence to the validity of the industrial sector counterfactuals. Table 1 contains the impact estimates of program commitment, in levels of electricity consumption, or GWh, as well as in electricity intensity, for the SM states, based on the estimated sector-level models for the W states. The W quartile models for the base period are provided in Table A2 of the Appendix A, and Table A3 contains the models for the treatment period.
Changes in Electricity Demand in the United States / 103
Table 1. Impacts of Energy Efficiency Program Commitment Level of Electricity (GWh) Use Electricity Intensity Statistic
Base
Treatment % Change
Base
Treatment % Change
Residential Factual Counterfactual 95% C.L. (+/-) Net Impact
670,272 693,203 15.1%
949,968 919,711 8.2%
41.7% 32.7% 9.1%
12.1 11.8 6.1%
14.5 13.9 3.1%
19.3% 17.7% 1.6%
627.4 689.6 3.4%
578.7 692.2 5.2%
-7.8% 0.4% -8.1%
2,928 2,787 2.9%
2,898 3,088 4.1%
-1.0% 10.8% -11.8%
Commerical Factual Counterfactual 95% C.L. (+/-) Net Impact
558,303 664,761 22.4%
865,047 1,111,229 34.5%
Factual Counterfactual 95% C.L. (+/-) Net Impact
675,353 925,015 21.9%
793,024 1,175,135 31.0%
54.9% 67.2% -12.2%
Industrial 17.4% 27.0% -9.6%
For each model, the reported R-squared is derived from a deviation-frommeans specification, all of the variables are in natural log form, and a GLS fixed effects, weighted least squares estimator is employed. This estimator minimizes heteroscedasticity by allowing for different residual variances for each cross section; in addition, the fixed effects accounts for unobserved, state-specific factors, such as location-specific and regulatory conditions, that permanently affect state electricity use. Diagnostics indicate that the estimated equations need not be corrected for unit roots or autocorrelation. All the standard errors are White-corrected. For each sector, the first row contains the factual, annual mean GWh use and electricity intensity for the SM states for the base period and the treatment period, and the percentage change from the base period. The second row provides the same information for the counterfactuals, and the third row contains an approximation of the 95 percent confidence levels of the counterfactuals based on the average standard errors, or relative standard errors in the case of electricity intensity, of the individual observations. Finally, the fourth row contains the difference in differences, or net impact, in percentage terms. This statistic represents the difference between what the change for the SM quartiles actually was, and what the simulation said it would have been had these states behaved like the states in the W quartile. By definition, the net impact is the aggregate effect of state energy efficiency program commitment on electricity demand. In the residential sector, the findings indicate that the net effect of program commitment for the SM states was an increase of 9.1 percent in average annual GWh use from the base to the treatment period. One interpretation of this findings is that program commitment in the residential sector has the reverse effect
104 / The Energy Journal than intended. However, this interpretation is called into question by the parallel finding for electricity intensity, whose change, while also in the opposite direction intended by the programs, is a much smaller 1.6 percent. Unlike in the residential sector, both the energy intensity and energy consumption findings in the commercial and industrial sectors are consistent with the expected impacts of energy efficiency program commitment. In addition, their estimates of energy intensity and energy consumption impacts do not widely differ. Net program impacts reveal that even though energy use grew for the SM states, even greater growth would have occurred had their program commitment been weak. In the commercial sector, the net impact of program commitment was a reduction of 12.2 percent in electricity use from the base to the treatment period, and a reduction of 8.1 percent in energy intensity. In the industrial sector, the net impact of program commitment was a reduction of 9.6 percent in electricity use, and a reduction of 8.1 percent in energy intensity relative to what it might otherwise have been with weak program commitment. 4. Decomposition and Demand Transformation The findings of the impact analysis raise several issues that can be investigated further via decomposition analysis. Chief among these is the extent to which the net impacts are determined by differences in average state characteristics or differences in electricity demand behavior. Also, since behavioral responses to the X’s may be conditioned by program commitment, it is instructive to measure and compare the relative importance of each variable’s coefficient to the difference in the means of the dependent variables in each period. Such a comparison permits exploration into the phenomenon of demand transformation, a topic of particular interest to policymakers who, since the mid-1990s, have sought to develop programs and policies that have long-lasting, or transformative, effects on the markets for energy and energy-related equipment. While market transformation can mean different things depending on the context, in economic terms one way to define it is as a long-term change in electricity demand behavior. Different ways of addressing this subjects can be found in Duke and Kammen (1999) and Horowitz (2001). For this study the electricity intensity models estimated for the S quartile and the W quartile are used to decompose the differences in the means of the dependent variables. Note that the S quartile electricity intensity model does not include the observations for the two M quartiles. This is because the M states fall between program commitment extremes and dilute the contrast provided by the two ends of the program commitment spectrum. The analysis begins with the estimation of the electricity intensity equations presented in equation (6) and (7). For these equations it can be observed that the difference of the two group means is calculated as: –
–
–
–
EI W – EI S = (bW0 + ΣbWj XjW ) – (bS0 + ΣbSj XjS).
(10)
Changes in Electricity Demand in the United States / 105 Then, by subtraction and addition of terms, the definition can be rearranged such that: – – – – – EI W – EI S = ΣbWj (XjW – XjS) + Σ(bWj – bSj) XSj + (bW0 – bS0).
