Electric Utility Demand Side Management in Canada

Nic Rivers* and Mark Jaccard**

Government, utility, and private subsidies for energy efficiency play a

prominent role in current efforts to reduce greenhouse gas emissions, yet the
effectiveness of this policy approach is in dispute. One opportunity for empirical
analysis is provided by the past energy efficiency subsidies, called demand-side
management programs, offered by electric utilities in North America over several
decades. Between 1990 and 2005, most electric utilities in Canada administered
such programs, with total spending of $2.9 billion (CDN$2005). This paper uses
the significant inter-annual variation in demand side management spending dur-
ing this period to econometrically estimate the effectiveness of these subsidies.
The resulting estimates indicate that these programs have not had a substantial
impact on overall electricity consumption in Canada.
doi: 10.5547/ISSN0195-6574-EJ-Vol32-No4-5

1. INTRODUCTION

Electric utility demand side management (DSM) programs were con-

ceived following the dramatic energy price increases of the 1970s and early 1980s.
Increases in fuel prices during this period were accompanied by high interest rates
that significantly increased the cost of building and operating new power plants
[Gellings, 1996]. Responding to the increase in energy prices, the US government
implemented a wave of new policies aimed at stimulating energy efficiency, in-
cluding The Energy Policy and Conservation Act (1975), The Energy Conser-
vation and Production Act (1976), The National Energy Conservation Policy Act
(1978), and The Public Utilities Regulatory Policy Act (1978) [Gillingham et al.,
2006]. Electric utility regulators, especially in the US, were also concerned with

The Energy Journal, Vol. 32, No. 4. Copyright 䉷2011 by the IAEE. All rights reserved.

* Corresponding author. Graduate School of Public and International Affairs and Institute of the
Environment, University of Ottawa. E-mail: [email protected].
** School of Resource and Environmental Management, Simon Fraser University. E-mail:
[email protected].
doi: 10.5547/ISSN0195-6574-EJ-Vol32-No4-5

93
94 / The Energy Journal

rising energy prices and with potential misinvestment risks of large-scale gener-
ation, and so were persuaded that utilities should be required to foster improved
efficiency as well as load shifting by their customers. (By reducing peak demand,
load shifting delays the need for new power plants.) Pushed by these interrelated
drivers, electric utilities in the US initiated DSM programs that initially focused
on load shifting and then increasingly on electric end-use efficiency. Although
load shifting programs remain important today, most of the DSM focus through-
out North America in the past two decades has been on end-use energy efficiency.
In Canada, DSM developed more slowly [Jaccard, 1993]. First, the elec-
tricity industry, being especially based on hydropower and coal, did not experi-
ence the financial crisis of its US counterpart, although Ontario Hydro was caught
with significant cost overruns in its nuclear investments. Second, most of the
electricity industry was (and still is) publicly owned with little oversight by utility
regulatory agencies, so there was no external force to require consideration of
energy efficiency. Third, authority over the electricity system is more decentral-
ized in Canada than in the US, so there is no equivalent federal act in Canada to
those in the US prompting energy efficiency investments. The development of
DSM in Canada therefore depended to a significant degree on personalities and
political preferences. In 1986, the CEO of BC Hydro took a personal interest in
DSM and launched Power Smart, a division that eventually had influence across
the country. Similar developments in Ontario and eventually Quebec led to a
patchwork of DSM efforts across the country.
The first utility DSM programs were implemented in California and the
US northwest, and have since spread throughout the United States and Canada
[Nadel and Geller, 1996]. Most large electric distribution utilities in these two
countries now have some experience with such programs. Total electric utility
spending on DSM in the US between 1990 and 2005 was $36 billion (US$2005),
while total spending in Canada over the same period was about $2.9 billion
(CAD$2005) (see Figure 1 for sources). Once the relative size of the two countries
is accounted for, expenditures are relatively similar, although the US out-spent
Canada somewhat: Canada spent about $6.01 (CAD$2005) per person per year
on DSM between 1990 and 2005, compared to $9.21 (US$2005) in the US. The
most aggressive utilities in both countries spend up to 4 percent of their total
revenue on DSM, with most utilities spending less than 1.5 percent of total rev-
enue.
As shown in Figure 1, spending on DSM in both the US and Canada
peaked in 1993, with US utilities spending about $3.7 billion (US$2005) and
Canadian utilities spending about $550 million (CAD$2005). As electricity mar-
ket restructuring efforts intensified throughout much of the US in the mid-1990s,
spending on DSM by utilities fell significantly [Nadel and Geller, 1996]. Despite
the more tentative nature of electricity restructuring in Canada, spending fell even
more precipitously than in the US during the latter half of the 1990s. This is
particularly as a result of the cessation of all DSM programs by Ontario Hydro
(formerly the largest-spending entity in Canada) in the mid-1990s, as government
 Electric Utility Demand Side Management in Canada / 95

Figure 1: Electric Utility Demand Side Management Expenditures in U.S.

and Canada, 1990–2005

U.S. data from U.S. Energy Information Administration Electric Power Annual. Canadian data was
collected for this study from the sources described herein.

prepared it for restructuring. However, in both countries, DSM programs have

enjoyed a resurgence during the last few years. The recent increase has been
prompted in part by the volatility and high prices in international energy markets,
but also by new concerns about energy-related greenhouse gas emissions. Based
on recent utility resource plans, it is likely that DSM spending in Canada in the
next few years will surpass the previous high set in 1993.
Despite the substantial investment in DSM over the past two decades,
the impact of the spending is not possible to discern directly and remains the
subject of significant controversy, both in academic literature and in utility reg-
ulatory processes. Unlike supply side investments in new generation capacity,
DSM investments do not result in tangible, utility-owned assets whose output can
be measured directly. Instead, electric utilities must infer the results of DSM
programs through program evaluations that seek to estimate the incremental (or
96 / The Energy Journal

additional) effect of DSM programs on the underlying natural evolution of energy

or electricity efficiency in a given economy. While these have improved over the
last two decades, effective evaluation of DSM programs remains a difficult and
contentious exercise.
This paper aims to estimate the degree to which utility DSM programs
have reduced electricity consumption in Canada over the past two decades using
a method significantly different than that used by utilities. Following the example
of some studies of the US, we use the significant inter-temporal variation in DSM
effort by the utilities in Canada as a quasi-experiment to econometrically infer
the effectiveness of these programs. Unlike previous studies, we estimate a partial-
adjustment dynamic model of electricity demand, from which both short-run and
long-run responses of the economy to DSM expenditure can be predicted. While
this method is associated with its own set of challenges, it avoids some of the
weaknesses of more common forms of program-focused DSM evaluations.
This type of study can contribute to two related public policy debates.
First, because utility DSM spending, by both electricity and natural gas utilities,
is projected to increase rapidly in Canada during the coming years, it is important
to have an understanding of the potential reduction in energy demand that results
from such spending, and the cost of energy efficiency relative to new investments
in electricity supply. By using a method that is not normally employed by utilities,
this study can provide a confirming or contrasting estimate about the degree to
which DSM programs can reduce energy demand in a given region.
Second, and perhaps more importantly, this study can provide an indi-
cation of the likely effectiveness of any subsidy program aimed at inducing energy
efficiency. With increasing concerns about climate change, subsidy programs for
energy efficiency are increasing in scope and magnitude. These may be govern-
ment subsidy programs, utility DSM subsidy programs or even the increasingly
popular carbon offsets programs, in which individuals or firms voluntarily pay
offset companies to subsidize actions by third parties that apparently reduce GHG
emissions from what they otherwise would be. Energy efficiency is frequently a
target of these voluntary subsidies. Since there is little in the way of formal ex
poste analysis by governments or independent agencies to measure the effective-
ness of these subsidies in inducing energy efficiency and thus reducing GHG
emissions, this evaluation can provide an independent indication of their likely
effectiveness, with important implications for the choice of policy instrument for
achieving provincial, national and global GHG reduction targets.
The remainder of the paper is structured as follows. In section 2, we
summarize estimates made by utilities relating to the effectiveness of DSM pro-
grams, and also describe the challenges associated with such evaluations. In sec-
tion 3, we describe academic studies that have sought to estimate the effectiveness
of US DSM programs. In sections 4 and 5, we outline the method and data that
are used in this study to assess the effectiveness of Canadian DSM programs. In
section 6, we present the results of the model, and in section 7, we conclude.
 Electric Utility Demand Side Management in Canada / 97

