Você está na página 1de 14

Energy Efficiency

DOI 10.1007/s12053-010-9082-6

Measuring the savings from energy efficiency policies:


a step beyond program evaluation
Marvin J. Horowitz

Received: 17 March 2009 / Accepted: 19 March 2010


# Springer Science+Business Media B.V. 2010

Abstract Recent worldwide events are producing other energy resource. Wind, solar, coal, water, oil,
growing agreement that rapid, large-scale energy effi- atoms, corn, or what have you, are physical objects,
ciency is a necessary, feasible, and economical way of as is the energy that comes from their conversion.
reducing the need for new energy supplies. Yet, many However, energy efficiency is not a physical object;
obstacles must be overcome for commitments to energy for any desired level of light, heat, or other service
efficiency to be successful, perhaps none more impor- that requires energy as an input, improved energy
tant than measuring policy-related energy savings. To efficiency means that less energy and fewer
address this topic, this paper provides several examples resources are demanded. This creates a dilemma.
of how readily available MWH sales data can be Simple counting by arithmetic cannot measure the
employed to monitor and verify energy efficiency volume of fuel and energy that would have been
policies and detect changes in trends at the state and demanded but was not. Measurement can only be
utility levels. The analyses used for this kind of savings done indirectly by statistical inference or what is
verification are collectively referred to in this paper sometimes called counterfactual reasoning. Like every
as counterfactual simulation. The goal of this paper is “what if…” question, the answer to how much energy
to demonstrate the practicality and value in using an energy efficiency measure saves requires a leap of
standardized research designs and econometric models reasoning.
to inform energy efficiency policymakers and the Uncertainty notwithstanding, recent worldwide
public as to the benefits of energy efficiency policies. events are producing growing agreement that rapid,
large-scale energy efficiency is a necessary, feasi-
Keywords Energy savings measurement . ble, and economical way of reducing the need for
Energy efficiency policy . Counterfactual simulation . new energy supplies. With greater enthusiasm than
Econometric models ever before, energy efficiency is viewed by many
as among the first options that ought to be chosen
for attempting to lessen greenhouse gas emissions
Introduction and mitigate global climate change. This is clearly
displayed by rising popular interest in national
It has long been recognized that it is difficult to energy efficiency policies and international climate
measure the energy savings that come from energy agreements. What is more, numerous efforts are
efficiency measures. Energy efficiency is unlike any underway via proposed state/regional power and
emissions pacts to make energy efficiency a tradable
commodity equal to physical supplies of fuel or
M. J. Horowitz (*)
electricity.
Demand Research LLC,
Fairfax, VA, USA For energy efficiency to succeed as a major energy
e-mail: mhorowitz@demandresearch.net resource many obstacles must be overcome, perhaps
Energy Efficiency

none more difficult than producing credible measure- Counterfactual simulation: a step beyond
ments of energy savings. A market for energy
efficiency is no different than a market for any Despite decades of experience with impact evaluation
commodity or financial instrument. Mutually agreeing methods that mimic the laboratory research techni-
to reductions in energy use or greenhouse gases, or ques found in the physical sciences (e.g., pre and post
swapping emissions with a cap and trade system, can period monitoring, treatment and control group
only work well when all the engaged parties are comparisons) this subject now requires a fresh look.
prudent and open about the amounts, terms, and risks Widely acknowledged limitations in existing impact
involved. Energy savings contracts cannot be priced evaluations include two fundamentally unrealistic
efficiently, and should not be bought or sold, without assumptions; first, that most programs can be reliably
such information. The same holds true for voluntary evaluated as stand-alone efforts even though many
national and global climate change agreements; pacts programs overlap each other, and second, that
cannot be meaningful unless trusted procedures are in consumer preferences and behaviors are fixed and
place for estimating and verifying the quantity of unresponsive both to market conditions and changes
energy efficiency resources that flow from large-scale in public policies. A third limitation is the inability of
public efforts. these evaluations to produce “net” savings using
This paper contributes to the advancement of revealed, rather than stated intentions, and a fourth
energy efficiency program evaluation by providing is the lack of standardized technical definitions,
several examples of how evaluation focused on research designs, models, and reporting conventions.
aggregate impacts rather than individual program This latter issue is of paramount importance to energy
impacts can overcome several of the most significant system planners and policymakers wanting to com-
challenges facing energy savings measurement. These pare the value of different energy efficiency program
are the challenges of dealing with program overlap, resources across sectors, years, and/or locations.
changes in consumer behavior, net savings, and lack Initiatives in the United States and Europe address-
of rigorous evaluation standardization and reporting. ing these shortcomings include the European Com-
To avoid confusion, in the context of this paper an mission’s project on evaluation and monitoring for the
aggregation of programs is referred to as an energy end-use directive on energy end-use and energy
efficiency policy. The term “portfolio” is eschewed services (EMEEES 2008), the US Department of
because it connotes a collection of programs imple- Energy’s Office of Energy Efficiency and Renewable
mented by a single organization. The term energy Energy publication “International Performance Mea-
efficiency policy, on the other hand, is meant to surement and Verification Protocol” (IPMVP 2007),
encompass the sum of energy efficiency measures, the National Action Plan for Energy Efficiency
projects, programs, codes and standards, and other Leadership Group’s, “The Model Energy Efficiency
efforts implemented in a defined geographic territory Program Impact Evaluation Guide,” (NAPE 2007),
by all federal and local public agencies, regulated and the Northeast Energy Efficiency Partnership’s
utilities, and third parties. Evaluation and Monitoring Forum, the mission of
The next section provides brief background infor- which is described in “Regional EM&V Forum’s
mation on impact evaluation and the reasons for Three Year Plan” (NEEP 2008).
employing counterfactual simulations for estimating Unfortunately, efforts to improve evaluation have
aggregate savings. This is followed by four related thus far focused only on enhancing well-known
analyses demonstrating the adaptability of the coun- microdata-based methods, such as those described in
terfactual approach to different sampling frames, the IPMVP handbook. They fail to embrace the
different data availability, and different historical possibility that with the amount of funding for public
conditions. As a proof of concept intended as the energy efficiency policies now being negotiated at the
first step towards large-scale deployment of this state, regional, and national levels, permanent energy
evaluation approach, the analyses in this study focus savings in aggregate could very well be much larger
on residential MWH sales in many different states and than that which appears from the summation of
exploit empirical data and tested models rather than savings derived from individual program evaluation
fictional examples. findings. Underestimation using individual program
Energy Efficiency

