NCEE#
NATIONAL CENTER FOR
ENVIRONMENTAL ECONOMICS
Technology Diffusion and Environmental Regulation: The
Adoption of Scrubbers by Coal-Fired Power Plants
Elaine Frey
Working Paper Series
Working Paper # 08-04
March, 2008
£
M'i
1 *
U.S. Environmental Protection Agency
National Center for Environmental Economics
5 " 1200 Pennsylvania Avenue, NW(MC 1809)
Washington, DC 20460
¦h
pro^^ http://www.epa.gov/economics
-------
Technology Diffusion and Environmental Regulation: The
Adoption of Scrubbers by Coal-Fired Power Plants
Elaine Frey
NCEE Working Paper Series
Working Paper # 08-04
March, 2008
DISCLAIMER
The views expressed in this paper are those of the author(s) and do not necessarily represent
those of the U.S. Environmental Protection Agency. In addition, although the research described
in this paper may have been funded entirely or in part by the U.S. Evironmental Protection
Agency, it has not been subjected to the Agency's required peer and policy review. No official
Agency endorsement should be inferred.
-------
Technology Diffusion and Environmental Regulation:
The Adoption of Scrubbers by Coal-Fired Power Plants
Elaine F. Frey1
George Washington University
and
U.S. EPA National Center for Environmental Economics
2115 G Street, NW #308
Washington, DC 20052
(301) 910-5799
effrey@gwu.edu
Keywords: technology diffusion; environmental regulation; electric generating industry
Subject Matter Classification: Pollution Control Options and Economic Incentives; Environmental Policy
Abstract
This research examines the technological diffusion of scrubbers, a sulfur dioxide (S02) abatement
technology, in response to Title IV of the Clean Air Act. Title IV implemented a tradable pollution
permit system for S02, which is radically different from the state emission rate standards that were in
place previously. I find that power plants with strict state regulations (or command-and-control
regulations) and low expected scrubber installation costs have a high probability of installing a scrubber.
These findings suggest that, although Title IV has encouraged diffusion, some scrubbers have been
installed because of state regulatory pressure. Since policies are often evaluated based on the incentives
they provide to promote adoption of new technologies, it is important that policy makers understand the
relationship between technological diffusion and regulatory structure to make informed decisions.
Although tradable permit systems are thought to give firms more flexibility in choosing abatement
technologies, I show that interactions between a permit system and pre-existing command-and-control
regulations can limit that flexibility.
1 This research is an essay from my dissertation and has been funded in part by the U.S EPA, National Center for
Environmental Economics. Opinions expressed in this paper are solely those of the author and do not necessarily
represent those of the U.S. EPA. I would like to thank my dissertation advisor, Aran Malik, dissertation committee
members Frederick Joutz and Wallace Mullin, Ron Shadbegian, Cynthia Morgan and members of the GWU
Microeconomics Seminar for useful comments and suggestions for this paper. I would also like to thank Nathaniel
Keohane for providing data.
-------
I. Introduction
Research that concerns the effects of policy on technology diffusion, especially in the area of
environmental policy, has increased in recent years. Economists' interest in this area spawns from the
fact that the diffusion of new technologies has the potential to change environmental conditions and
because implementing environmental regulation often induces or alters the diffusion path of an abatement
technology. This study will contribute to this body of work by determining the effects of regulatory type
and regulatory stringency on the diffusion of technology. Specifically, I will examine the effect of sulfur
dioxide (S02) emission restrictions on the diffusion of scrubbers by coal-fired power plants.
Scrubbers, or flue gas desulfurization units, are designed to remove S02 from exhaust gases of
electric generating units. Scrubbers are a highly effective abatement technology and can be designed to
remove up to 99 percent of S02 from the emissions stream. Within a power plant, scrubbers are installed
on an individual generating unit. Power plants typically contain several generating units which can vary
significantly in specifications such as age, size, technology usage and fuel type; units within the same
power plant may even face different environmental restrictions. As a result, a given power plant may
have separate compliance strategies assigned for each unit, which may include installing scrubbers on one
or more generating units.