2. ELECTRIC UTILITY ASSESSMENT OF DSM EFFECTIVENESS

Most utilities that invest significantly in DSM also invest in measurement

and verification programs to determine the effectiveness and cost-effectiveness
of such spending. Eto et al. [1995], in a survey of commercial sector DSM pro-
grams in the US, report that the mean utility-estimated cost for conservation
through DSM programs is $0.027/kWh. This is similar to recent estimates in
Canada. For example, Manitoba Hydro, in its 2006 DSM plan, estimated a utility
cost of $0.019/kWh of avoided demand. BC Hydro, in its 2006 Integrated Elec-
tricity Plan, reported costs of $0.032–0.076/kWh. Similarly, the Ontario Power
Authority estimated a levelized cost of $0.020/kWh for DSM programs offered
by local distribution companies in 2006. In nearly all cases, estimates made by
utilities suggest that reducing electricity demand through DSM expenditures is
significantly less expensive than investing in new generation.
Utilities use a variety of techniques to estimate the effectiveness and
cost-effectiveness of DSM programs, both prior to program implementation and
also after the program has been running. While these techniques can be quite
sophisticated, and benefit from the detailed micro-data on electricity sales avail-
able to the utility, such program level evaluations must make difficult judgements
about important factors that are key to program effectiveness: the free ridership
rate, the spillover rate, and the rebound effect. Unfortunately, there is no widely
accepted method for estimating these factors.

Free Riders

DSM programs are generally targeted at providing incentives for con-

sumers or businesses to adopt more energy efficient equipment. However, adop-
tion of energy efficient equipment also occurs in the absence of DSM programs,
as old technologies become obsolete and are replaced with newer and more energy
efficient technologies. For example, Natural Resources Canada [2006] estimates
that houses built in 2002 are about 35 percent more energy efficient than similar
houses built in the 1970s, and that new models of many major household appli-
ances were almost twice as efficient in 2002 compared to 1990. These changes
in efficiency can be the result of (1) an evolution toward stricter government
efficiency regulations in buildings and equipment, (2) utility and government
DSM programs, (3) increases in energy prices and energy price expectations, and
(4) natural gains in energy productivity (as well as labour and materials produc-
tivity) that occur as firms and households adopt newer technologies.
Given this mix of potential causal factors for improvements in energy
efficiency, the main challenge is to estimate how much adoption of energy effi-
cient technologies would have occurred without the DSM program, and thus how
much can be attributable to the program [Horowitz, 2004]. There is an important
temporal dimension to this challenge: given the general trend towards improved
98 / The Energy Journal

efficiency of technologies, part of the effect of a DSM program is likely to shift

the timing of investments, rather than the overall magnitude. When a DSM pro-
gram is applied, it cannot normally distinguish between individuals who would
have adopted the energy efficient technology anyway, called free riders, and those
who required the DSM subsidy to do so [Loughran and Kulick, 2004]. Free riders
add to the utility cost of a subsidy program without contributing to its effective-
ness.1
Although accounting for free riders was less common when DSM pro-
grams were first introduced, most large utilities now attempt to account for them,
typically by conducting follow up surveys of program participants asking whether
they would have adopted a particular energy efficiency measure in the absence
of the DSM program. Those participants who indicate they would have adopted
the technology even in the absence of the subsidy are considered free riders. As
an alternative, some utilities calculate free riders on a given program by compar-
ison with a group of non-participants, either within the region or in a different
region. The percent of non-participants that adopts the energy efficiency measure
is considered to be the free rider percentage. Calculations by utilities using these
methods often show fairly low free ridership levels. For example, a survey of the
40 largest commercial sector DSM programs in the US by Eto et al. [1995] found
an average free ridership rate reported by utilities of 12.2 percent.
However, both of these methods have drawbacks, which result in un-
certainty and potential bias in free ridership estimates. In particular, stated pref-
erence data can be biased because when answering a survey, consumers do not
face real world constraints (e.g., time, budget, or information constraints). Errors
or bias may arise if consumers do not understand the survey properly, have dif-
ficulty recalling historical decisions, or if they purposefully bias their answers to
alter the survey results [Louviere et al., 2000].
Estimating free ridership can equally be a challenge when comparing
adoption of energy conservation measures associated with the DSM program with
adoption by a control group within or outside the region with the program. Non-
participants within a region are by definition those customers who are less likely
to undertake conservation activities than program participants, and are therefore
an inappropriate control group [Hartman, 1988]. Non-participants outside of the
region may not be comparable with participants within the region for reasons that
are both observable (income, house size, etc.) as well as unobservable (attitudes
towards conservation, willingness to adopt new technologies, etc.). While an in-
ter-region comparison can account for observable differences, it cannot usually
account for unobservable differences in the cross-sectional approach that is used
in most program evaluations. To the extent that these are important in explaining

1. Formally, this is an adverse selection problem, created by differences in information between

the electric utility and its customers. However, the utility evaluation literature refers to the issue as a
free-ridership problem, and we follow this nomenclature here.
 Electric Utility Demand Side Management in Canada / 99

technology adoption, inferences made from such comparisons will yield incorrect
estimates of free ridership.
The academic literature contains a variety of techniques that attempt to
determine the number of free riders on a DSM subsidy program in a way that
avoids the issues discussed above. For example, Malm [1996] used a cluster
analysis to allocate households from the US Residential Energy Consumption
Survey into groups of similar households. With these groups in place, he exam-
ined heating system efficiency choice contingent upon the presence and magni-
tude of DSM programs. This technique aims to correct for some unobserved
drivers of energy efficient technology adoption that could be similar between
clusters, and results in a free ridership estimate of 89 percent for heating system
programs. Train and Atherton [1995] used combined market and survey data in
a discrete choice model to find free ridership rates of 36 percent for refrigerator
programs and 66 percent for air conditioner programs. Grosche and Vance [2009]
use a similar technique and estimate a free ridership rate of about 50 percent for
a recent German retrofit program. Loughran and Kulick [2004] estimated a panel
model relating electricity sales at all major US electric utilities to spending on
DSM programs. The results of this model suggest that DSM expenditures are
much less cost effective than claimed by utilities, indicating overall free ridership
rates (an average of all utility DSM programs in all sectors) on the order of 50–
90 percent.
In general, the free ridership rate calculated by utilities is usually sig-
nificantly lower than that calculated in independent academic research. By im-
plication, the DSM energy savings estimates of utilities will be higher and their
estimated cost of DSM lower.