findings is likely to occur for several reasons. For one, effects of rebound, which reduces energy savings, and
individual program evaluation research designs are spillover, which increases the energy savings that can
primarily intended to measure immediate, short-term be attributed to public programs. Fundamentally, the
savings. In other words, they are intentionally myopic. approach relies on the existence of variations in
For another, these research designs focus narrowly on governmental policies, and market factors, across
program participants, not on society in general, which time and cross-sections to explain changes in energy
means that they produce tunnel vision. As a result, demand trends and for forecasting what would have
though they achieve some useful goals, such as aiding in occurred in the absence of public policy interventions.
year-to-year program design and management and Numerous well-known studies use time series or
resolving shareholder cost recovery and incentives panel models to study energy demand, dating at least as
claims, they may inadvertently result in the undervalu- far back as Fisher and Kaysen (1962), Halvorsen
ation of the collective, long-term benefits of concerted, (1975), and Taylor et al. (1984). Recent studies include
multi-faceted public policy efforts. those of Bernstein et al. (2003), Loughran and Kulick
At present, the sum of individual energy programs in (2004), Horowitz (2004), and Metcalf (2008). Of
many jurisdictions and countries is expected to yield particular relevance for the present study is Horowitz
energy savings at rates of between 1% and 2% of energy (2007) which analyzes changes in national electricity
sales per year for the next decade. This is a far greater demand and estimates the impacts of electricity energy
amount of savings than ever before planned, with efficiency policies in the commercial, industrial, and
cumulative impacts substantial enough to help forestall residential sectors from 1992 to 2003 using the 48
the building of new electric generation and transmission mainland states. In this study, policy-related energy
facilities and to slow down the rate of growth in savings were estimated after controlling for energy
greenhouse gas emissions. Given the large expectations, prices, incomes, changes in equipment stock, and
new methods need to be cultivated for measuring energy weather. By the character of the research design and
savings. One method involves longitudinal statistical the econometric models, the analysis netted out the
analyses of the energy sales to entire populations of effects of market factors that both add to, and subtract
consumers, not just short-term analyses for selected from, policy-related energy savings. The end result
consumer segments. If, after controlling for market and was a comprehensive estimate of the long-term net
other factors, public programs are found to meet or impacts of energy efficiency policy.
exceed their long-term goals, reassurance would be The four examples provided in this study follow in
provided that energy efficiency programs produce this vein and are intended to demonstrate the
reliable, permanent results. If they are found to fall applicability of the econometric approach at various
short, this too would be of importance to know and levels of aggregation for which time series or panel
understand. data are available. Generally speaking, for most
Time series and pooled cross-section time series or countries there exists at least three geographic levels
“panel” econometric models are not only able to of publicly available data that are capable of being
overcome the short-sightedness and narrowness of used for energy efficiency policy evaluation. For
conventional program impact evaluations, but are also example, in the USA various kinds of public data are
able to eliminate the need for supplemental studies available at the national level, the state level, and
that help transform program energy savings to net utility service territory level, and even the county,
impacts from gross impacts. Typically, a net-to-gross municipal, and zip-code levels. Time wise, various
factor for energy efficiency programs is composed of kind of public data can be found at the decennial
a single estimate of free ridership derived via level, the annual level, the quarterly level, and the
consumer surveys of stated intentions. Econometric monthly level. Sector-wise, most of the data in the
models cancel the need for these surveys because free USA can be divided into that which relates to
ridership is a market effect driven by such variables as the residential, commercial, or industrial sector.
energy prices, consumer incomes, and capital stock For clarity and focus, all four of these analyses take
trends, and properly-specified econometric models electricity consumption in the residential sector as
control for these effect. As an added bonus, by their their subjects and all four research designs result in
very nature econometric models also incorporate the comparisons of actual energy sales data to simulated
Energy Efficiency