(11)
Expressing the difference of the means in this way, the first term on the right hand side of the equal sign is the portion of the average difference between the two groups that is due to differences in the means of the X’s, i.e., state characteristics, such as price and incomes. This is the explained portion of the difference in means. The second and third terms are the portion of the difference in means is that due to the model coefficients, i.e., to differential behavior between groups. Together, they make up the unexplained part of the difference in means. Of the two unexplained terms, the former represents that part of the difference that is due to differing behavioral responses to changes in the X’s, that is, differences in the slopes or bj’s, while the latter expression represents that part of the difference in means that is due to the model intercept which, by definition, represents common, undefined fixed factors. Interpretation of the two unexplained components tends to be problematic; it is well-known that this additional level of decomposition yields values that depend on arbitrary modeling choices such as the unit of measurement of the variables (Jones, 1983). However, since the modeling choices in this study are invariant from period to period, intra-sectoral comparisons have consistent meaning. Table 2 contains the decomposition statistics for each sector for the base period, and Table 3 contains the parallel statistics for the treatment period. Table 2. Decomposition Analysis: Base Period Models (S and W Quartiles) Sector Model Residential Commercial Industrial Variables constant
X
b
I
-1.536 28.065
X
b
I
-0.559 -3.217
X
b
I
5.619 13.126
Electricity Price
0.000 0.048 -0.871
-0.003 0.463 2.662
0.011 -0.675 -1.576
Natural Gas Price
-0.033 0.178 -3.251
0.020 -0.202 -1.162
-0.032 0.313 0.731
Per Capita Inc/GSP
-0.012 -0.711 12.989
-0.003 -0.166 -0.958
-0.004 -3.272 -7.642
Technology Trend
0.000 0.138 -2.523
0.000 0.304 1.748
0.000 -0.794 -1.854
Heating Degree Days 0.083 1.313 -23.990
-0.015 0.300 1.729
0.026 -1.242 -2.901
Cooling Degree Days 0.021 0.515 -9.419
0.003 0.034 0.198
0.077 0.478 1.116
Raw Difference
0.004
1.000
0.175
1.000
0.506
1.000
Explained
0.059
0.002
0.078
-0.055
0.174
0.428
Unexplained
106 / The Energy Journal Table 3. Decomposition Analysis: Treatment Period Models (S and W Quartiles) Sector Model Residential Commercial Industrial Variables constant
X
b
I
-0.918 13.853
X
b
I
3.999 14.466
X
b
I
-5.610 -10.618
Electricity Price
0.038 0.436 -6.574
0.025 -0.046 -0.166
-0.043 2.260 4.277
Natural Gas Price
-0.013 0.075 -1.138
-0.005 -0.059 -0.212
-0.010 0.179 0.338
Per Capita Inc/GSP
-0.011 -0.036 0.537
0.007 -1.992 -7.206
0.129 1.219 2.307
Technology Trend
0.000 0.032 -0.490
0.000 -0.039 -0.142
0.000 0.836 1.582
Heating Degree Days 0.092 0.349 -5.261
-0.026 -0.997 -3.608
0.003 1.604 3.037
Cooling Degree Days 0.007 -0.005 0.073
-0.004 -0.589 -2.132
-0.029 0.041 0.077
Raw Difference
0.048
1.000
0.274
Explained
0.114
-0.066
Unexplained
1.000
0.579
1.000
-0.002
0.276
0.050
0.528
The columns labeled X represent the weights given to the means of the X’s in explaining the difference of the dependent variable means, and the columns labeled b represent the weights given to the different estimated coefficients in explaining the difference of the means. As the models for the W states are always used as the starting point, the X’s are always weighted by the coefficients of W states, and the b’s are always weighted by the means of the S states. Hence, all things being equal, a positive statistic always favors higher electricity intensity for the W states, and a negative statistic always favors higher electricity intensity for the S states. By definition, the negative and positive values cancel each other so that when combined, the individual decomposition statistics sum to the difference of means between the groups. For the base period, the decomposition analysis shows that controlling for differential behavioral responses (the b’s), the means of the residential sector variables favors higher use for the W states; on the other hand, controlling for the mean level of the variables (the X’s), the behavioral responses favor higher use for S states. On the whole, these two factors are approximately equal, leaving the difference of the dependent variable means close to zero. The decomposition analysis for the treatment period is contained in Table 3. It reveals that in the treatment period the differences in the means of the X’s have much greater weight than the differences in the behavioral coefficients. In other words, the models of the residential sector S and W quartiles, though relatively similar in the base period, appear to have grown even closer in the treatment period. This findings hints at an alternative explanation for why the impact analysis results appear negative or negligible for the residential sector, namely, the presence of a high degree of nationwide spillover. This would occur if the influence of strong program commitment spread beyond state borders, and moreover, if the influence spread quickly