Title IV of the 1990 Clean Air Act Amendments, also known as the Acid Rain Program, presents
a unique regulatory regime which established the first nationwide tradable pollution permit system in the
U.S. This was significantly different from the emission rate standards and abatement technology
requirements that were in place previously. The purpose of Title IV was to mitigate the negative effects
of S02 emissions from electric power plants by cutting emissions by 40 percent from 1980 levels. Title
IV was implemented in two phases. Phase I required the 263 largest and dirtiest generating units to
reduce emissions in a tradable pollution permit system, beginning in 1995. Under Phase I, each unit was
allocated permits sufficient to achieve an emission rate of 2.5 lbs or S02 per million Btu (lbs of
S02/mmBtu), where one permit allows the unit to emit one ton of S02. Permit allocations are based on
each unit's average heat input from 1985 to 1987. Allocated permits can then be used in the allocated
E. Frey
1
-------
year, banked for use in future years or traded with other firms. Phase II began in 2000 and includes all
units larger than 25 MW (megawatts of nameplate capacity2). Under Phase II, each unit is allocated
permits sufficient to emit at a rate of 1.2 lbs of S02/mmBtu. Prior to Title IV, newly constructed
generating units had to follow New Source Performance Standards (NSPS), which established rigid
technology and emission restrictions that required the installation of scrubbers. In addition, existing
sources were subject to state level emission rate standards or command-and-control regulations. State
level emission rates were mandated under State Implementation Plans (SIPs) as part of the National
Ambient Air Quality Standards (NAAQS). NSPS were essentially overridden under Title IV, although
existing sources still have to comply with state regulations.
Before Title IV was enacted, scrubbers were adopted primarily to comply with NSPS. In 1971,
NSPS required a subset of large, new coal-fired units to have emission rates lower than 1.2 lbs of
S02/mmBtu. Since there was no restriction on the compliance method, some scrubbers were installed
during this period but the technology was not widespread. In 1977, NSPS were amended such that all
new units constructed after 1978 essentially had to install scrubbers. As a result, most scrubbers that were
installed after 1978 were mandated by NSPS until Title IV began. Title IV departs from previous
regulations in that the method of pollution abatement is flexible and the amount of abatement for each
unit is not predetermined.
Theory suggests that market-based policies such as Title IV give firms more incentives to adopt
new technologies, relative to command-and-control regulations (see Milliman and Prince, 1989; Jung et
al., 1996; Downing and White, 1986; Fischer et al., 2003; and Keohane, 2006). But because there have
been so few market-based regulations in place in the U.S., there is little empirical evidence to support this
claim. Accordingly, this research will examine the technological diffusion of scrubbers, a S02 abatement
technology, under the market-based regime established under Title IV of the Clean Air Act.
2 Nameplate capacity is the maximum rate output of a generator under specific conditions designated by the
manufacturer.
E. Frey
2
-------
The theoretical model that I utilize in this paper depicts technology diffusion as a dynamic
optimization problem where a scrubber is installed in a given period if the value of the benefit stream
from scrubbing is greater than the cost of installing a scrubber. This decision is dependent upon unit
specific characteristics, regulatory restrictions and scrubber installation costs, which can vary for each
generating unit. Under this framework, since net benefits from scrubber installation can differ across
generating units, there will always be some units at which it is not profitable to install a scrubber. A
similar model is adopted by Karshenas and Stoneman (1993), Kerr and Newell (2003) and Popp (2006),
but my model emphasizes the importance of the rank effect, which assumes that firms have heterogeneous
characteristics that influence the net benefits of adopting a technology.