Spillover

While free riders can erode the impact of a DSM program, spillover does the
opposite. Spillover occurs if a DSM program induces energy efficiency improve-
ments in addition to those directly caused by the program. For example, a par-
ticipant in an energy efficient lighting program, impressed with energy savings
from efficient lighting in a given facility, might decide to install efficient lighting
in other facilities. Additionally, a DSM measure could induce non-participants
to improve efficiency, for example by changing the market for a given technol-
ogy. In BC Hydro’s 2002–2005 seasonal light emitting diode (LED) program,
the utility estimates that virtually all seasonal LEDs sold in the province were
indirectly a result of the utility’s efforts, so that although the utility’s demand
side management program was estimated to directly reduce demand by only
0.861 GWh, the overall impact of the program was estimated to be 13.9 GWh
[Sampson Research, 2005]. Such impacts are referred to as participant spillover
and non-participant spillover respectively, and improve the effectiveness of the
program without adding to its costs. Non-participant spillover is not limited to
spillover within a jurisdiction. For instance, BC Hydro’s evaluation of its seasonal
100 / The Energy Journal

LED program attributes increased penetration of seasonal LEDs throughout Can-

ada to the program [Sampson Research, 2005].2
Estimation of spillover rates by electric utilities usually follows the same
general techniques as estimation of free ridership rates. In this case, both partic-
ipants and non-participants are surveyed with hypothetical questions about tech-
nology adoption with and without the DSM program. As with free ridership
estimates, such questions can result in poor estimates of true spillover rates if
respondents have difficulty with purchase recollection or are uncomfortable with
the hypothetical nature of the questions.
Utilities often correct estimates of DSM effectiveness for estimated spill-
over. However, we are not aware of any academic studies that attempt to estimate
spillover rates empirically in a rigorous way. Largely, this is because it is empir-
ically difficult to distinguish between real spillover effects and autonomous mar-
ket developments that would have occurred without any program in place. Ho-
rowitz [2007] comes closest, conducting a comparative study of energy efficiency
trends in each of the US states, and speculating that residential energy efficiency
programs in particular have impacts that spread rapidly between states, implying
a high spillover rate.

Rebound Effect

The potential reduction in energy consumption resulting from adoption

of an energy efficiency measure can be eroded as a result of the rebound effect.
The rebound effect exists because improving the energy efficiency of a given
energy service (while holding real fuel prices constant) reduces the cost of using
that energy service, which can increase demand for the service, and therefore
erode the net energy savings from the adoption of the energy efficiency measure.
With a high enough elasticity of substitution between energy services and other
goods, it is theoretically possible that the rebound effect overwhelms energy sav-
ings, such that increases in energy efficiency lead to increases in overall energy
consumption [Saunders, 1992].
Empirical estimates of the direct rebound effect vary significantly be-
tween studies, but Greening et al. [2000] and Sorrell [2007] conclude from a
review of several hundred empirical studies that the magnitude of the rebound
effect is ‘insignificant to moderate’, but that available studies are associated with
significant uncertainty. Specifically, for space heating in the residential sector,
they report a rebound effect of 10–30 percent (meaning that this percentage of
energy savings from adoption of energy efficiency measures is eroded due to the
rebound effect), for space cooling 0–50 percent, for lighting, 5–20 percent, and
for water heating, 10–40 percent. Results for firms are less conclusive, but suggest
a rebound effect less than 30 percent in the short run, and somewhat larger in the
long run.

2. Although the utility claimed these savings were due to its DSM efforts, these savings were not
included explicitly in its estimates of savings.
 Electric Utility Demand Side Management in Canada / 101

Although the available empirical and theoretical evidence suggests that

up to one third of the energy savings due to adoption of energy efficiency mea-
sures are eroded due to the rebound effect, utility evaluation protocols often do
not include the rebound effect in calculating the effectiveness of DSM programs,
in part because of the difficulty of obtaining appropriate estimates at a program
level. As a result, it is likely that utility estimates of energy savings overestimate
actual energy savings.
Although we will not be able to directly estimate the magnitude of the
rebound effect from our data, Sorrell and Dimitropoulos [2007] show that the
own price elasticity of energy demand is an upper bound on the direct rebound
effect. Specifically, the absolute value of the own price elasticity of energy de-
mand is equal to the direct rebound effect, provided that (among other assump-
tions) the consumer’s response to increases in energy efficiency is equivalent to
reductions in energy price, and that the consumer responds to these changes by
altering intensity of equipment use but not equipment efficiency. Since the last
of these assumptions is likely to be incorrect, the own price elasticity is likely an
overestimate of the direct rebound effect. However, the total rebound effect (in-
cluding indirect rebound) is likely to be larger than the direct rebound effect for
a variety of reasons.

3. AGGREGATE ECONOMETRIC ESTIMATES OF DSM

EFFECTIVENESS

Because of these factors that complicate program level analyses of DSM,

several studies have attempted to develop and estimate empirical models that
indirectly account for these difficulties in a way that avoids making restrictive
assumptions. Most of these studies use a time-series cross-section (panel) ap-
proach, which greatly increases the number of observations available, allows for
a more detailed model specification to be tested with greater reliability, helps to
address unobserved heterogeneity, and also enables testing of key time-dependent
lags. In particular, these studies have estimated parameters for variants of the
following relationship:

EFFit⳱DSMit bⳭX⬘it ␣Ⳮeit (1)

where EFFit isa measure of energy efficiency in region i at time t, DSMit is a

measure of DSM effort, possibly lagged by one or more years, Xit is a vector of
characteristics of the electric utility and customers in the region, eit is an error
term that may be serially correlated and heteroscedastic, and ␣ and b are param-
eters to be estimated.
Parfomak and Lave [1996] estimated a variant of equation (1) in which
DSM effort is given by utilities’ reported estimates of savings from commercial
and industrial DSM programs, and energy efficiency (EFFit ) is proxied by utility
electricity sales. They collected data from a sample of 39 utilities in the US
102 / The Energy Journal