Changes in Electricity Demand in the United States / 107 and strongly. All in all, the findings suggest that investigation into nationwide spillover is warranted, about which more will be discussed below. Decomposition of the means of the commercial and industrial sector dependent variables yields different patterns from those found in the residential sector. For these sectors, almost all of the differences in means are unexplained, that is, attributed to differences in the models. As can be seen in Table 3, this pattern is even more pronounced in the treatment period. While in the residential sector the proportion of the difference of means accounted for by the models declines in the treatment period, in the commercial and industrial sectors the high proportion accounted for by the models grows slightly, and the absolute amount accounted for by the models grows moderately. In short, differential demand behavior is the main source of the differences in means in the commercial and industrial sectors, more so in the treatment period than in the base period. An investigation into demand transformation, or the extent to which program commitment has been responsible for long-term changes in the demand for electricity, is a natural extension of the decomposition analysis. Further information about the models that relate to demand transformation can be derived by comparing the changes in components of the models from the base to the treatment period. As an indicator of the importance of the individual behavioral responses, the column labeled I represents the relative importance of each b, calculated as the ratio of each individual b to the total unexplained difference. Changes in these statistics are direct indicators of changes in demand behavior. In the residential sector models, it appears from Tables 2 and 3 that the model constants are responsible for the largest share of the unexplained difference in the means of the dependent variables of the quartiles, followed by heating degree days and per capita income. This pattern holds across periods, except that electricity price in the treatment period is now second to the model constant in importance, and heating degree days is of far lesser importance. Only one coefficient increases its share of the unexplained difference from the base period to the treatment period; it is the one associated with electricity price. This uniformity does not carry over to the commercial and industrial sectors. For these sectors, additional changes appear in the relative importance of specific coefficients in the treatment period. In the commercial sector, four of the coefficients have greater relative importance in the treatment period, and in the industrial sector two coefficients are relatively more important in the treatment period. These statistics reveal that for the residential sector it is the differences in the state characteristics of the quartiles, as opposed to differences in electricity demand behavior, that drive differences in electricity demand, especially in the treatment period. This implies that the residential quartile models are similar, and that whatever transformation occurred in the S quartile also seems to have occurred in the W quartile. However, this does not appear to be the case for the commercial and industrial sectors. For these sectors, it is the differences in demand behavior, not state characteristics, that largely drive the differences in electricity demand. Moreover, the models move farther apart in the treatment period, imply-
108 / The Energy Journal ing that demand behavior is transforming at a faster rate in the S quartile than in the W quartile. To complete the analysis of demand transformation it is necessary to examine whether or not specific components of demand behavior are moving in the direction intended by energy efficiency programs. To do so, the direction of the changes in the model coefficients from the base to the treatment period are calculated, as shown in Table 4. Table 4. Changes in Coefficients from Base to Treatment Period Sector
Residential
Commercial Industrial
Coefficients
S
W
S
W
S
W
0.715 -0.283 -0.039 -0.310 -0.029 0.134 0.034
1.333 -0.162 -0.089 -0.083 -0.096 0.016 -0.043
-2.043 -0.140 0.068 -0.347 -0.047 0.134 0.033
2.514 -0.296 0.142 0.184 -0.152 -0.025 -0.059
5.793 -0.384 0.030 0.462 -0.151 -0.278 -0.033
-5.436 0.667 -0.034 -0.776 0.316 0.053 -0.102
constant Electricity Price Natural Gas Price Per Capita Inc/GSP Technology Trend Heating Degree Days Cooling Degree Days
In this table, the values in italics indicate that the change, based on the ttest, is statistically significant at the 95 percent level or better. Initial inspection of these statistics confirm the findings of the decomposition analysis, i.e., that in the residential sector the treatment period model for the S quartile hardly differs from the W quartile model, whereas in the commercial and industrial sectors, the treatment period models diverge. More specifically, in the residential sector all of the changes in coefficients for the S states and the W states are in the same direction, save cooling degree days. This is not true for the other sectors. In the commercial sector, only three of the seven changes are in the same direction; in the industrial sector, only one change is in the same direction. The direction of the change in the coefficients of three key economic variables, i.e., electricity price, per capita income or GSP, and technology trend, reveal whether the observed demand transformation favors lower electricity demand and thus meets the expectations of program commitment. With respect to energy prices, economic theory suggests that if programs intended to lower electricity demand are successful, energy price elasticity will increase as consumers become more sensitive to the possibility of substituting capital, in the form of more energy efficient equipment, for fuel. Of course, other pathways for increased price elasticity are possible, too, such as altered equipment operation. As can be seen in Table 4, demand has, in fact, become more price elastic from the base to the treatment period for both residential and both commercial groups, and for the strong program commitment states in the industrial sector. In addition, all of the changes, save for the price elasticity of the strong program commitment quartile in the commercial sector, are statistically significant.