In the empirical analysis, first I estimate expected installation costs of scrubbers for all generating
units, including those that did not install a scrubber, using unique cost data for scrubbers that were
actually installed between 1972 through 2002. Then, I use a Cox proportional hazards model to
determine how technology diffusion responds to changes in regulatory structure and changes in regulatory
stringency, controlling for the expected cost of installing a scrubber. I also explore a model that controls
for firm-level unobserved heterogeneity. I find that electric generating units with strict state command-
and-control regulations have a high probability of scrubber installation. Moreover, the expected cost of
installing a scrubber is an important factor in this decision; as expected, those units that have very low
expected costs are more likely to adopt. Finally, the age of an individual generating unit is also an
important factor in this decision. My findings suggest that, although Title IV has encouraged diffusion,
some scrubbers have been installed because of state regulatory pressure. Theoretically, tradable permit
systems should give firms more flexibility in choosing abatement technologies, which leads to lower total
abatement costs. However, empirical evidence shows that interactions between a permit system and pre-
existing command-and-control regulations seem to limit that flexibility.
The remainder of this paper is organized as follows: Section II provides a brief overview of
technological diffusion models and empirical evidence, Section III establishes the theoretical and
E. Frey
3
-------
empirical model and discusses the data used in this study, Section IV presents the results of the empirical
estimation and Section V concludes.
II. Literature Review
In this section, I will give an overview of relevant empirical findings that relate to my research.
Throughout the text, I will follow Schumpeter's definition of technology change and technology
diffusion. Schumpeter (1942) describes three stages of technological change: i) invention, the creation of
a new product or process; ii) innovation, advances in technologies and techniques used in production; and
iii) diffusion. I will use the term adoption as the act of installing a particular technology; a single firm
adopts a technology and the combined adoption decisions of firms are referred to as technological
diffusion. Much of the work on diffusion does not explicitly model policy effects, but rather, concerns
itself with explaining why some technologies are diffused more quickly than others and why some firms
will adopt more quickly than others. Four types of effects have been developed to explain these patterns
of technology diffusion overtime: epidemic, order, stock and rank effects. Some empirical research
exists, which tests one or more of these effects on diffusion and a small subset of these studies also
incorporates the effect of policy of the diffusion process. I will first discuss the epidemic, order, stock
and rank effects in more detail and then give an overview of the major empirical findings in this area.
II.A. Epidemic. Order. Stock and Rank Effects
Epidemic, order, stock and rank effects have traditionally been the basis for technology diffusion
models for the purpose of explaining the pattern of diffusion over time.3 These models each follow the
well-known diffusion pattern, the S-curve, which reveals that firms find it optimal to adopt a particular
technology at different times. The first generation of models used the epidemic approach, finding that
technological diffusion is self-propagating and occurs through a process of information exchange among
firms (Griliches, 1957). Under this framework, diffusion is analogous to the spread of an infectious
3 See Stoneman (2001), Hall and Khan (2003) and Geroski (2000) for an in-depth review of the diffusion literature.
In addition, Jaffe et al. (2002) provide an overview of the relationship between technological change and
environmental issues.
E. Frey
4
-------
disease; as the infected (adopters) come into contact with non-infected individuals (non-adopters), they
spread the disease (technology). However, the epidemic model does not explain any underlying
economic behavior, assumes that all firms are homogeneous and provides no theoretical explanation for
why some firms would adopt a technology faster than others.
Models that incorporate stock and order effects assume that a firm's adoption decision is
dependent on the actions and adoption decisions of other firms. For this reason, both types of models are
often termed game-theoretic models. To make the adoption decision, firms compare their profits with
adoption to their profits without adoption, given the decisions of other firms. In this model, the benefits
of adoption are also dependent on production levels of the firm and production levels are dependent on
other firms' adoption behavior. Although profits differ between adopters and non-adopters, firms within
each group have the same profit levels.4 The order model assumes that the present value of adoption
depends on the firm's order of adoption, relative to other firms. Thus, benefits of adoption decrease as
adoption increases and eventually, no firms find it profitable to adopt (barring any other changes in the
quality or cost of the technology). Unlike the stock model, Fudenberg and Tirole (1985) argue that the
profit levels of firms differ such that those firms that adopt early have higher profits than firms that adopt
later. The model implies that there are first-mover advantages to adoption, in that the first to adopt can
influence the outcomes of other rival firms.