or forecasted energy sales. Also, their econometric represented by (G); climatic conditions as represented
models use similar explanatory variables, functional by (W); and time trend related to changes in
forms, and estimators. However, standardization must equipment stock as represented by (T). Estimating a
not be confused with rigidity. The optimal methodol- model based on this function for the years prior to
ogy must be replicable from situation to situation, but energy efficiency policy implementation, Y can then
must also be adaptable. If they are to be true to reality, be projected or simulated in the post-policy years.
energy demand analyses must sometimes differ in This provides the necessary counterfactual by which
their sampling frames, their time frames, their models, the impacts of the policies, i.e., energy savings, can be
and other specifics. The examples provided here empirically measured. Note that it is occasionally
demonstrate these qualities; they take advantage of suggested that economic variables, especially energy
information and resources that are unique to individ- prices, should be omitted from this function so that
ual circumstances while remaining similar in key impacts of non-economic policies can better be
features. detected (e.g., Sheppard et al. 2009). However, this
suggestion is not consistent with economic and
econometric theory. Models containing economic
Single state, annual pre–post analysis variables and non-economic variables are readily
capable of being used to isolate the effects of each.
The simplest counterfactual simulation approach Given the basic function just described, the
uses a time series model for a single subject, such strength of this approach is easily demonstrated.
as a state or a utility, to compare its actual energy Figure 1 depicts actual and simulated electricity use
demand to that which would have been demanded (in BBTU) for four cross-section (state) subjects, i.e.,
in the absence of an energy efficiency policy. There New York, Oregon, Illinois, and Nevada. The
are two features of this approach that determine its electricity sales data, taken from EIA’s SEDS database
success. The first is historical knowledge. Knowl- covers 100% of all residential sector use in these
edge of a subject’s energy efficiency programs and states. For convenience, each state is assumed to have
policies is necessary for selecting a date, or dates, changed its residential sector electricity efficiency
when policy changes are likely to have occurred or policy after 1991 and to have an identical set of
when their effects are likely to be detectable. In determinants of aggregate demand. For the period
addition, background knowledge of other major from 1970 to 1991, the model is specified such that
events, or social and economic changes, that may annual state average residential electricity use per
have also affected demand over the study period is customer is regressed on average state electricity
critical for developing and testing hypotheses and prices, average state natural gas prices, average state
for interpreting findings. The second ingredient for heating and cooling degree days, average state per
success is knowing enough about a subject to be capita income, and a national equipment-related stock
able to specify a reasonable econometric model of time trend. These models, like all the models in this
aggregate energy demand. Given these prerequi- paper, are estimated in log–log form. Model statistics
sites, a general function for a reduced-form time for this analysis are contained in the “Appendix”.
series energy demand model, where the subscript In Fig. 1, the shaded area represents the pre-regime
t represents a given year is: change period over which the behavioral relationships
are estimated. Each state’s model is then used to
Yt ¼ f ðPt ; Gt ; Wt ; Tt Þ: retrospectively simulate the total annual residential
electricity use in the state from 1992 to 2005, taking
In this function Y could be an absolute level of into account the actual number of residential custom-
energy or of energy demand, or an energy intensity ers in each year. In the figures, the solid line
ratio of some kind. Determinants of energy demand or represents actual use and the dashed line represents
intensity may include and of course are not limited to simulated use through 2005. If the selected policy
electricity prices as represented by (P) and perhaps cutoff year is accurate, and the model does not suffer
also the price of the nearest substitutable fuel, such from misspecification, the simulation should produce
natural gas prices; absolute or relative incomes as reasonably accurate trends showing what total elec-
Energy Efficiency

240,000 72,000
New York Oregon
220,000 68,000
64,000
200,000
60,000
180,000 56,000
160,000 52,000

140,000 48,000
44,000
120,000
40,000
100,000 36,000
80,000 32,000
1970 1975 1980 1985 1990 1995 2000 2005 2010 1970 1975 1980 1985 1990 1995 2000 2005 2010
180,000 40,000
Illinois Nevada
160,000 35,000

30,000
140,000
25,000
120,000
20,000
100,000
15,000

80,000 10,000

60,000 5,000
1970 1975 1980 1985 1990 1995 2000 2005 2010 1970 1975 1980 1985 1990 1995 2000 2005 2010

Fig. 1 Total electricity (BBTU) use by state, actual and predicted

tricity demand in each state might have been had the average, marginal effects of the independent
aggregate behavior from 1992 to 2005 been unaffect- variables across all the pre-period years. Just as the
ed by energy efficiency policy. If the simulated mean forecast for the pre-policy period will not be
demand trend is well above the actual demand in the equal to the actual value of the dependent variable for
later years of the policy, it can be concluded that the the mid-year, so too the coefficients do not necessar-
energy savings from the policy change is observed ily represent the exact mid-point of the pre-policy
and verified. time period. This feature does not in any way dilute
It is essential to understand that the baseline for the accuracy of findings—if anything, it provides a
these analyses (and unless otherwise noted for all the more honest interpretation of the baseline, since in
other analyses using this econometric approach) is reality the baseline is a moving target that should not
approximately the mid-point of the pre-policy period, be viewed as a simple, single point in time.
in this case around the early 1980s, since the pre- From the individual state figures it can be seen that
policy period stretches from 1970 to 1991. This must policy savings appears to have occurred in New York
be so because the coefficients of the models that are and Oregon. These findings seems reasonable given
used to forecast demand in the post-policy period are that these two states had vigorous residential energy
Table 1 Four state energy
efficiency policy impacts for State Actual Simulated Rel. SE (%) 90% error bound 2005 policy impact (%)
2005
Use (BBTU) Use (BBTU) Lower Upper

NY 172,418 221,997 8.7 190,193 253,801 −28.8


OR 62,572 71,665 4.4 66,526 76,804 −14.5
IL 165,798 173,052 13.1 135,835 210,270 −4.4
NV 37,804 34,906 6.5 31,151 38,662 7.7
Energy Efficiency