Changes in Electricity Demand in the United States / 109 The finding of an income effect in the expected negative direction is an equally important indicator of demand transformation. Greater income inelasticity will occur if programs encourage consumers with rising incomes to substitute higher energy efficiency equipment for existing equipment, thereby turning energy use from a superior good into a normal one, all things being equal. Other pathways, such as sophisticated operations and maintenance, are possible too. As can be seen in Table 4, a change in the expected direction, as expressed through the per capita income or GSP coefficient, occurs for both residential groups, and for the S quartile in the commercial sector, and for the W quartile in the industrial sector. Further, these changes are statistically significant at the 95 percent level or better for the S quartile in the residential and commercial sectors. A third important economic indicator that program commitment led to demand transformation is the technology trend coefficients. This continuous variable varies by sectors and years, but not by state. It provides more precise insight into the relationship between the changes in the adoption of electronic technologies and electricity demand than would be possible through either a simple annual time trend or a set of annual dummy variables. Once again, it is to be expected that successful programs will reduce the elasticity of demand with respect to technological change. In other words, it is expected that energy efficiency programs will help slow the increase in electricity demand associated with the increasing production and use of electronic equipment. As can be seen in Table 4, demand becomes more inelastic from the base to the treatment period in all but the industrial sector W states, and the changes are statistically significant in all but the S quartile of the commercial sector. To summarize, with few exceptions, the findings related to the changes in the coefficients of three key economic variables in all three sectors suggest that demand transformation due to energy efficiency program commitment has occurred. However, in the residential sector the decomposition and demand transformation analyses suggest that the changes in demand in the weak program commitment states closely parallel those of the strong program commitment states. This lends credence to the idea that there is a high degree of spillover in the residential sector, and that spillover is perhaps taking place much more strongly, and much faster, than in the other two sectors. 5. Nationwide Spillover Spillover, as applied to energy efficiency programs, is a positive externality that increases the cost-effectiveness of public programs by creating unanticipated social benefits. Under a narrow definition, spillover occurs when energy efficiency program participants go beyond what is expected of them, for example, by purchasing an eligible energy efficient item without applying for a rebate after they have already purchased a similar item with the assistance of a rebate. A broader definition of spillover is that it occurs when, for example, individuals who have never participated in a program purchase an energy efficiency item because
110 / The Energy Journal they have been influenced by the existence of a program. There are many pathways through which this may occur, say through a program’s explicit endorsement of a line of products, through its diffusion of product information, or even through its influence on producer and supply chain strategies. Efforts to estimate the benefits of spillover typically involve local area surveys of program participants and non-participants. While no empirical studies of spillover have been published in journals, there are many professional reports containing anecdotal or survey information on these topics, summaries of many of which can be found in conference proceeding such as ACEEE (2006) and IEPEC (2005). However, because most energy efficiency programs are geographically circumscribed, there are no reports or studies that have attempted to measure the extent to which areas of the country with strong or moderate energy efficiency program commitment have encouraged energy efficiency in areas of the country with weak program commitment. Nationwide spillover is particularly relevant to the residential sector because news and information flow almost effortlessly via mass media and mass marketing from one socio-economic or demographic group to another, and from one geographic area to another. Adding to the rapidly-spread influence of lifestyle and taste trends, consumers are able to execute decisions regarding equipment investments and equipment operations far faster than are most business organizations. Also, in the residential sector it is often the case that there are but a few major manufacturers of equipment, or a few major retailers with large market shares, and their inventories tend to be relatively standardized throughout the country. At the other extreme, spillover is probably least likely to occur in the industrial sector because these businesses have the most heterogeneous and complex needs and uses for energy, and require the most capital. The commercial sector is probably somewhere in between the two with regards to spillover; building stock and building functions vary, yet end uses and equipment suppliers tend to be similar. Also, equipment tends to be more costly and longer lived than in the residential sector, yet commercial buildings must continually modernize if they are to keep tenants. In light of the possibility of rapid residential spillover, the findings of the residential sector impact and decomposition analyses might be interpreted as supporting the hypothesis that nationwide spillover is ubiquitous in the residential sector, and that spillover has led to the W quartile’s demand behavior being similar in that of the S states, even in the base period. This would help explain why positive impacts cannot be detected for the SM states, and why the decomposition analysis finds that the energy intensity models are similar across quartiles and move towards convergence in the treatment period. To explore this phenomenon, a new base period is constructed for the residential sector, one that goes as far back as 1970 and ends in 1979. In these years, residential programs were rare even in the S states. Re-analysis of this sector using the earlier years makes particular sense because it is likely that for this sector, more than the two others, the 1977 to 1992 base period was contaminated by strong program commitment in some states. Electricity efficiency programs began largely