Finally, models that integrate rank effects, beginning with the work of David (1969), assume that
potential adopters have heterogeneous characteristics, which affect the benefits they receive from
adoption. The benefits of adoption can be influenced by the firm's location, previous investments and
technologies, management decisions and knowledge, market conditions, the regulatory environment, firm
organizational type, expectations of future market conditions and the firm's discount rate. In this model,
adoption occurs if the net present value of adoption is positive and it is not more profitable to put off the
adoption decision until another period. A diffusion path is created because over time, the benefits or costs
4 See Reinganum (1981) and Stoneman (2001) for more detail on stock models.
E. Frey
5
-------
of adoption change. Like the order model, those firms with the highest net benefits will adopt the
technology first.
Stoneman (2001) suggests that empirical research on technological diffusion aims to prove that
data is consistent with theory and to identify which factors are the most important in the diffusion
process. Despite these objectives, there is no certainty as to which effect (epidemic, stock, order or rank)
describes technology diffusion best. In the majority of empirical studies, the researcher simply chooses
one model and applies the model to a particular diffusion case. In many instances, this is due to limited
availability of data, which precludes the researcher from using a particular model. One exception is
Karshenas and Stoneman (1993), who nest all four effects into one model. They find evidence to support
rank and epidemic effects, but no evidence that the stock or order effects describe the diffusion process in
that particular application.
II.B. Empirical Evidence
Empirical studies often utilize probit or survival models, which contain variables that take into
account one or more of the four effects described above. Variables that have been used to incorporate the
rank effect (i.e. firm heterogeneity) include firm characteristics such as firm size, age, ownership type and
the complexity of a firm's operations. Below I will discuss each of the characteristics that I use in my
analysis and how it is expected to affect technology diffusion.
The size of the unit, which is measured by the unit's nameplate capacity in megawatts (MW), is
expected to have a positive coefficient (note that I use the natural log transformation of this variable).
Karshenas and Stoneman (1993) argue that size should have a net positive effect on the hazard rate
because of three effects on adoption behavior. First, large firms are more likely to adopt because there are
positive scale effects of adoption and can adopt new technology at a lower cost. Also, large firms are
willing to undertake more risk than smaller firms, and are therefore more likely to adopt. In the opposite
direction, larger firms may have more rigid managerial approaches, and may actually be less likely to
adapt to changes, relative to smaller firms. I do not have data on managerial preferences, but in the frailty
models I do control for unobserved firm-level effects, which may include these sorts of managerial
6
-------
predispositions. In the literature, a number of authors find that large firms are more likely to adopt than
smaller firms (Rose and Joskow, 1990; Karshenas and Stoneman, 1993; Popp, 2006; Snyder et al., 2003;
Kerr and Newell, 2003; and Fuglie and Kascak, 2001).
The age variable, measured as of 2002, is expected to have a negative effect on technology
diffusion. Karshenas and Stoneman (1993) hypothesize that newer firms may be more likely to adopt
new technologies than older firms because adoption often requires reorganization, which new firms are
more likely to take on. There is some empirical evidence that firm age has a negative effect on the
instance of adoption (Popp, 2006), but, Karshenas and Stoneman (1993) and Dunne (1994) find that age
is not statistically significant. Rose and Joskow (1990) find that ownership structure affects diffusion
rates; however, results from Karshenas and Stoneman (1993) do not show ownership type to have a
statistically significant effect. Complexity of the firm's operations is also thought to be an important
predictor of adoption but the results are mixed. Snyder et al. (2003) find that complexity has a negative
effect on the probability of adoption, but Popp (2006) and Kerr and Newell (2003) show a positive effect.