efficiency policies for at least the past 15 years. It is unlikely that any other non-energy policies or major
also no surprise that in Illinois and Nevada, two states events affected their trends.
without aggressive residential energy efficiency pol- As stated above, this analysis of each state’s
icies, actual and predicted total use do not differ by policies is a starting point and should not be taken
much. Table 1 presents the data for these four states as a definitive policy evaluation. Obviously, much
for 2005, with Rel. SE, or relative standard error, more time and effort must be put into studying the
defined as the standard error of the simulated use impacts of energy efficiency policies in these and all
divided by the level of simulated use. In both New the other states, sectors, and fuels. The analysis is
York and Oregon, actual use is below the 90% lower offered to demonstrate the flexibility of this approach
error bound of predicted use for 2005, whereas in and that it need not be overly costly or complex to
both Illinois and Nevada, actual use is well within the provide accurate, consistent, and unbiased estimates
expected range of predicted use. As such, the of total, long-term policy-related energy savings.
estimated 29% policy impact in New York in 2005, Before going on to describe a second application of
and the estimated 15% savings impact in Oregon are counterfactual simulation it is worth emphasizing that
statistically significant at the 90% level, whereas the these measured savings, like those for all counterfac-
estimated impacts in Illinois and Nevada are not. tual simulations using aggregate data, are net savings
Again it should be emphasized that these findings do in the broadest sense of the term. This cannot be
not represent the incremental savings for a single stressed enough given the current controversy in the
year, but rather the cumulative savings relative to the energy efficiency program evaluation field over the
approximate mid-point of the pre-policy period. inherent weaknesses in calculating of net-to-gross
Introducing additional information to further ex- factors for adjusting mean short-term savings. Using
plain the quantitative results is beyond the scope of the counterfactual approach, the confusion and con-
this discussion. However, it is worth noting that an troversy surrounding this issue is avoided.
obvious objection to this approach might have to do
with the issue of attribution. Attribution hinges on the
degree of faith that is placed in the model specifica- Single state, monthly pre–post analysis
tion. If it is believed that the differences found in New
York and Oregon between simulated and actual use A more detailed counterfactual analysis can be
are due to structural changes in the residential sector conducted using monthly rather than annual aggregate
rather than policy, this points to two likely modeling energy consumption data. To illustrate the adaption of
problems; (a) either the model is misspecified, the econometric approach to monthly data and the
meaning it has omitted key variables or suffers from added value it provides, historical data for the
endogeneity bias, or (b) the true demand parameters residential sector for all the Class A utilities in
have changed over the study period for reasons other Wisconsin are analyzed. These consumption levels
than energy policy. The first of these problems is represent approximately 75% of the annual residential
easily correctable by revisiting the model. The second sector electricity use in this state.
problem is also easily correctable. Economic theory To initiate this analysis, the historical relationship
argues that if demand parameters change at all, they between average residential energy prices (in constant
do so very slowly. The exceptions are when a dollars) and non-weather adjusted residential electric-
government policy targets demand or when major ity use are inspected from the earliest month avail-
events occur that change consumer preferences or able, January 1980, to the latest date available,
benefit and cost considerations. It so happens that in December 2008. Figure 2 vividly illustrates the
these analyses what is being maintained is that policy pattern of the two variables, with GWHR on the
regime change has affected demand. To confirm that left-hand axis representing aggregate residential con-
it is not another policy or set of events that affected sumption and REALKWHP on the right-hand axis
demand in New York and Oregon, all that is needed is representing real average cents per kWh. The indi-
a history of the period under inspection. In these four vidual figures represent the shoulder months of May
state examples, notwithstanding that the policy cutoff and September. These similar patterns are typical for
dates may not be as accurate as they could be, it is all the months. They show rising levels of consump-
Energy Efficiency

13 13
12 12
11 11
10 10
1,200 1,400
9 9
1,100
8 1,200 8
1,000
900 1,000
800
800
700
600 600
86 05
88 05

96 05

92 09

08 09
04 05
06 05

84 09
86 09

00 09
02 09
82 05

98 05

08 05

88 09

94 09
8 4 05

90 05
92 05
9 4 05

00 05
02 05

05

82 09

9 0 09

96 09
98 09

04 09
0 6 09

09
M
M

M
M
M

M
M
M

M
M
M

M
M

M
M

M
M
M

M
M

M
M
M

M
M

M
M
80

80
GWHR REALKWHP GWHR REAKWHP

Fig. 2 WI class A utilities residential energy sales and prices, May and September (1980 to 2008)

tion over the three decades, coupled with declining are presented in Table 2 where GWHRF represents
prices until 1998, and rising prices thereafter. forecasted use and GWHRFSE represents the forecast
Clearly there was a dramatic shift in the relation- standard error.
ship of the two variables; this is confirmed with The findings show that savings varied by month.
diagnostic tests. The Quandt-Andrews breakpoint test However, caution should be used in overinterpreting
reveals what the naked eye can observe more quickly; the results and trying to look for specific meaning in
that a naive regression analysis that combines all the estimates for each month. Rather, the overall
years is likely to run into serious estimation and 2-year savings should be observed. It was about
interpretation problems. The most plausible explana- 2.8%, plus or minus approximately 0.5% at the 95%
tion for the structural break around 1998, backed by confidence level, relative to the baseline, roughly the
actual data, is that by the late 1990s the state of mid-year of the pre-policy period. Model statistics for
Wisconsin lacked its own generating capacity to meet this analysis are contained in the “Appendix”.
its increased demand and was forced to import higher-
priced electricity from Canada and other out-of-state
providers (Wisconsin Division of Energy 2006).
While a broad historical analysis may be of interest
2,000
for many other purposes, for the purposes of Actual Monthly GWH = Solid 2007m01
Predicted Monthly GWH = Dotted Line
estimating the recent impacts of energy efficiency 1,800
policies, this historical knowledge and empirical
evidence helps direct the focus of the demand analysis 1,600
to the post-1997 period.
To estimate the aggregate impact of energy 1,400
efficiency programs in 2007 and 2008, a monthly
model is estimated using the first month in 1998 up 1,200
through the last month in 2006. Then, using the actual
values of the independent variables in 2007 and 2008, 1,000
the monthly energy use for these years are forecasted.
Figure 3 provides a visualization of the analysis, with 800
the solid area representing actual monthly consump- 98 99 00 01 02 03 04 05 06 07 08

tion and the dotted line representing the monthly Fig. 3 WI energy demand model actual and predicted:
forecasts while the numerical results for the 24 months estimation period—1998 to 2006 (108 months)
Energy Efficiency