Changes in Electricity Demand in the United States / 111 in the mid-1970’s with information and audit programs for the residential sector. In time, efficiency programs spread to the commercial and industrial sectors. However, residential programs began earlier, were larger in scope, and spread faster from state to state than any other programs. By 1992, according to Form EIA-861, 18 states recorded residential sector DSM savings of more than 100,000 MWh. Also, by 1992 DSM savings were present in all states in the residential sector. Descriptions of many of these residential programs can be found as far back as Berman, et al. (1987). Therefore, most especially for the residential sector, shifting backwards and shortening the base period can produce a period with less confounding effects and greater contrast to the treatment period than the original base period. Shifting the base period backwards, Table 5 restates all of the residential findings from Tables 1, 2, 3, and 4, using the 10 years from 1970 to 1979 as the new base period, referred to as base70. As above, descriptive statistics for the variables in the new models are contained in Table A1 and the models themselves are contained in the rightmost columns of Table A2 in Appendix A. The findings of the re-simulation, found in the top portion of Table 5, indicate that the residential energy consumption impacts are far less negative than when using the original base, an increase in electricity consumption of 2.3 percent rather 9.1 percent. And, they reverse for energy intensity, showing a decline in energy intensity of 4.4 percent, rather than an increase in energy intensity of 1.6 percent. Since the treatment period’s counterfactual for the SM quartile remains the same as in the original impact analysis, the new residential sector impact findings derive from the differences in the the base70 period counterfactuals. This means that the base and base70 period W quartile models differ in important details, detection of which is possible by contrasting the original residential sector Table 5. Re-analysis of Residential Sector Impact Analysis Level of Electricity (GWh) Use Electricity Intensity (SM Quartiles)
Base70
Factual Counterfactual 95% C.L. (+/-) Net Impact
496,863 486,993 18.3%
Decomposition (S and W Models)
Treatment % Change 949,968 919,711 8.2%
91.2% 88.9% 2.3%
Base70
Base70 Treatment % Change 9.9 9.3 8.9%
14.5 13.9 3.1%
45.9% 50.3% -4.4%
% Change (Base70 to Treatment)
X
b
I
S
W
constant Electricity Price Natural Gas Price Per Capita Inc. Technology Trend Heating Degree Days Cooling Degree Days
-0.012 -0.012 -0.008 0.000 0.164 0.005
-4.018 -0.913 0.145 0.741 -0.055 3.486 0.417
20.416 4.637 -0.739 -3.765 0.279 -17.712 -2.117
0.030 -0.601 0.074 -0.081 -0.258 0.208 0.112
3.130 -0.175 0.023 -0.360 -0.198 -0.174 0.050
Raw Difference Explained Unexplained
-0.060 0.137
-0.197
1.000
112 / The Energy Journal decomposition and demand transformation analysis with the re-analysis using the base70 models. The findings of the re-analysis show that unlike in the base period models, the unexplained portion of the difference of the means of the S and W quartiles is larger, not smaller, than the explained portion of the difference. In other words, differences in the coefficients of the S and W quartile models are more pronounced in the base70 period. Further, three base70 coefficients, rather than just one in the original analysis, are relatively more important in the treatment period than they are in the earlier period. These are indicated by the bold values in the I column for electricity price, natural gas price, and technology trend. Finally, the analysis of the changes in S and W quartile model shows that the components of both the S and W quartile’s demand behavior follow the same directions of change. This is consistent with the original findings and supports the impression that demand transformation is occurring in the W quartile despite weak program commitment. Moreover, the three key economic variables, i.e., electricity price, per capita income, and technological change moved in directions that are consistent with the expected impacts of program commitment for both the S and W quartiles, and these movements are statistically significant in all but one case. 6. How Goes California To come full circle, it is instructive to return to the introductory discussion to see whether California’s accomplishments, as suggested by simple trend lines in Graph 3.A to 3.C, are supported by a more detailed analysis. As can be seen in these graphs, once economic variables are introduced, discussion of the impacts of energy efficiency program commitment becomes more complex. Since 1980, per capita income or GSP rose while, on average, real electricity prices were flat, albeit volatile. The patterns of these economic variables suggest that California’s electricity intensity trends for the past three decades should not have been so favorable. Yet, without controlling for these and other factors for other states, it is not obvious that California’s trends are as extraordinary as they seem. In other words, without a more detailed analysis, conclusions related to the relative impacts of California’s program commitments are speculative. Graph 3.A – 3.C. California Economic Trends, by Sector
Changes in Electricity Demand in the United States / 113 To gain a more complete understanding of California’s accomplishments, the impact analysis findings are estimated for California alone, as seen in Table 6. The findings for California are dramatic; net impacts in the residential sector show a reduction in electricity use of 42.8 percent and a reduction in electricity intensity of 27.7 percent. In the commercial sector, electricity consumption shows a reduction of 28 percent and a reduction in electricity intensity of 19.3 percent. In the industrial sector, electricity consumption shows a reduction of 46 percent and a reduction in electricity intensity of 32.2 percent. To assess the validity of the counterfactuals for California, descriptive statistics for the state are contained in Table A4 of Appendix A. In comparing them to the base period descriptive statistics for the W quartiles in Table A1, it can be seen that some of the raw means differ for several of the independent variables in the model. However, other than for residential sector per capita income and industrial sector electricity price, none of the means of California’s variables fall outside two standard deviations of the means of the W quartile variables in the base or base70 period. This suggests that the California counterfactuals, while more uncertain than the estimates for the full SM quartile, are nevertheless valid estimates of what California electricity demand would have been in the absence of strong program commitment. Table 6. Evaluation of California Program Commitment California, only Level of Electricity (GWh) Use Electricity Intensity Sector
Base
Treatment % Change
Base
Treatment % Change
Residential (base70) Factual (CA) 44,049 Counterfactual (CA) 50,665 95% C.L. (+/-) 22.37% Net Impact
73,474 106,183 8.21%
Factual (CA) 72,142 Counterfactual (CA) 103,391 95% C.L. (+/-) 21.65% Net Impact
95,474 165,757 32.69%
66.80% 109.58% -42.77%
7.0 8.0 10.2%
7.6 11.0 3.4%
9.0% 36.7% -27.7%
464 645 3.5%
367 634 5.3%
-20.9% -1.7% -19.3%
1,328 2,512 2.8%
1,214 3,108 4.0%
-8.5% 23.7% -32.2%
Commercial 32.34% 60.32% -27.98%
Industrial Factual (CA) 52,547 Counterfactual (CA) 99,777 95% C.L. (+/-) 22.45% Net Impact
57,319 154,752 34.38%