northeast and California from 1970 to 1993 to estimate the model. This formu-
lation allows them to interpret b as a ‘realization rate’: essentially the relationship
between utility estimates of conservation and the ‘true’ realized level of conser-
vation. Estimation of the model suggests that the realization rate is very close to
100 percent, with a 95 percent confidence interval of 43 to 156 percent. Provided
utility costs for DSM programs are reported accurately, this estimate suggests that
utility cost effectiveness estimates should likewise be accurate.
Loughran and Kulick [2004] estimated equation (1) using utility-re-
ported expenditures on DSM programs as a measure of effort (DSMit) and elec-
tricity sales as a measure of efficiency (EFFit). Their data covers all large electric
utilities in the US from 1989 to 1999, and is derived from form EIA-861—a
census of electric utilities which includes DSM expenditures and savings that is
annually administered by the Energy Information Administration. With this for-
mulation, b is interpreted as the elasticity of electricity sales with respect to DSM
investments (the equation is estimated in log-log form). From this measure,
Loughran and Kulick inferred the cost effectiveness of DSM programs, given as
the cost of reducing electricity demand by one kilowatt-hour. To account for the
fact that DSM investments may influence electricity sales not just in the year of
program implementation, but also in future years, they included DSM expenditure
as a stock variable and include DSM expenditures lagged by two years in their
effort variable (DSMit), and tested (but did not report) variations of the model
with DSM lagged by as much as four years. Using this formulation, they reported
point estimates suggesting that DSM programs have cost between $0.06 and
$0.22/kWh of reduced electricity consumption, significantly higher than most
utilities themselves reported during this time period. They concluded that utilities’
estimates of ex poste free ridership rates are likely biased downwards.
The Loughran and Kulick study has been controversial, attracting re-
search that counters and supports its findings. Geller and Attali [2005] noted that
since the specification estimated by Loughran and Kulick only includes two lags
of DSM expenditures, it is likely to underestimate the cost effectiveness of DSM
programs, since these are likely to have effects that last longer than 3 years. While
Loughran and Kulick discussed model specifications involving as many as four
years of lagged DSM expenditures, Geller and Attali claim that even this may be
insufficient to fully capture the potentially highly persistent impact of DSM pro-
grams. However, this criticism was countered by Gillingham et al. [2006], who
suggested that Geller and Attali misunderstood the estimation approach used by
Loughran and Kulick. In particular, by treating DSM expenditures as a stock as
opposed to a flow in their first differenced model, Loughran and Kulick accounted
for the persistence of DSM expenditures, beyond even the two lags explicitly
estimated in their model.
Auffhammer et al. [2008] provided two alternative criticisms. First, they
suggested that the calculation used by Loughran and Kulick to convert DSM
elasticity to DSM cost effectiveness should weight utilities according to size,
rather than equally. Correcting for this makes DSM appear more cost effective
 Electric Utility Demand Side Management in Canada / 103

than in the original Loughran and Kulick study. Second, they implemented a
bootstrapping procedure to develop confidence intervals around Loughran and
Kulick’s point estimates. Since these confidence intervals include the cost effec-
tiveness estimates of utilities, frequently $0.01–0.03/kWh, Auffhammer, Blum-
stein, and Fowlie concluded that the results from Loughran and Kulick’s study
are insufficiently precise to statistically reject utility estimates of DSM effective-
ness.
Horowitz [2007] converted DSM energy savings from the EIA-861 data-
base to an ordinal measure of DSM effort at the state level (DSMit) and used state
level energy intensity as a measure of energy efficiency (EFFit) to estimate equa-
tion (1). He found that states with utilities that have a strong commitment to DSM
programs have reduced energy intensity much faster than those with a weak com-
mitment to energy efficiency programs. Although this suggests that DSM pro-
grams have been effective at reducing energy intensity, it is not possible to infer
either a realization rate or a cost effectiveness estimate from this study.
Horowitz [2004] produced another similar study, but focusing on the
commercial sector only, and using continuous, rather than ordinal, values as a
measure of DSM effort. This study suggests that DSM expenditures from 1989
to 2001 reduced commercial sector electricity intensity by roughly 1.8 percent,
suggesting a realization rate of 54 percent. This is relatively similar to the Lough-
ran and Kulick estimate, again suggesting that utilities have overestimated cost
effectiveness of DSM programs. Like Loughran and Kulick, however, Horowitz
does not provide confidence intervals around the estimate, making it hard to reject
utility estimates of cost effectiveness or realization rate.
Most recently, Arimura et al. [2009] have updated the analysis of Lough-
ran and Kulick as well as Auffhammer, Blumstein, and Fowlie. Starting from the
same base data set on utility DSM spending (but with a longer time dimension),
they include contemporaneous energy efficiency regulations as an additional ex-
planatory variable, and also include state-level spending on energy efficiency
programs that falls outside of utility DSM expenditures. Additionally, they esti-
mate the model in non-linear form, allowing for declining marginal effects of
DSM expenditures as the level of expenditures increases. They find an average
DSM cost effectiveness of about $0.06/kWh, with a 95 percent confidence interval
of about $0.025–0.11/kWh, and they find that the marginal effectiveness of DSM
expenditures decreases with additional expenditures, as predicted by theory.
Overall, these aggregate studies suggest that electric utility DSM pro-
grams in the US over the past three decades have been between 50 and 100 percent
as effective as utilities themselves have estimated. The cost of such programs per
unit of electricity consumption reduced would therefore be between equal to and
double what utilities have estimated. However, there is significant uncertainty in
these estimates, suggesting the value of additional research. This is particularly
the case since virtually all studies to date have used the same data set, derived
from EIA’s form 861. Given the ongoing nature of this dispute, and its critical
importance for future energy efficiency policy, testing of different data sets, model
formulations, and jurisdictions may help to clarify some of the key issues.
104 / The Energy Journal

4. THE MODEL

This section uses an approach similar to those described in the previous

section to estimate the impacts of DSM expenditures on energy consumption. The
estimates are based on a data set containing electric utility DSM expenditures and
electricity sales in Canada from 1990 to 2005. Since DSM programs were initiated
only in the late 1980s in Canada, this period covers nearly the entire Canadian
experience with such programs up to 2005.
The impact of utility DSM expenditures on electricity sales can be cap-
tured in the following equation:

log kWHit⳱ ␾itⳭDSMit bitⳭX⬘it ␣itⳭeit (2)

Here, kWH*it is the consumption of electricity per capita in province i in year t,

measured in kWh, DSMit is the per capita expenditure on demand side manage-
ment, Xit is a vector of length k of observed characteristics of the utility and the
customers it serves, ␾it, bit, and ␣it are coefficients, and eit is an error term.
The ‘*’ on the left hand side variable indicates desired, rather than actual,
electricity consumption. Because electricity-using capital generally lasts for sev-
eral years, and because firms and households cannot immediately adjust electricity
consumption in response to changes in electricity price, demand side management
incentives, or other variables, actual electricity consumption may deviate from
desired electricity consumption. We assume a constant-elasticity adjustment pro-
cess between actual and desired electricity consumption, such that actual elec-
tricity consumption can be modeled as:

冢 冣
t it
kWHit kWH*it
⳱ (3)
kWHi,t – 1 kWHi,t – 1

where tit is a parameter that can take on values between 0 and 1. A value of 1
indicates that full adjustment to desired energy consumption takes place imme-
diately, and implies that there is no inertia in adjusting energy consumption fol-
lowing an exogenous shock. A value of 0 indicates that no adjustment towards
desired energy consumption following a shock is possible. Taking logarithms and
substituting (2) into (3) gives:

log kWH*it ⳱tit( ␾itⳭDSMit bitⳭX⬘it ␣it –log kWHi,t – 1)

Ⳮlog kWHi,t – 1Ⳮtit eit (4)

Letting xit⳱tit␾it, dit⳱titbit, git⳱tit␣it, kit⳱1–tit, and lit⳱titeit, we obtain:

log kWHit⳱xitⳭDSMit ditⳭX⬘it gitⳭlog kWHi,t – 1 kitⳭlit (5)