Table 2 WI energy
efficiency policy impacts Obs. GWHR GWHRF GWHRFSE Obs. GWHR GWHRF GWHRFSE
for 2007 and 2008
2007M01 1,555 1,502 125 2008M01 1,688 1,527 129
2007M02 1,506 1,333 111 2008M02 1,458 1,400 117
2007M03 1,336 1,518 126 2008M03 1,375 1,511 127
2007M04 1,156 1,195 99 2008M04 1,183 1,205 101
2007M05 1,173 1,306 109 2008M05 1,157 1,267 107
2007M06 1,474 1,514 127 2008M06 1,290 1,547 132
2007M07 1,650 1,637 137 2008M07 1,617 1,716 146
2007M08 1,729 1,641 137 2008M08 1,567 1,676 142
2007M09 1,338 1,318 110 2008M09 1,243 1,335 113
2007M10 1,243 1,331 111 2008M10 1,172 1,268 106
2007M11 1,284 1,508 128 2008M11 1,350 1,486 126
2007M12 1,664 1,548 132 2008M12 1,651 1,530 130
Total 17,108 17,349 1,454 Total 16,751 17,469 1,475
Savings 242 718
% savings 1.4 4.3
Rel. SE 8.4 8.4

Multiple utilities, annual pre–post analysis Utility District (SMUD). Together, these utilities
represent 82% of 2007 residential electricity sales in
An extension of the single time series counterfactual California.
simulation involves using several cross-section sub- To produce utility-specific data, the California
jects, be they nations, states, utilities, etc., to estimate Energy Commission (CEC) map of electric utility
jointly the impacts of energy efficiency policies. The service territories was used to allocate each of the 58
general function for a cross-section time series energy counties in California to a single utility. Where two or
demand model, where the subscript t represents a more utilities served the same county, the utility that
given year and s represents a give cross-section and served the largest share of the county’s population
the variables as defined above is: was designated the utility for that entire county.
 Monthly heating and cooling degree days (HDD and
Yt;s ¼ f Pt;s ; Gt;s ; Wt;s ; Tt;s : CDD at base 65°F) for all the major weather stations
This version of the counterfactual approach is more in California whose data covered the dates used for
complicated than the former because it hinges on this analysis (these are known as WBAN stations,
finding compromises that work for all the cross- short for “Weather Bureau Army Navy”) were then
sections, be they in establishing reasonable dates for acquired from the National Climatic Data Center
regimes changes, specifying a common model, or (NCDC). Based on weather zones and proximities, a
identifying and controlling for variations in energy total of 23 of these weather stations were assigned to
efficiency policies. However, it can also yield higher the 58 counties. To derive the average HDD and CDD
quality coefficients than a single time series because it for each utility, the final step in this process was to
boosts the amount of information available for weight the weather data by the population in each
analysis. county, a procedure that is similar to the one used by
To demonstrate the feasibility and value of using a NCDC to derive state-level averages.
time series cross-section approach at the utility level, Annual residential electricity sales and prices data
residential electricity sales are analyzed for the four for each utility was taken from the CEC’s Energy
major utilities in the state of California, i.e., Pacific Consumption Data Base (ECDMS). These data
Gas and Electric Company (PG&E), Southern Cali- extend as far back as 1990 and, for the residential
fornia Edison (SCE), San Diego Gas and Electric sector, are almost identical to the data available from
Company (SDG&E), and Sacramento Municipal the US Energy Information Administration’s Form
Energy Efficiency

18
Inspection of California’s electricity prices are
instructive on many levels. They suggest that analyses
16 of California’s residential energy demand, at least
Real Cents per kWh

since the late 1990s, are fraught with danger and they
14 illustrate why some of the major features of the
econometric analyses used to derive counterfactuals
for estimating energy efficiency impacts cannot be
12
standardized but rather must take into consideration
the unique circumstances of the subject at hand. No
10 matter what the results of short-term evaluations,
none are capable of anticipating the additional
8 savings, or the reduction in savings, that can occur
1 - 90
1 - 94
1 - 98
1 - 02
1 - 06
2 - 91
2 - 95
2 - 99
2 - 03
2 - 07
3 - 92
3 - 96
3 - 00
3 - 04
3 - 08
4 - 93
4 - 97
4 - 01
4 - 05
from dramatic changes caused by economic, social,
technical, and political conditions.
1 = PG&E 2 = SCE 3 = SDG&E 4 = SMUD In implementing this multi-utility time series
analysis, a similar set of variables and a similar
Fig. 4 CA annual average residential electricity prices (1990–2007)
functional form is used as in the previous analyses for
the individual states. Only here, a weighted least-
EIA-861 database. The final data used for the cross- squares, fixed effects model is specified, and the pre-
section time series utility-level energy demand anal- policy period is the span of years from 1990 through
ysis was taken from the US Bureau of Economic 2005 with energy efficiency impacts estimated for
Analysis Regional Economic Information System 2006 and 2007. Note that since residential energy
which contains economic data for all the counties in efficiency programs have been widespread and
the USA. These county data were summed to arrive at growing in the California since the late 1970s, the
utility totals for variables, such as population and estimated impacts must be interpreted as the marginal
personal income. impact of the energy efficiency programs run in 2006
Prior to undertaking the energy demand analysis, and 2007 relative to the average impacts of the
average annual prices per utility were plotted. These programs run between 1990 and 2005. Table 3
prices are presented in Fig. 4, in which it can be seen presents the California utility results. Again, model
that California residents did not face the slowly statistics for this analysis are contained in the
changing administered prices usually found in most “Appendix”.
states and at most electric utilities. Rather, real In total, it appears that residential energy efficiency
average prices were erratic, most likely due to the programs in California led to a 2-year decrease in
upheaval caused by electric utility deregulation in the energy as of 4.8%. As noted previously, these results
mid-1990s that culminated in California’s disruptive are not interpretable as the incremental savings for
energy episodes in the early part of the present these 2 years combined. Rather, they are the cumu-
decade. lative savings of energy efficiency policies relative to