9.08% 55.10% -46.02%
This analysis of California electricity demand is illuminating for at least two reasons. For one, it confirms that the speculations based on the popular graphs of California electricity demand trends are not unjustified. But perhaps more importantly, since California is the state with the longest history and the strongest
114 / The Energy Journal degree of energy efficiency program commitment of any state, the analysis uncovers what is likely to be, to date, the upper bounds of the impacts of strong statelevel energy efficiency program commitment. 7. CONCLUSION This study presents rigorous empirical estimates of the changes in electricity demand in the United States from the 1970s to the beginning of the 21st century. It introduces a counterfactual difference in differences estimator for measuring energy efficiency program commitment impacts that is designed to overcome the shortcomings of more conventional difference in differences estimators. Using this approach, the evidence suggests that, on average, those states that have moderate to strong commitment to energy efficiency programs reduce energy intensity in their states relative to what it would have been with weak program commitment. In addition, these findings hold for changes in the levels of energy consumption in the commercial and industrial sectors, if not the residential sector. As such, the results appear to be robust and are likely, if anything, to understate net impacts because of the inherent imprecision in state rankings and the uniform cutoff year. Ironically, the findings of this study simultaneously confirm and refute Jevons’ paradox. On the one hand, the evidence supports the paradox by revealing that energy efficiency program commitment in the residential sector is associated with a decrease in energy intensity alongside an increase in energy consumption. However, in the commercial and industrial sectors energy efficiency program commitment is associated with decreases in both energy intensity and energy consumption. Together, this collection of empirical findings suggests that Jevons’ paradox is no paradox at all, but rather a description of one of many possible outcomes of increases in energy efficiency. On a more practical note, the results of this paper have potentially important implications for how energy policymakers should view energy efficiency programs. The evidence presented in this paper largely indicates that energy efficiency program commitment in all three sectors of the economy has transformed electricity demand with respect to three key economic variables; electricity price, income as measured by per capita income or gross state product, and technological change. Furthermore, a comparison of the statistical findings from two different base periods suggests that nationwide spillover from energy efficiency programs may be rapid and ubiquitous in the residential sector. Awareness of these longterm impacts should encourage policymakers to take into serious consideration the benefits of different levels of program commitment, where estimates of the impact of California’s energy efficiency programs likely reveal the upper bounds of these benefits to date, given California’s particular history and the size and mix of its energy efficiency program portfolio. Finally, the methodology used in this study should have broad interest both within the U.S. and internationally. Until now, most effort to measure changes in electricity demand have ignored the impacts of energy efficiency program
Changes in Electricity Demand in the United States / 115 commitment, or have focused on limited geographic areas, or have not simultaneously controlled for the key determinants of energy demand. Moreover, they have not been able to address, or measure, demand transformation and nationwide spillover. The analyses in this study show that wide-ranging empirical studies of these subjects are not only feasible, but create an opportunity to explore difficult and important issues that have heretofore been unapproachable. REFERENCES American Council for an Energy Efficient Economy (ACEEE) (2006). Proceedings of the Summer Study on Energy Efficiency in Buildings. Washington, D.C. Becker, Gary S. and Kevin M. Murphy (1993). “A Simple Theory of Advertising as a Good or Bad.” Quarterly Journal of Economics, 108(4): 941-964. Berman, Ellen, Marc Cooper, and Howard Geller (1987). A Compendium of Utility-Sponsored Energy-Efficient Rebate Programs. American Council for Energy Efficient Economy, Washington, DC: Report Number U871. Bernstein, Mark, K. Fonkych, S. Loeb, and D. Loughran (2003). State-Level Changes in Energy Intensity and Their National Implications, Rand, MR-1616-DOE, Santa Monica, CA. Blinder, Alan S. (1973). “Wage Discrimination: Reduced Form and Structural Estimates.” Journal of Human Resources, 8: 436-455 Bosseboeuf, Didier, B. Lapillonne, and W. Eichhammer (2005). “Measuring Energy Efficiency Progress in the EU: The Energy Efficiency Index ODEX,” European Council for an Energy Efficient Economy, Proceedings of the 2005 ECEEE Summer Study, 1127-1135, Stockholm, Sweden. California Public Utilities Commission and California Energy Commission (2006). Energy Efficiency, California’s Highest-Priority Resource. Dahl, Carol (1993). A Survey of Energy Demand Elasticities in Support of the Development of the NEMS. Washington, DC: U.S. Department of Energy. Duke, R. and M. Kammen (1999). “ The Economics of Market Transformation Programs.” Energy Journal, 20(4): 15-64. Gillingham, Kenneth, Richard Newell, and Karen Palmer (2004). Retrospective Examination of Demand-Side Management Energy Efficiency Policies. Resources for the Future, Discussion Paper 0419, Washington D.C. Global Energy Partners (2003). California Summary Study of 2001 Energy Efficiency Programs, Final Report. Lafayette, CA. Herring, Horace (2006). “Confronting Jevons’ Paradox: Does Promoting Energy Efficiency Save Energy.” International Association for Energy Economics Newsletter, Volume 15, 4th quarter, 14-15. Horowitz, Marvin J. (2001). “ Economic Indicators of Market Transformation: Energy Efficient Lighting and EPA’s Green Lights.” Energy Journal, 22(4): 95-122. Horowitz, Marvin J. (2004). “ Electricity Intensity in the Commercial Sector: Market and Public Program Effects.” Energy Journal, 25(2): 115-137. International Energy Efficiency Program Evaluation Conference (IEPEC) (2005). Conference Proceedings. Madison, WI. Johnson, Justin P. and David P. Myatt, (2006). “On the Simple Economics of Advertising, Marketing, and Product Design.” American Economics Review, 96(3): 757-789. Jones, F.L. (1983). “On Decomposing the Wage Gap: A Critical Comment on Blinder’s Method.” Journal of Human Resources, 18: 126-130 Kushler, Martin, and Patti Witte (2001). A Revised 50-State Status Report on Electric Restructuring and Public Benefits. American Council for an Energy Efficient Economy, Washington, DC: Report Number U005. Loughran, David S. and Jonathan Kulick (2004). “Demand Side Management and Energy Efficiency in the United States.” The Energy Journal, 25(1): 19-43.