I expect that higher installation costs would have a negative effect on the decision to adopt a new
technology, which is intuitive. Empirical evidence has in fact shown that the cost of installing new
technology has a negative effect on technology adoption (Stoneman and Karshenas, 1993; Stoneman and
Kwon, 1994; Kerr and Newell, 2003); but measuring costs accurately is often difficult, due to data
constraints. Previous studies have used firm characteristics, such as firm size, as proxies for the
differences in adoption costs across firms. My research differs in that the expected cost for scrubber
installation will be explicitly estimated.
Stoneman and Karshenas (1993) show that the number of firms that have previously adopted the
technology can be used to measure both stock effects and order effects. Order effects are also captured by
a variable for the expected number of adopters in the next period. Although, as Popp (2006) points out,
including both number of firms that have adopted the technology and regulation variables in the
estimation may lead to endogeneity, since regulatory stringency may directly affect the number of
7
-------
adopters. For this reason, my estimation will exclude the number of previous adopters as an explanatory
variable.
Other empirical work has supported the theoretical finding that market-based policies, such as
Title IV, create more incentives for technological adoption than command-and-control policies. Popp
(2001) uses detailed data on patents of electric generating units that have installed scrubbers since 1985 to
test if Title IV caused innovation and adoption of more efficient technologies. He finds that although the
number of patents did not increase because of the permit system, more efficient scrubbers were installed
because of Title IV, leading to lower abatement costs. Bellas (1998) finds no evidence of technological
advances in scrubbers in new electric generating units under the command-and-control regulations in
place before Title IV was implemented.
Keohane (2005) provides an empirical simulation of abatement choices under different regulatory
schemes. He finds that the type of regulation implemented is an important determinant of abatement
choice. Specifically, he finds evidence that total abatement costs are lower under a permit system because
of the types of technologies; his primary focus is on coal-switching technologies and the installation of
scrubbers. This provides evidence that permit systems grant more incentives for increased technological
adoption than an alternative command-and-control policy. Similarly, Gray and Shadbegian (1998) find
that in the paper industry, state environmental regulations do have an effect on the technology choice of
firms. In particular, strict regulations have a large impact on technology choices of new plants, but
regulatory stringency has little to no effect on existing plants.
Kerr and Newell (2003) provide empirical evidence on the dynamic effects of environmental
regulation on technology diffusion, using data from oil refineries. Their survival model differs from
similar models in that it incorporates the effect of market-based regulation versus command-and-control
regulation on technological adoption. Consistent with theory, they find that increased regulatory
stringency has a positive effect on the adoption of the new technology and that firms are more likely to
adopt the technology under the permit system than under a performance standard.
E. Frey
8
-------
Popp (2006) uses a survival model for technological adoption of two types of nitrous oxide (NOx)
emission controls for electric power plants. His estimation controls for expected quality increases of
scrubbers, using patent data as a proxy for increased future technological improvements. His results show
that regulatory stringency is the most important predictor of technology choice. In addition, he finds that
available knowledge of emerging technologies, from domestic innovations and knowledge from abroad,
has a small but significant effect on adoption decisions.
Finally, Snyder et al. (2003) use a survival model to estimate the effects of environmental
regulations on the diffusion of membrane cell technology, a cleaner technology for the production of
chlorine. They model both the adoption of membrane cell technology at existing plants, as well as the
exit of older plants using a survival model. Dummy variables are used to control for the effects of each
type of regulation on diffusion of new technologies and the exit of older chlorine facilities. They find that
the presence of environmental regulations did not facilitate technological adoption at existing facilities,
but did increase the exit rate of facilities using older technologies.
My research will use the methods developed in the diffusion literature to explore the adoption of
scrubbers by coal-fired electric generating units in response to Title IV of the Clean Air Act. This
research differs from existing research in that it will provide empirical evidence that market-based
policies, particularly pollution permit systems, encourage technological diffusion. Moreover, I will
explicitly estimate the expected installation cost of scrubbers, which I believe is an influential determinant
in the decision to adopt but has been missing from the literature. Because diffusion of abatement
technologies can have a great impact on environmental quality, this research is not only important to the
diffusion field but also contributes to the environmental field.