Table 3 CA energy effi-


ciency policy impacts for Util Obs. GWHR GWHRF GWHRFSE Av. Rel. SE (%) Total diff Total % diff
2006 and 2007 (sum of four
largest electric utilities) PG&E 1-06 30,823 32,981 1,731 5.13 −5,444 −8.8
PG&E 1-07 30,846 34,132 1,714
SCE 2-06 30,431 30,378 1,821 5.94 −847 −1.4
SCE 2-07 30,212 31,111 1,834
SDG&E 3-06 7,522 7,776 501 6.28 −416 −2.8
SDG&E 3-07 7,541 7,703 471
SMUD 4-06 4,747 4,816 277 5.57 −318 −3.4
SMUD 4-07 4,634 4,883 264
2-year total −7,024 −4.8
Energy Efficiency

the approximate mid-point of the designated pre- group includes Portland General Electric and Pacificorp,
policy period, perhaps 8 or 9 years earlier. As a result, both of whom are investor-owned utilities and have
roughly speaking the annual savings from residential been involved in the Energy Trust of Oregon (ETO).
energy efficiency policies since the late 1990s seems The third utility in the group is the Eugene Water and
to average about one-half of a percent for the state as Electric Board, a public utility that is not involved in
a whole. Food for additional thought is whether or not ETO, but that nevertheless maintains comprehensive
inter-utility comparison can be made to determine energy efficiency policies. Together, in 2006 these three
why savings are higher at some utilities compared to utilities were responsible for providing approximately
others. For this purpose, additional econometric work 80% of the residential electricity in the state.
and the introduction of individual program evaluation Using a similar set of variables to those employed
findings could be quite instructive. in the prior simulations, a weighted fixed effects
model is estimated for the small group of utilities with
significant residential electricity efficiency policies
Two groups of states, annual pre–post analysis and then the total energy demand of the group of 27
utilities is simulated. For the analysis, the estimation
To provide another example of the flexibility and power period is 1989 through 2002, and the simulation
of econometric analysis and counterfactual simulation, period is 2003 through 2006. Table 4 contains the
the differential effect of utility energy efficiency policies results of the counterfactual simulation in which the
is analyzed for Oregon electric utilities. This example behavioral coefficients of the small group of energy
also shows how the information generated from this kind policy intensive utilities were used to simulate the
of analysis can be used for understanding and quantify- energy demand of the other group of 27 utilities for
ing the behavioral effects of energy efficiency policies. all the years from 2003 through 2006. Model statistics
For consistency, the model used for this analysis derives for this analysis are also contained in the “Appendix”.
from the same function described above, the difference As can be seen in the column labeled “% diff.,” in
here being the inclusion of a subscript, R, to represent 2006 the difference between actual and simulated
entirely different energy efficiency policy regimes. The total residential electricity demand in the 27 utilities
function now appears as: was 8.8%. In other words, had the 27 utilities behaved
 like the three utilities in the 1989 to 2002 period, their
Yt;s;R ¼ f Pt;s;R ; Gt;s;R ; Wt;s;R ; Tt;R energy use would have been lower by almost 9%.
where policy regime change is assumed to occur for However, it would be a mistake to attribute all of this
some cross-sections and years and not others. difference to a lack of energy efficiency policies.
The residential energy demand model specified for Comparison of actual and simulated use prior to 2003
Oregon electric utilities consists of two groups of for the 27 utilities reveals that simulated use is
utilities in Oregon, one containing three utilities with consistently lower by an average of 2.4%. This
large, comprehensive residential electricity efficiency suggest underlying systemic differences between the
programs, the other containing 27 utilities with fewer two groups of utilities that could be done prior to
and less heavily promoted programs. The former energy efficiency policies or differences in the market

Table 4 OR energy efficiency policy impacts for 2003 through 2006 (sum of three largest electric utilities)

Year Actual use (GWH) Simulated use (GWH) % diff. Rel. SE 90% error bound Implied policy impact (%)

Lower Upper

Pre-2003 (Avg.) 15,643 15,275 −2.4 2.5 14,642 15,907 –


2003 16,477 15,925 −3.5 2.6 15,253 16,597 −1.1
2004 16,723 15,851 −5.5 2.5 15,196 16,505 −3.1
2005 17,011 16,242 −4.7 2.5 15,576 16,909 −2.3
2006 17,596 16,169 −8.8 2.5 15,494 16,844 −6.4
Energy Efficiency