116 / The Energy Journal Meyer, Bruce D. (1995). “Natural and Quasi-Experiments in Economics.” JBES, 13(2): 151-161. Rosenfeld, Arthur H. (2005). “Extreme Efficiency: Lessons from California.” California Energy Commission. Sacramento, CA. World Energy Council (2001). Energy Efficiency Policies and Indicators. London, United Kingdom. York, Dan and Martin Kushler, (2006). A Nationwide Assessment of Utility Sector Energy Efficiency Spending, Savings and Integration of Utility System Resource Acquisition. American Council for an Energy Efficient Economy, Proceedings of the Summer Study on Energy Efficiency in Buildings. Washington, D.C.
APPENDIX A Table A1. Descriptive Statistics, by Quartiles Sector
Residential
Variables
Mean
Electricity Intensity Electricity Price Nat. Gas Price Per Cap Inc/GSP Tech. Trend HDD CDD
12.12 29.25 7.82 19.61 7.20 5176 1070
Electricity Intensity Electricity Price Nat. Gas Price Per Cap Inc/GSP Tech. Trend HDD CDD
14.46 25.89 7.69 26.15 112.14 5062 1076
Electricity Intensity Electricity Price Nat. Gas Price Per Cap Inc/GSP Tech. Trend HDD CDD
12.68 28.04 6.49 19.18 7.20 5722 1121
Electricity Intensity Electricity Price Nat. Gas Price Per Cap Inc/GSP Tech. Trend HDD CDD
15.35 22.93 6.55 25.12 112.14 5621 1099
Std. Dev.
Commercial Industrial Mean
Std. Dev.
Mean
Std. Dev.
2928 19.70 5.50 0.02 17.88 5811 871
2407 6.16 1.56 0.005 8.04 1986 700
169 6.20 1.35 0.006 45.41 2009 788
2898 15.77 4.50 0.031 131.10 5700 873
2492 5.38 1.26 0.005 80.43 1996 720
125 3.43 1.32 0.005 11.39 1728 607
3075 18.89 4.32 0.022 17.88 3818 1719
1001 3.29 1.20 0.004 8.05 1323 603
3273 13.19 4.05 0.028 131.10 3708 1708
1063 2.12 1.16 0.007 80.62 1318 664
SM - BASE Period (n=540) 3.57 7.00 1.86 3.60 3.83 2096 848
627 27.69 6.83 0.02 26.26 5681 918
201 7.17 1.61 0.00 11.37 1996 748
- TREATMENT Period (n=432) 4.13 7.00 1.79 4.62 65.83 2087 868
579 22.51 6.23 0.031 108.02 5565 931
W - BASE Period (n=180) 3.19 5.75 1.36 2.89 3.83 1779 463
739 25.42 5.63 0.023 26.26 4209 1577
W - TREATMENT Period (n=144) 3.29 3.88 1.26 3.38 65.98 1839 508
743 19.8 5.81 0.028 108.02 4112 1535
140 2.64 1.33 0.004 45.52 1726 617
Changes in Electricity Demand in the United States / 117 Table A1. Descriptive Statistics, by Quartiles (continued) S - BASE Period (n=180) Electricity Intensity Electricity Price Nat. Gas Price Per Cap Inc/GSP Tech. Trend HDD CDD
13.01 28.72 8.06 20.23 7.20 4253 1243
Electricity Intensity Electricity Price Nat. Gas Price Per Cap Inc/GSP Tech. Trend HDD CDD
15.13 25.94 8.13 26.99 112.14 4133 1258
4.39 7.99 1.56 3.28 3.83 2050 998
566 29.36 7.13 0.026 26.26 5049 995
170 8.77 1.6 0.004 11.39 2215 973
2037 20.80 5.64 0.025 17.88 5859 668
1230 7.49 1.50 0.004 8.05 1648 353
1983 16.62 4.52 0.034 131.10 5733 657
1050 6.02 1.18 0.005 80.62 1636 335
S - TREATMENT Period (n=144) 4.73 7.35 1.84 3.76 65.98 2020 1012
498 24.11 6.35 0.035 108.02 4922 1022
141 6.96 1.35 0.005 45.52 2199 998
RESIDENTIAL BASE70 Period Electricity Intensity Electricity Price Nat. Gas Price Per Capita Inc. Tech. Trend HDD CDD
SM (n=360) 9.91 25.85 5.26 15.59 2.89 5256 1024
W (n=120)
3.13 6.60 1.70 2.34 0.56 2136 812
10.28 24.98 4.08 15.87 2.89 5845 1064
3.39 5.18 1.20 2.55 0.56 1855 436
S (n=120) 11.14 24.44 5.46 16.08 2.89 4327 1191
4.07 7.98 1.63 2.25 0.56 2087 963
Table A2. Electricity Intensity Models, Base Period Base Period: 1977 to 1991 Sector Variables
Residential S W
Base70 Period
Commercial Industrial S W S W
Residential S W
constant s.e. Electricity Price s.e. Natural Gas Price s.e. Per Capita Inc/GSP s.e. Technology Trend s.e. Heating Degree Days s.e. Cooling Degree Days s.e.