III. Modeling the Decision to Install a Scrubber
This section will outline the decision to install a scrubber and describe how diffusion will be
modeled in this context. First, I will formally introduce a theoretical model, which provides a foundation
for the focus of this research, the empirical model. I will also discuss how my model compares to other
E. Frey 9
-------
models in the diffusion literature. Furthermore, I will discuss how epidemic, rank, order and stock effects
drive technological diffusion in my model. Specifically, I argue that rank effects will dominate the
decision to adopt in the case of scrubbers. Then, I will describe the empirical model, which uses survival
analysis to model the time to adoption of scrubbers. Finally, I will describe the data used to estimate the
empirical model.
III.A. Theoretical Model
In my model, which follows from Kerr and Newell (2003), each plant manager, has the option
to choose T, the adoption time of the scrubber. It is possible that firms with high expected installation
costs (or low benefits) will never find it optimal to adopt. The expected cost of installing the technology,
at installation time T for each /' is CiT . Zit is a vector of unit-specific characteristics that influence the
cost of installing a scrubber. In every period t. the manager faces regulatory restrictions Rit and decides to
install a scrubber or to delay the decision to install until next period. Note that regulations Rit represent
the unique regulatory requirements that were established by the State Implementation Plans (SIP), which
vary across units and can vary across time for a single unit.5 This vector of regulatory variables also
reflects whether the unit was regulated under Phase I or Phase II of Title IV. This decision can be
represented in the following dynamic optimization problem,
T co
max|(s"(Z„Ji(„0>-"A-C,re-" +\B'{Z„,R„ty"dt (1)
0 T
where r is the unit's discount rate. Each manager is choosing the time of adoption, T, which will
maximize its total net present value of returns.6 The first term on the RHS captures z"s present value of
benefits before the technology adoption, the second term represents the costs of adoption at time T, while
the third term is the present value of /'s benefit stream, after the adoption takes place.
5 SIPs are established by each state in order to comply with National Ambient Air Quality Standards (NAAQS) and
are usually imposed in the form of standard emission rates. These emission rates were set anywhere between 0.006
and 12.6 lbs S02/mmBtu. It is not uncommon for these emission rate standards to change over time for a given unit,
as states try to adjust for any noncompliance areas or other local circumstances that may arise.
6The time of adoption, T. varies for each plant manager, Thus, the precise notation would include a subscript for /'
on T. To simplify notation, I have omitted this subscript, but it is implied throughout the text.
E. Frey 10
-------
Adoption is dependent on two conditions. First, adoption will occur at T as long as it is not more
profitable to wait and adopt at a later period. This is commonly known as the arbitrage condition and is
given by the derivative of Equation (1) with respect to T.7 Second, adopting the technology must result in
positive profits (the second order condition).8 The arbitrage condition is a sufficient condition if the
installation cost is nondecreasing and convex and the gross value of adoption is nondecreasing with
respect to time. In other words, the second derivative of Equation (1) must be greater than or equal to
zero.
Equation (1) captures the rank effect that was described in the literature review. Generating units
are heterogeneous and therefore have different benefits and costs of technology adoption. The plant
manager will choose to adopt at time t if the net present value of adoption is positive and it is not more
profitable to postpone adoption. This simple model suggests that as the costs of adoption increase, z"s
benefit from adoption will decrease. Therefore, for any adoption cost, there will always be some units at
which it is not profitable to adopt at any given time, due to heterogeneous benefits across firms. My
model indicates that technological diffusion is affected by unit and plant level characteristics, which are
represented by the vector Zit, regulatory structure and stringency encapsulated in Rit and the expected cost
of the technology, CjT . The empirical model will provide a method to test these assertions.