penetration of electric heating or air conditioning, or a there are common sets of data collected over the same
host of other factors. As such, to derive net policy time frequencies for all the subjects. Although an
impacts the systemic difference is subtracted out, inventory of such data is far from the scope of this
leaving an implied policy impact for 2006 of 6.4%. study, it is clear that much of these kinds of data
These findings suggest that if the three largest already exist at international and national agencies.
utilities’ residential programs were adopted by the Also, it is not hard to imagine that with reasonable
group of 27 utilities, their combined annual energy effort, much of these local data can be collected from
demand would show statistically significant decreases state agencies such as state energy offices, public
at the 90% confidence level. utility commissions, and local electric and gas utilities.
This counterfactual simulation is useful not only
for verifying the impacts of energy efficiency policies,
but for understanding energy demand behavior. For Conclusion
example, the same energy demand model can be
estimated for both groups of utilities and their The fundamental issue that this paper addresses is
coefficients compared (see Appendix). For these two how to measure or verify the total, long-term energy
groups of Oregon utilities, the major difference in the savings that come from energy efficiency policies.
models is in the elasticities representing the impact of Towards this end four examples are provided that
changes in economic growth, a proxy for income. A demonstrate the feasibility, and the flexibility, of
simple comparison indicates that increases in eco- achieving this goal using time series and panel data
nomic growth in the three largest utilities are econometric analyses. Not only has this paper shown
associated with relatively less residential electricity that counterfactual simulation is adaptable to many
use per customer. This is consistent with the hypoth- different geographic circumstances and data constraints,
esis that energy efficiency policies have a beneficial but it has also shown that these kinds of analyses yields
influence on the relationship between increasing valuable information and insights that are unattainable
affluence and energy use by encouraging capital with surveys and pseudo-experimental evaluation
purchases, and behaviors, that slow energy use. designs for individual programs.
Further, to get a quantitative sense of the importance For governments or public agencies wishing to
of the differential elasticities between the two policy extend this approach further, the next step in the
regimes it is possible, using the formula below, to development of energy efficiency policy evaluation is
calculate how a change in the value of each variable, to create a location-oriented database that matches
holding all other variables constant, would affect Y, historical energy sales data with economic, demo-
energy use per residential customer: graphic, and policy-related variables. In the USA this
database can consist of more than 2,300 electric
% change in Y ¼ 1  ð1 þ % change in X ÞbX : utilities and more than 1,500 natural gas utilities, most
Using this formula and the estimated model of whom have at least 10 years of energy sales and
coefficients for economic growth, an assumed future revenues information available for analysis. With this
increase in economic growth statewide of 25% is degree of differentiation and data depth, many
projected to lead to a decline in average residential different research designs are possible depending on
customer use of over 11% in the top three utility the nature of the counterfactual being investigated.
service territories, whereas in the other utility service On the international scene, the levels of differentia-
territories, average residential customer use would tion between locations and policies may be quite
hardly decline at all. different than that found in the USA. Nevertheless, it is
These illustrations show that a wealth of informa- likely that every country collects time series data that are
tion can be derived from the counterfactual simulation related to annual energy sales and other basic economic,
approach as it is applied to groups of subjects with demographic, and policy variables. With even rudimen-
different policy regimes. It should also be obvious tary data, the counterfactual simulation approach to
that this approach is applicable not only to groups of estimating policy impacts can and should be tried.
utilities, but to groups of states, or nations, or Finally, it should be noted that the approach to
communities. The only technical requirement is that evaluating energy efficiency programs in aggregate that
Energy Efficiency

is described and advocated in this study is not intended as complimentary and as providing new kinds of
to replace existing methods for evaluating individual program net energy savings estimates, ones that are not
energy efficiency programs. Individual program impact satisfactorily provided by current methods.
evaluations are essential to program management and
design, not to mention cost recovery and incentives. Acknowledgements This research was funded by the U.S.
Environmental Protection Agency. The author would like to
They provide immediate feedback to program imple- thank Andy Henderson and Emanuel Klein for their sugges-
menters and policymakers, and permit programs to be tions and advice, Anna Horowitz for her research assistance,
modified, improved, and optimized on a regular basis. and Jim Mapp for providing the monthly MWh sales data for
Counterfactual simulations for measuring policy savings Wisconsin. The author takes sole responsiblity for all errors and
opinions found within.
cannot provide the same information as do these
program evaluations. They should therefore be viewed

Appendix
Single state, annual pre–post analysis

Dependent variable: residential electricity consumption


(BBTU) per capita. Estimation time period: 1970–1991,
forecast time period 1992–2005

Illinois Actual values


Model variables Coeff. SE Prob. 1970 1991 2005
Constant −8.254 0.829 0.000
Electricity price −0.137 0.119 0.265 28.954 34.251 21.643
Natural gas price 0.058 0.055 0.306 3.706 5.750 10.135
Per capita income 0.849 0.311 0.016 16.596 25.123 32.294
Equip. stock trend −0.016 0.055 0.775 1.476 11.574 123.826
Heating degree days 0.036 0.089 0.689 6,290 5,836 5,929
Cooling degree days 0.150 0.040 0.002 973 1,153 1,137
Adj. R-squared 0.932 DV= 0.007 0.011 0.013

Nevada
Model variables Coeff. SE Prob. 1970 1991 2005
Constant −11.766 2.097 0.000
Electricity price 0.104 0.150 0.499 16.192 20.451 26.443
Natural gas price −0.227 0.095 0.030 5.059 6.410 10.452
Per capita income 0.805 0.365 0.043 17.927 24.586 33.161
Equip. stock trend −0.080 0.056 0.174 1.476 11.574 123.826
Heating degree days 0.462 0.177 0.019 4,321 4,133 3,501
Cooling degree days 0.199 0.170 0.260 1,536 1,675 2,088
Adj. R-squared 0.450 DV= 0.014 0.015 0.016

New York
Model variables Coeff. SE Prob. 1970 1991 2005
Constant −7.106 1.436 0.000
Electricity price 0.225 0.129 0.101 32.066 41.548 40.779
Natural gas price −0.038 0.074 0.612 4.973 8.469 12.821
Per capita income 0.042 0.177 0.817 17.700 28.380 36.309
Energy Efficiency

Equip. stock trend 0.181 0.049 0.002 1.476 11.574 123.826


Heating degree days 0.092 0.107 0.406 6,092 5,262 5,984
Cooling degree days 0.013 0.050 0.799 750 804 944
Adj. R-squared 0.960 DV= 0.005 0.007 0.009