Equation (5) is a partial adjustment model of electricity demand, of the type

commonly used in studies of the electricity sector as well as for other energy
 Electric Utility Demand Side Management in Canada / 105

commodities [Houthakker et al., 1974]. Although it is an ad hoc model, not

directly derived from consumer theory, it is commonly used because it is parsi-
monious and specifies an explicit dynamic adjustment process. This allows both
short-run and long-run responses to exogenous shocks to be determined. In the
current context, it allows us to estimate both short- and long-run responses to
demand side management expenditures. Assuming the demand side management
variable is measured in levels (rather than logs), a $1 expenditure on DSM in
province i at time t is associated with a dit percent change in electricity demand
in period t.3 This is the short-run response. In the long-run, the same expenditure
d
is associated with a percent change in electricity demand.
1– k
Estimation of (5) in its current form however, is impossible, since the
number of coefficients to be estimated exceeds the number of data points. To
make estimation possible, some restrictions have to be put on coefficients.
An obvious starting point is to restrict coefficients to be time invariant:

log kWHit⳱xiⳭDSMit diⳭX⬘it giⳭlog kWHi,t – 1 kiⳭlit (6)

In this formulation, coefficients are allowed to vary across utilities, so for example
the effectiveness of DSM expenditures is able to vary from one utility to another.
Estimation of (6) is possible, provided that 4Ⳮk ⬍ T. It is exactly equal to esti-
mation of separate time series models for each of the utilities in the sample. This
strategy is attractive because it puts a relatively small number of restrictions on
the model coefficients. However, in the current sample, which has a relatively
short time dimension (16 years), estimation is unlikely to yield precise coefficient
estimates. Much more precision could be gained if it were possible to restrict
coefficient values (except for intercepts) to the same value for all utilities, as in
the following:

log kWHit⳱xiⳭDSMit dⳭX⬘it gⳭlog kWHi,t – 1 kⳭlit (7)

We test whether the joint restrictions imposed by (7) are warranted by using an
F-test.
In (7), xi captures unobserved time-invariant heterogeneity across util-
ities. Because it is unobserved, we can include it with the error to form a com-
pound error term mit⳱litⳭxi. Problems can arise in estimation of (7) if mi is
correlated with the other covariates. Two data transformations are generally em-
ployed to deal with this issue: differencing and within-group demeaning. We use
the former here:4

3. We measure DSM in levels rather than logs because so that we can include utility-year pairs
where DSM expenditures are zero. Our results are qualitatively unchanged when we conduct the
analysis using logged DSM expenditures.
4. We conduct unit root tests to ensure that the variables of concern are stationary. A Levin-Lin-
Chu panel unit root test rejects the null hypothesis of a unit root for the sales variable (p⳱0.0188)
106 / The Energy Journal

log kWHit –log kWHit⳱(DSMit – DSMi)dⳭ(X⬘it – X⬘i)g

Ⳮ(log kWHi,t – 1 –log kWHi, – 1)kⳭ(lit – li) (8)

In (8), the overbar denotes a time-average. Both differencing and demeaning

eliminate xi and should eliminate this source of bias in model coefficients.5
It is possible that the idiosyncratic error term is serially correlated, such
that E(lit lis)⬆0 ∀ t⬆s. We therefore estimate the model using generalized least
squares, allowing for a first order autoregressive error structure, in addition to the
ordinary least squares approach described above.
The presence of the lagged dependent variable introduces additional
complication into the model estimation. As Nickell [1981] shows, in the presence
of a lagged dependent variable, the demeaning strategy generates correlation be-
tween the lagged variable and the error term, which results in biased coefficients.
Several strategies have emerged to deal with this problem, mostly involving in-
strumenting the lagged dependent variable with additional lags that are truly ex-
ogenous [Arellano and Bond, 1991, Blundell and Bond, 1998]. However, these
general method of moments strategies rely on a large number of cross-sectional
observations, and so are inappropriate in the current context where the number
of cross sectional units is small and fixed. Instead, Kiviet [1995] derives a ‘cor-
rection’ for the bias present in the demeaned model and adjusts coefficient esti-
mates with this correction. Monte Carlo evidence suggests that this correction is
appropriate for the dimensions of data in this study [Judson and Owen, 1999].
As a result, we report coefficient estimates corrected for the bias resulting from
the presence of lagged dependent variables.
We include key variables that are expected to influence electricity sales
in Xit, with all variables in log form:

1. the number of heating degree days, measured from an 18oC base

2. the number of cooling degree days, measured from an 18oC base
3. the residential retail electricity price, in 2005 Canadian dollars per
kWh
4. gross provincial product per capita, in 2005 Canadian dollars
5. the price of the closest substitute energy for electricity (natural gas in
all but Atlantic provinces; heating oil in Atlantic provinces, where
natural gas is unavailable) in dollars per GJ
6. the percentage of total end-use electricity consumption by the resi-
dential sector

and for the DSM expenditure variable (p⳱0.0020). Thus we do not transform our variables using
differences.
5. Both first differencing and demeaning should eliminate the problem of correlated errors; the
relative desirability of the two depends on error structure. While the results in the paper are for the
model estimated using demeaning, we also estimated the model in differences; our main conclusions
are qualitatively unchanged by the estimation strategy.
 Electric Utility Demand Side Management in Canada / 107

Finally, we note that the price variables in the model are not strictly
exogenous, since the price and quantity of electricity consumption may be co-
determined. We tried to find appropriate instruments for the electricity price in
the model, including the cost of fossil fuels and capital inputs. However, most
electricity in Canada is produced from hydro and nuclear sources, so fossil fuel
prices are very poorly correlated with electricity prices, and we could not find an
appropriate capital cost series that was adequately disaggregated. Consequently,
we use electricity prices directly in the model, with the caveat that inferences on
coefficients may overlook the interdependence of electricity price and quantity.
We think that this problem is less severe than in other markets, however, because
electricity prices throughout Canada are generally regulated, with prices set in
the year or years prior to the actual year that consumption takes place. To the
degree that regulators are not forward-looking, electricity price can therefore be
considered relatively exogenous. Paul et al. [2009] use a similar identifying as-
sumption in their recent study of electricity demand in the US.
A similar problem may exist for the variable of interest—demand side
management expenditures. In particular, it is possible that demand side manage-
ment expenditures are influenced by electricity sales, perhaps because jurisdic-
tions with robust growth in electricity are more likely to implement DSM pro-
grams. If this is the case, estimation of the current model will result in biased
coefficient estimates. We discuss the implications of such endogeneity further
below.

5. DATA SOURCES

Although Statistics Canada requires regular reporting by utilities on elec-

tricity generation capacity, electricity sales and revenues, and other information,
it does not make this data public, except as provincial aggregates. Additionally,
neither Statistics Canada nor any other federal department maintains data on DSM
expenditures or energy savings from these programs. As a result, we conduct our
analysis at a provincial level, and like Parfomak and Lave [1996], we gathered
our data on DSM effort from individual electric utilities.
We gathered data on electric utility DSM spending between 1990 and
2005 from electricity distribution utilities in each Canadian province.6 Data were
gathered mostly from utility, utility regulator, and government publications and
were supplemented with a structured spreadsheet survey and unstructured inter-
views of electric utility staff. For most utilities, data were not available to show
the breakdown of historical DSM spending between sectors or by activity type.
Additionally, we were not able to collect data on utility estimates of energy sav-
ings from DSM programs, but only on DSM expenditures. Finally, we were not