0.986 -0.550 0.248 0.554 -0.237 -0.222 0.034 0.047 0.065 0.151 0.019 0.027 0.470 0.233 0.043 0.118 0.036 0.112 0.012 0.021 0.040 0.201 0.018 0.043 0.053 0.129 0.016 0.022
4.667 0.343 -0.065 0.030 -0.002 0.021 -0.590 0.036 0.074 0.010 -0.074 0.032 0.011 0.014
4.108 0.708 0.077 0.074 -0.108 0.031 -0.545 0.076 0.170 0.013 -0.037 0.054 0.016 0.040
3.878 9.498 0.652 0.863 -0.204 -0.433 0.029 0.058 -0.066 0.119 0.024 0.032 -0.847 0.033 0.070 0.083 0.157 -0.131 0.017 0.015 0.085 -0.059 0.063 0.074 -0.002 0.074 0.023 0.048
1.671 0.400 0.081 0.045 -0.048 0.026 0.242 0.107 0.266 0.041 -0.034 0.023 -0.026 0.026
-2.347 1.004 -0.210 0.041 0.040 0.036 0.510 0.147 0.213 0.060 0.391 0.093 0.036 0.038
D.V. Mean n adj. R2
2.502 180 0.91
6.451 180 0.88
6.462 180 0.54
7.469 180 0.75
2.341 120 0.96
2.281 120 0.99
2.506 180 0.79
7.976 180 0.72
118 / The Energy Journal Table A3. Electricity Intensity Models, Treatment Period Treatment Period: 1992 to 2003 Sector Variables
Residential S W
Commercial Industrial S W S W
constant s.e. Electricity Price s.e. Natural Gas Price s.e. Per Capita Inc/GSP s.e. Technology Trend s.e. Heating Degree Days s.e. Cooling Degree Days s.e.
1.701 0.489 -0.519 0.055 0.026 0.015 0.161 0.079 0.008 0.008 0.174 0.029 0.087 0.015
0.783 0.440 -0.384 0.046 0.063 0.017 0.150 0.075 0.015 0.008 0.217 0.032 0.086 0.018
2.623 0.794 -0.204 0.071 0.066 0.023 -0.937 0.156 0.027 0.030 0.060 0.033 0.044 0.018
6.622 1.318 -0.219 0.110 0.034 0.034 -0.361 0.170 0.018 0.037 -0.062 0.096 -0.043 0.048
9.672 1.214 -0.588 0.097 -0.037 0.028 -0.385 0.297 0.005 0.037 -0.192 0.077 -0.035 0.028
4.062 1.418 0.234 0.136 0.084 0.042 -0.743 0.271 0.186 0.034 -0.006 0.090 -0.028 0.062
D.V. Mean n adj. R2
2.658 144 0.91
2.705 144 0.92
6.339 144 0.77
6.426 144 0.39
7.461 144 0.54
8.040 144 0.69
Table A4. Descriptive Statistics for California Base (Base70 for Res.)
CA-Res (n=10) Mean Std. Dev.
CA-Com (n=15) Mean Std. Dev.
CA-Ind (n=15) Mean Std. Dev.
Electricity Intensity Electricity Price Nat. Gas Price Per Cap Inc/GSP Tech. Trend HDD CDD
6.98 26.39 3.86 19.14 2.89 2737 805
464 31.92 6.98 0.028 26.26 2508 883
1328 27.78 5.66 0.028 17.88 2508 883
Treatment
CA-Res (n=12) Mean Std. Dev.
CA-Com (n=12) Mean Std. Dev.
CA-Ind (n=12) Mean Std. Dev.
Electricity Intensity Electricity Price Nat. Gas Price Per Cap Inc/GSP Tech. Trend HDD CDD
7.60 35.18 7.37 28.81 112.14 2451 943
367 31.96 6.86 0.034 108.02 2451 943
1178 23.53 4.44 0.034 131.10 2451 943
0.44 2.31 0.46 1.42 0.58 219 76
0.17 2.26 1.07 2.70 68.67 209 108
51 2.35 1.65 0.003 11.75 167 113
21 2.98 0.96 0.003 47.37 209 108
159 3.46 1.31 0.003 8.31 167 113
128 3.06 1.16 0.003 83.91 209 108
Changes in Electricity Demand in the United States / 119 Table A5. Program Commitment Quartiles (States in alphabetical order) Quartile 1 Quartile 2 Quartile 3 Quartile 4 (S States) (M states) (M states) (W States) RES COM IND
RES COM IND
RES COM IND
AL
CA
CA
AZ
AZ
CO
IN
DE
CA
CT
CT
CO
CO
FL
KY
FL
FL
IA
CT
IA
ID
LA
GA MA
IN
IA
IN
ME
MA MD
MA
ID
ME
NY
RES COM IND
GA
AR
AL
AL
ID
IL
DE
AR AR
IL
MD
IL
GA AZ
ME
KY
MO
KS
KS
DE
MI
ND
MT
MO
LA
KS
NC MN
MI
MD
MI
OH
MS
NH
ND
ND
MO KY
NY
NJ
MN
MN
MT
PA
NH
PA
NH
NE
MS LA
OH NY
NC
MT
NC
RI
OK
SD
NV
NM
NE MS
OR OR
NJ
NJ
NV
SC
PA
UT
SD
NV
NM NE
TX
TX
OR
SC
OH
UT
RI
VA
TX
SD
OK NM
WA WA
WA
UT
RI
VA
VT
VT
VT
TN
TN OK
WI
WI
VA
SC
WY
WV WV WV
WI
WY WY TN
120 / The Energy Journal