Oregon
Model variables Coeff. SE Prob. 1970 1991 2005
Constant −8.085 0.824 0.000
Electricity price −0.165 0.091 0.089 13.268 16.704 18.817
Natural gas price −0.003 0.032 0.935 5.251 7.047 10.934
Per capita income 0.413 0.132 0.007 14.252 21.940 27.950
Equip. stock trend −0.006 0.021 0.795 1.476 11.574 123.826
Heating degree days 0.383 0.068 0.000 5,240 5,056 4,937
Cooling degree days 0.008 0.017 0.651 220 267 282
Adj. R-squared 0.734 DV= 0.016 0.019 0.017

Single state, monthly pre–post analysis

Dependent variable: monthly residential electricity


consumption (GWH) per state personal income.
Estimation time period: 1998–2006, forecast time
period 2007–2008

Wisconsin Actual values


Model variables Coeff. SE Prob. Jan. 1998 Jun. 2002 Dec. 2006
Constant 3.409 0.264 0.000
Summer (J,J,A) 0.088 0.030 0.004 0 1 0
Winter (N,D,J,F,M) 0.223 0.019 0.000 1 0 1
Electricity price −0.373 0.173 0.033 8.273 9.185 10.281
Equip. stock trend 0.132 0.046 0.005 66.882 90.352 172.866
Heating degree days −0.043 0.010 0.000 1,278 68 1,068
Cooling degree days 0.012 0.006 0.045 0.000 162 0.000
Adj. R-squared 0.634 DV= 25.548 23.615 23.887

Multiple utilities, annual pre–post analysis mation time period: 1990–2005, forecast time period
2006–2007
Dependent variable: utility residential electric con-
sumption (GWH) per utility personal income. Esti-

4 largest CA Utilities Actual mean values


Model variables Coeff. SE Prob. 1990 1998 2005
Electricity price −0.103 0.080 0.205 13.995 12.267 13.488
Shock2000–2003*Ele. price −0.024 0.007 0.001 0 0 0
Equip. stock trend −0.091 0.009 0.000 9.738 70.409 131.815
Heating degree days −0.063 0.070 0.373 2,194 2,339 1,900
Cooling degree days 0.035 0.044 0.421 1,309 1,037 1,151
Adj. R-squared 0.976 DV= 83.3 75.3 65.8
Energy Efficiency

Two groups of states, annual pre–post analysis

Dependent variable: residential electricity consump-


tion (MWH) per customer. Estimation time period:
1989–2002, forecast time period 2003–2006

3 largest electric utilities Actual mean values


Model variables Coeff. SE Prob. 1989 2002 2006
Electricity price −0.119 0.029 0.000 0.057 0.069 0.067
Natural gas price −0.147 0.021 0.000 7.606 9.774 12.149
Economic growth −0.543 0.114 0.000 7.356 7.237 7.353
Equip. stock trend −0.029 0.003 0.000 8.373 100.000 143.356
Heating degree days 0.406 0.037 0.000 5,217 4,993 4,933
Adj. R-squared 0.981 DV= 13.841 11.926 11.817

27 smaller electric utilities


Model variables Coeff. SE Prob. 1989 2002 2006
Electricity price −0.220 0.021 0.000 0.059 0.063 0.061
Natural gas price −0.089 0.012 0.000 7.606 9.774 12.149
Economic growth −0.024 0.012 0.044 9.312 8.753 8.188
Equip. stock trend −0.003 0.002 0.116 8.373 100.000 143.356
Heating degree days 0.404 0.024 0.000 5,217 4,993 4,933
Adj. R-squared 0.971 DV= 15.098 14.435 14.162

References Metcalf, G. (2008). An empirical analysis of energy intensity


and its determinants as the state level. The Energy Journal,
29(3), 1–27.
Bernstein, M., Fonkych, K., Loeb, S., & Loughran, D. (2003). National Action Plan for Energy Efficiency. (NAPE, 2007).
State-level changes in energy intensity and their national Model energy efficiency program impact evaluation guide.
implications. Santa Monica: Rand, MR-1616-DOE. Northeast Energy Efficiency Partnerships. (2008). Regional
EMEEES (2008). Harmonised methods for evaluating energy EM&V forum’s three year plan.
end-use efficiency and energy services—the European Sheppard, Chamberlin, & Jacobson (2009). Exploring strate-
Commission’s project on evaluation and monitoring for gies for implementing a performance based state efficiency
the end-use directive on energy end use and energy program: State energy consumption metrics—residential
services, Final Conference, Brussels, BE. sector analyses. Schatz Energy Research Center, Humbolt
Fisher, F. M., & Kaysen, C. (1962). A study in econometrics: State University.
The demand for electricity in the United States. Amsterdam: Taylor, L. D., Blattenberger, G. R., & Rennhack, R. K.
North Holland Publishing Co. (1984). Residential energy demand in the United States:
Halvorsen, R. (1975). Demand for electric energy in the United introduction and overview of alternative models. Advan-
States. Southern Economic Journal, 42(4), 610–625. ces in the Economics of Energy and Resources, 5, 85–
Horowitz, M. J. (2004). Electricity intensity in the commercial 105.
sector: market and public program effects. The Energy U.S. Department of Energy’s Office of Energy Efficiency and
Journal, 25(2), 115–137. Renewable Energy (IPMVP, 2002 and updated in 2007).
Horowitz, M. J. (2007). Changes in electricity demand in the United States International Performance Measurement & Verification
from the 1970s to 2003. The Energy Journal, 28(3), 93–119. Protocol.
Loughran, D. S., & Kulick, J. (2004). Demand side management Wisconsin Division of Energy. (2006). Wisconsin energy
and energy efficiency in the United States. The Energy statistics (2006). Madison: Wisconsin Department of
Journal, 25(1), 19–43. Administration.

Você também pode gostar