6. In some cases, a province is served by more than on distribution utility. In these cases, we
gathered data from each of the utilities individually or from sources that tabulated DSM activity for
all utilities in a province (this was the case in Ontario).
108 / The Energy Journal

always able to distinguish between DSM spending aimed at load shifting and
DSM spending aimed at energy efficiency, but where such disaggregation was
available, the majority of expenditures (about 75 percent) were directed towards
energy efficiency programs. Thus, we include all demand side management ex-
penditures in our DSM variable. We aggregate the data on DSM spending to a
provincial level to match the dimensions of the rest of the data in our set.
Other data are sourced from Statistics Canada and Natural Resources
Canada. In particular, the volume of electricity sales is derived from Statistics
Canada Table 127-0001, and is given by ‘Total available’ electricity minus ‘Total
industrial generation’. Sales are broken down by end user based on data from
Statistics Canada Tables 128-0002 and 128-0009. Population is from Table 051-
0001, gross domestic product is from Table 384-0001, heating and cooling degree
days are by request from Natural Resources Canada. Natural gas and heating oil
prices are from Tables 129-0003 and 326-0009, respectively, and electricity price
is by request from Natural Resources Canada.
In total, our data set forms a balanced panel consisting of 160 observa-
tions from 10 provinces over 16 years. A summary of the data is provided in
Figure 2 and Table 1. Several trends are obvious in the data. First, in most prov-
inces, electricity consumption is increasing gradually throughout the 16-year pe-
riod of observation.7 However, in most provinces electricity consumption has
fallen for at least part of the period under observation. Second, prices in most
provinces are very stable over time since utility prices were regulated, suggesting
that it will probably be difficult to obtain precise estimates of price elasticity.
Third, most utilities have engaged in demand side management spending during
the sample period (Alberta was the only province with no DSM spending at all
during the period), and there is substantial intertemporal variation in DSM spend-
ing in most provinces, suggesting that if DSM has an effect on electricity sales,
it should be possible to identify it in the data.

6. RESULTS

We begin our analysis by testing whether model coefficients for different

provinces can be pooled together (i.e., is (7) appropriate, or is it necessary to
work with (6)?). An F-test on the joint coefficient restrictions implied by (7)
suggests that these are appropriate, so we work with the pooled models (with
fixed effects) only in this section.8
The first column of Table 2 reports coefficient estimates for the fixed
effect (8) model, with standard errors robust to heteroscedasticity. Most coeffi-
cients are generally of the expected sign, and all except the coefficient on demand
side management expenditures and on cooling degree days are significant at stan-
dard levels. The results suggest a short run price elasticity of –0.06 and a long

7. Although, as discussed earlier, it is stationary.

8. F⳱0.5563, df1⳱85, df2⳱50, p⳱0.991.
 Electric Utility Demand Side Management in Canada / 109

Figure 2: Summary of Demand Side Management Data, Canadian

Provinces, 1990–2005

Per capita electricity sales in each province are shown as an index with an average value of unity.
Per capita demand side management expenditures are shown as a provincial index that takes on a
value of unity in the year of maximum provincial DSM expenditure.

Table 1: Summary of Data

Variable Mean Std. Dev. Min. Max
Sales per capita (GWh) 0.015 0.005 0.006 0.025
DSM expenditure per capita (2005$) 3.51 5.57 0 33.13
GDP per capita (2005$ millions) 0.031 0.007 0.019 0.061
Electricity price (2005c/kWh) 9.37 1.87 6.09 14.56
Price of electricity substitute (2005c/kWh) 3.20 1.43 0.81 7.21
Heating degree days 4615 852 2627 6773
Cooling degree days 123 99 10 493
Fraction of sales to residential/agricultural 0.328 0.073 0.163 0.474

run price elasticity of –0.2. These estimates are fairly consistent with other recent
work. For example, in a recent study of US electricity demand Alberini and
Filippini [2010] find a price elasticity of –0.09 to –0.15 in the short run and
110 / The Energy Journal

Table 2: Results of Regression Analysis

LSDV LSDVAR1 LSDVc
Lagged log of per capita sales 0.7005*** 0.6329*** 0.7722***
(–0.0949) (–0.0622) (–0.0631)
Log of GDP per capita 0.1362** 0.1433*** 0.1110***
(–0.0441) (–0.0411) (–0.0405)
Log of electricity price –0.0637* –0.0820* –0.0664
(–0.0306) (–0.0444) (–0.0455)
Log of substitute energy price –0.0358** –0.0349** –0.0345**
(–0.0144) (–0.0145) (–0.0144)
Log of heating degree days 0.1128* 0.1245** 0.1115**
(–0.0594) (–0.048) (–0.0509)
Log of cooling degree days 0.0031 0.0061 0.003
(–0.0045) (–0.0069) (–0.0059)
Log of percent sales to residential sector –0.0864* –0.0969*** –0.0705**
(–0.0409) (–0.0301) (–0.0287)
Per capita demand side management 0.0001 0.0002 0
(–0.0003) (–0.0007) (–0.0007)
N 150 140 150

* p⬍0.1; ** p⬍0.05; *** p⬍0.01

–0.43 to –0.73 in the long-run. In a meta-analysis, Espey and Espey [2004]

report a mean short run price elasticity of –0.35 and a mean long run elasticity
of –0.85, but note that there is significant variance in both of these estimates.
Our estimated income elasticity is likewise somewhat below published results, at
0.14 in the short run and 0.47 in the long run. For example, in the same meta-
analysis, Espey and Espey [2004] report a mean short-run income elasticity of
0.28 and a long-run elasticity of 0.97. Counter intuitively, the sign on the substi-
tute energy price is negative and significant, implying that increases in the price
of substitute fuels are associated with reductions in the consumption of electricity.
This contradicts other studies, which typically find that electricity and other heat-
ing fuels are substitutes rather than complements [Serletis et al., 2010].9 More
heating degree days are associated with larger electricity sales in the model, which

9. This finding is potentially caused by mis-specification of the model. If electricity and other
fuels are actually substitutes, an upward shock to fuel demand would increase fuel price (provided
fuel supply is not perfectly elastic) and reduce electricity demand (since fuel and electricity are
substitutes). However, our single-stage model will report a negative correlation between fuel price
and electricity demand, suggesting the two are complements.
 Electric Utility Demand Side Management in Canada / 111

makes sense: a one percent increase in the number of heating degree days is
associated with about a 0.1 percent increase in electricity sales. In contrast, the
coefficient on cooling degree days is not significant (although it is of the correct
sign). Finally, the coefficient on the percent of sales going to the residential sector
is significant and of the correct sign, and suggests that as the residential sector
makes up a larger share of total sales, the utility’s sales fall.
The coefficient on the demand side management variable is close to zero
and statistically insignificant. Despite the fact that the variable is not precisely
estimated, it is possible to estimate confidence bounds on the effectiveness of
DSM over the period analyzed. In the short run, the standard error on the coef-
ficient suggests a 95 percent confidence bound of [–0.00054, 0.00079]. This
implies that a $6 per capita DSM expenditure, which was the sample mean during
1990–2005, is estimated to increase electricity consumption in the short run by
0.00075 percent (with a 95 percent confidence interval of [–0.0032,0.0047]). In
the long run, the same expenditure is estimated to increase electricity consumption
by 0.0025 percent (with a 95 percent confidence interval of [–0.018,0.023]).10
It is useful to consider the economic implications of these findings. By
weighting the estimated long-run (cumulative) percentage savings resulting from
demand side management expenditures by the quantity of electricity sales at each
utility in each year, we obtain an estimate of the total reduction in electricity sales
resulting from DSM.8 When this estimate is divided by the total expenditures on
demand side management, we obtain an estimate of the cost of reducing electricity
demand through DSM expenditures.11 Because the parameter on DSM is not
estimated precisely in the model, we apply these calculations to the 95 percent
confidence bounds suggested by the estimate, which allow us to identify a range
of DSM savings and cost effectiveness that are consistent with our data and model.
Based on these calculations, we find that DSM expenditures from 1990 to 2005
have resulted in changes in electricity sales between –1186 GWh (a saving) and
Ⳮ1519 GWh (an increase). Both of these are trivial in size compared to total
Canadian domestic utility sales of over 500 TWh (roughly one quarter of a percent
of total Canadian electricity consumption). The lower 95 percent confidence in-
terval implies a cost effectiveness of reducing electricity demand through demand
side management of over $2/kWh, much higher than other sources of supply (we
do not calculate a cost-effectiveness for the upper end of the confidence interval,
since this estimate suggests that DSM expenditures are associated with electricity
demand increases, rendering a cost-effectiveness calculation meaningless).
The second two columns of Table 2 implement alternative estimation
strategies described earlier, designed to correct for serial correlation in the model

10. The confidence interval of the long-run response to demand side management expenditure is
the ratio of two random variables, and so was determined by bootstrapping with 1000 replications.
11. This estimate is based on the weighting scheme proposed by Auffhammer et al. [2008] and
implemented by Arimura et al. [2009]. Formally, cumulative per capita savings in kWh S are cal-
d
culated for each utility in each year based on model parameters: Sit⳱ kWHitDSMit.
1–k
112 / The Energy Journal

residuals and to deal with bias introduced as a result of the lagged dependent
variable. Accounting for serial correlation using a first-order autoregressive error
term does not substantially change parameter estimates (see the second column
of the table). The coefficient for demand side management is actually reduced in
absolute value in this approach, and the confidence interval is somewhat widened.
These changes do not affect the conclusions drawn above. Correcting for the bias
caused by inclusion of the lagged dependent variable using the method proposed
by Kiviet [1995] does affect parameter estimates somewhat. In particular, the
coefficient on the lagged dependent variable increases substantially, implying that
electricity consumption is more persistent than suggested in previous models.12
The long-run price and income elasticities in this new model are 0.29 and 0.48
respectively, somewhat higher than in the uncorrected models, and more consis-
tent with the evidence from other studies described above. However, the coeffi-
cient on demand side management expenditures remains statistically insignificant
and economically small (even when evaluated at the bounds of the confidence
interval), again suggesting that demand side management expenditures have not
played a substantial role in influencing electricity demand in Canada and that the
cost of reducing demand through demand side management programs has been
high.13
In sum, all of the models that were estimated suggest that demand side
management expenditures have had minimal impact on electricity demand. Our
best estimate, from our preferred model, is that demand side management expen-
ditures in Canada, which averaged roughly $6 per person per year across all
utilities during our sample timeframe, have resulted in an increase of electricity
demand of about one hundredth of a percentage point, an economically very small
amount. As a result, we estimate that reductions of electricity consumption using
demand side management have cost significantly more than estimated by utilities.
Although coefficients on DSM expenditure are estimated imprecisely, standard
confidence bounds around the estimates do not include low-cost estimates of
demand side management that are estimated by utilities themselves. Our data do
not allow us to suggest why this is the case, but we speculate that a combination
of underestimating free-ridership rates and neglecting to account for rebound
effects are the likely reasons. In particular, the estimates from our model suggest
an upper bound for the direct rebound effect of about 30 percent, and other studies
cited in this paper have shown that free ridership rates are often 50 percent or
higher. The combination of these two effects could worsen the cost-effectiveness
of a utility DSM program to the degree that we have reported here.

12. To determine cost effectiveness, total expenditures on DSM across the sample are divided by
兺兺 i t DSM itPOP it
the sum of total savings: CE⳱ .
兺兺 i t S it POP it
13. As briefly noted above, we tested a specification of the model using logged DSM spending
as the explanatory variable, and also tested a differencing strategy. Neither of these variations pro-
duced meaningful differences in the coefficient on the demand side management variable.
 Electric Utility Demand Side Management in Canada / 113

Several potential problems may contaminate our findings. First, although

demand side management expenditures are treated as an exogenous variable here,
it is possible that they could be endogenous. A possibility is that jurisdictions
with high electricity sales are more likely to implement DSM programs, or else
that in years when electricity sales are high, DSM programs are more likely to
be implemented. If either of these is the case, our estimates of the impact of DSM
are likely to be biased towards zero. In theory it is possible to find an instrument
to account for this source of endogeneity, but in practice, especially given our
small sample, we could not find an appropriate one. However, we feel that it is
unlikely that DSM is endogenous in our formulation. DSM effort in Canada has
been driven especially by the regulatory environment and by personalities, and
anecdotal evidence as well as our experience in the industry suggests that causality
in the relationship runs from DSM expenditures to electricity sales, and not the
reverse.
A second problem that our study could suffer from is that it does not
account for regulations, standards, and policies that affect electricity sales other
than utility demand side management. For example, both provincial and federal
governments offer tax credits to encourage energy efficiency, as well as implement
standards and regulations that affect energy sales. Local and provincial govern-
ments also regulate land use and building codes, each of which can affect energy
consumption. If changes in these variables are correlated with demand side man-
agement expenditures, then omitting them from the model will cause the coeffi-
cient on the DSM variable to be biased. In particular, if other policies are imple-
mented simultaneously with DSM expenditures, then the (absolute value of) the
DSM coefficient will be biased upwards. On the other hand, if government reg-
ulations are used as a substitute for DSM expenditures, such that they increase in
intensity when DSM intensity is reduced, then the DSM coefficient will be biased
towards zero. We have no reason to suspect either case. The federal government
regulates appliances and industrial equipment on a regular cycle in Canada, with
new regulations issued every year and existing regulations updated every several
years.
Third, we note that our data on demand side management expenditures
include all demand side management—in particular it includes both load man-
agement expenditures as well as energy efficiency expenditures. Since load man-
agement expenditures are not aimed at curtailing electricity demand explicitly,
including these could lead us to suggest that demand side management is more
costly than is really the case. However, in utilities that were able to provide us
with data (as well as in US utilities), load management expenditures amounted
to less than 25 percent of the total, so error in our estimates should not be too
severe, and in particular should not change the nature of our conclusions.

7. CONCLUSIONS

The estimates from our analysis indicate that in aggregate DSM expen-
ditures by Canadian electric utilities have had only a marginal effect on electricity
114 / The Energy Journal

sales. Thus, our mean estimates for the cost effectiveness of DSM spending, for
the period covered by our study, suggest that the costs for reducing overall elec-
tricity demand through DSM subsidies are high in comparison to the values es-
timated by utilities themselves. Additionally, although the coefficient on DSM
spending in our estimated models is not estimated precisely, at conventional levels
of statistical significance we are able to reject the DSM cost effectiveness values
that have been estimated by the utilities.
The method we use, which because of its aggregate nature directly ac-
counts for the net effect of free ridership, rebound effect, and within-jurisdiction
spillover, provides a useful comparison to utility estimates of DSM effectiveness.
However, because of the aggregate nature of our approach, we are not able to
determine if the cost effectiveness of DSM has changed over time, or whether
DSM expenditures in some utilities or sectors are more effective than those in
others. Indeed, our data set almost surely includes some DSM programs that are
very cost effective, as well as others which are not. Our analysis is only able to
estimate the aggregate effectiveness and cost effectiveness of all DSM programs
in Canada over a 16-year period.
It is important to emphasize that the cost effectiveness estimates gen-
erated by this research do not indicate the full technico-economic cost of im-
proving energy efficiency, but rather reveal the cost of pursuing energy efficiency
by a utility, which is typically conducted with subsidies. Subsidies are less effec-
tive than often assumed because they cannot avoid distributing the subsidy to
those who would have undertaken the energy efficiency measure in the absence
of the subsidy. These free riders weaken the effectiveness of the subsidy. Addi-
tionally, the rebound effect, which occurs as consumers adjust purchases in re-
sponse to changes in relative prices, further erodes the effectiveness of an energy
efficiency subsidy. Our analysis is suggestive that free ridership and rebound may
have been substantial during the period of electric utility DSM subsidies covered
by this study. This suggests that policy initiatives that rely on subsidies to promote
energy efficiency may be much less effective than their supporters claim. This
holds equally for government subsidy programs and for private subsidy programs,
such as offset programs currently associated with efforts to reduce GHG emissions
via subsidy.
The results should not be taken, however, to imply that investments in
energy efficiency are undesirable from a social perspective. In estimating the cost-
effectiveness of DSM programs, electric utilities are confined to a narrow defi-
nition of cost effectiveness that typically excludes environmental and social ex-
ternalities. If these are taken into account, there is considerable research indicating
that energy efficiency remains cost effective relative to new conventional gener-
ation from a social perspective. The challenge for utilities, governments and con-
cerned individuals is to find policies that will actually ensure greater energy ef-
ficiency than that which naturally occurs as the capital stock evolves.
 Electric Utility Demand Side Management in Canada / 115

ACKNOWLEDGMENTS

Nic Rivers acknowledges funding from the Trudeau Foundation. Jonn

Axsen provided excellent research assistance.

REFERENCES

Alberini, Anna and Massimo Filippini (2010). Response of residential electricity demand to price:
The effect of measurement error. Working paper 75, Centre for Energy Policy and Economics,
Swiss Federal Institutes of Technology.
Arellano, M. and S. Bond (1991). “Some tests of specification for panel data: Monte carlo evidence
and an application to employment equations.” The Review of Economic Studies, 58(2):277–297.
Arimura, T.H., R.G. Newell and K.L. Palmer (2009). Cost-effectiveness of electricity energy effi-
ciency programs. Discussion Paper 09-48, Resources for the Future.
Auffhammer, Maximilian, Carl Blumstein, and Meredith Fowlie (2008). “Demand-side management
and energy efficiency revisited.” The Energy Journal, 29(3):91–104.
Blundell, R. and S. Bond (1998). “Initial conditions and moment restrictions in dynamic panel data
models.” Journal of Econometrics, 87(1):115–143.
Espey, James A. and Molly Espey (2004). “Turning on the lights: A meta-analysis of residential
electricity demand elasticities.” Journal of Agricultural and Applied Economics, 36(1):65–81.
Eto, Joseph, Suzie Kito, Leslie Shown and Richard Sonnenblick (1995). Where Did the Money Go?:
The Cost and Performance of the Largest Commercial Sector DSM Programs. Lawrence Berkeley
Laboratory.
Geller, Howard and Sophie Attali (2005). The experience with energy efficiency policies in IEA
countries: Learning from the critics. Technical report, International Energy Agency.
Gellings, Clark W. (1996). “Then and now : The perspective of the man who coined the term ‘DSM’.”
Energy Policy, 24(4):285–288.
Gillingham, Kenneth, Richard Newell and Karen Palmer (2006). “Energy efficiency policies: A ret-
rospective examination.” Annual Review of Environment and Resources, 31(1):161–192.
Greening, Lorna A., David L. Greene, and Carmen Difiglio (2000). “Energy efficiency and con-
sumption—the rebound effect—a survey.” Energy Policy, 28(6–7):389–401.
Grosche, Peter and Colin Vance (2009). “Willingness to pay for energy conservation and free-ridership
on subsidization: Evidence from Germany.” The Energy Journal, 30(2):135–153.
Hartman, R.S. (1988). “Self-selection bias in the evolution of voluntary energy conservation pro-
grams.” The Review of Economics and Statistics, 70(3):448–458.
Horowitz, Marvin J. (2004). “Electricity intensity in the commercial sector: Market and public pro-
gram effects.” The Energy Journal, 25(2):115–137.
Horowitz, Marvin J. (2007). “Changes in electricity demand in the united states from the 1970s to
2003.” The Energy Journal, 28(3):93–119.
Houthakker, H.S., P.K. Verleger and D.P. Sheehan (1974). “Dynamic demand analyses for gasoline
and residential electricity.” American Journal of Agricultural Economics, 56(2):412.
Jaccard, M. (1993). “Regulation of energy utilities in Canada: Where do we go from here?” Energy
Studies Review, 4(3):5.
Judson, R.A. and A.L. Owen (1999). “Estimating dynamic panel data models: a guide for macroe-
conomists.” Economics Letters, 65(1):9–15.
Kiviet, J.F. (1995). “On bias, inconsistency, and efficiency of various estimators in dynamic panel
data models.” Journal of Econometrics, 68(1):53–78.
Loughran, David S. and Jonathan Kulick (2004). “Demand-side management and energy efficiency
in the United States.” The Energy Journal, 25(1):19–43.
Louviere, Jordan J., D. A. Henscher, and Jeffrey D. Swait (2000). Stated Choice Methods. Cambridge
University Press, Cambridge, UK.
116 / The Energy Journal

Malm, Eric (1996). “An actions-based estimate of the free rider fraction in electric utility DSM
programs.” The Energy Journal, 17(3):41.
Nadel, Steven and Howard Geller (1996). “Utility DSM : What have we learned? Where are we
going?” Energy Policy, 24(4):289–302.
Natural Resources Canada (2006). Energy use data handbook: 1990 and 1998 to 2004. Technical
report, Canada.
Nickell, S. (1981). “Biases in dynamic models with fixed effects.” Econometrica: Journal of the
Econometric Society, 49(6):1417–1426.
Parfomak, Paul W. and Lester B. Lave (1996). “How many kilowatts are in a negawatt? Verifying ex
post estimates of utility conservation impacts at a regional level.” The Energy Journal, 17(4):59.
Paul, A.C., E.C. Myers, and K.L. Palmer (2009). A partial adjustment model of US electricity demand
by region, season, and sector. Discussion Paper DP 08-50, Resources for the Future.
Sampson Research (2005). Market and impact evaluation of BC Hydro’s seasonal LED initiative.
Consulting report, BC Hydro.
Saunders, Harry D. (1992). “The Khazzoom-Brookes postulate and neoclassical growth.” The Energy
Journal, 13(4):131–148.
Serletis, A., G. Timilsina, and O. Vasetsky (2010). International evidence on sectoral interfuel sub-
stitution. Working Papers.
Sorrell, S. and J. Dimitropoulos (2007). Technical report 2: Econometric studies, UKERC review of
evidence for the rebound effect. Working paper, UK Energy Research Centre.
Sorrell, Steve (2007). The rebound effect: An assessment of the evidence for economy-wide energy
savings from improved energy efficiency. Technical report, UK Energy Research Centre.
Train, Kenneth E. and Terry Atherton (1995). “Rebates, loans, and customers’ choice of appliance
efficiency level: Combining stated- and revealed-preferences.” The Energy Journal, 16(1):55.