Scrubbers typically have a lifespan of 15 years (Daniel Mussatti, U.S. Nuclear Regulatory
Commission, personal communication, May 16, 2007). Although, my data show that they often operate
longer than that, there was a 34 year-old scrubber that was still operating in 2003 and 50 percent of the
scrubbers in the sample were older than 20 years. However, in the model, I assume that the plant
manager faces an infinite time horizon, rather than a finite time horizon. In other words, the technology
adopted will have an infinite physical life and the technology does not become obsolete. Given a normal
7 The first order condition is: Ib1 (Z jT ,RjT ,T) - B° (Z iT ,RiT ,T)\- rC iT + ^iT > 0 • The expression in brackets
,T' ,T' ,T dT
can be interpreted as the gross value of the adopted technology at time T. while the second two expressions represent
the costs incurred from adoption at time T
8 The second order condition is:. ][B\z it ,Rit ,t) - B\Z it ,Rit dt -CiTe-T > 0
E. Frey
11
-------
scrubber lifespan, an infinite time horizon is a sensible approximation, given a reasonable value of the
discount rate. This also allows simplification of the problem without changing the implications of the
model, according to Stoneman (2001). This assumption is made throughout the literature, with the
exception of Ireland and Stoneman (1986), who explicitly examine technological diffusion by introducing
the probability that obsolescence can occur in any given time period. In their general equilibrium model,
they find that increasing firm expectations of the obsolescence of the technology will diminish the extent
of technology diffusion in an industry under certain market conditions.
As described in the previous section, epidemic effects depict the diffusion process as an
automated process where firms adopt as a result of coming into contact with firms that have already
adopted. This type of process may be justified if a technology is not known to all possible adopters; when
potential adopters meet adopters, information about the technology is disseminated and technology is
adopted by those who have acquired the new information. However, in the case of electric power plants,
scrubbers are a known abatement technology to all possible users. Because of the NSPS that began in
1978, all new units were required to install scrubbers. By 1990 when Title IV was passed, over 50
scrubbers had already been installed in the industry, ensuring that all potential adopters were well aware
of the technology. Therefore, my model does not incorporate epidemic effects.9
Although I do not include stock and order effects in my model, they can easily be incorporated
into the theoretical model presented above, following Stoneman and Karshenas (1993), by adding two
arguments that represent the stock and order effects. The first. S('I'), reflects the order effects, where for
a given firm benefits of adoption are dependent on the number of firms that have already adopted at z"s
adoption date. The second variable, K(t), reflects stock effects and captures the fact that the returns to a
firm will depend on the number of other users at that date t. To incorporate these effects in the model, the
9 Although epidemic effects are not incorporated in the theoretical model, Hannan and McDowell (1987) argue that
empirically, epidemic effects can be incorporated by using a logistic baseline hazard function or a hazard rate that is
increasing with duration. As 1 will discuss in the empirical results, 1 do not find that this type of model is
appropriate for the data.
E. Frey
12
-------
two variables would be arguments in the benefits before and after adoption or B° (•) and B1 (•),
respectively.10
The cost of adopting a new technology is thought to be a driving factor in a unit's technology
choice, but is not measured in much of the diffusion research. The cost estimates derived will be
controlled for directly in the scrubber choice econometric model. I will explicitly measure the expected
cost of scrubber installation, which is the relevant cost measure for making adoption decisions in the
diffusion literature (Stoneman, 2001).11 This is distinct from abatement costs, defined as the loss of
profits from reducing emissions, which are commonly used in the environmental literature. The
installation cost estimate, Cit (•), will simply be estimated by a log-linear function, which will depend on a
vector Z((, /' s unique characteristics. This vector will include factors such as capacity of the unit, quality
of coal used by the generating unit, the removal efficiency of the installed scrubber, the age of the
scrubber and regional variables, which are thought to be important determinants of cost (Bellas, 1998;
Keohane 2005). Cost data are only available for units that have already adopted the technology. But,
using the coefficients measured in the cost equation above, I derive expected installation costs for all
units, even those that did not install the technology. Due to space limitations, I omit the results for the
installation cost equation, but results are available upon request.12
My model is distinct from other models in the use of the installation cost of scrubbers,
represented in Equation (1) by Cit (ZiT,/). I assume that the installation cost depends upon time and on
10 Thus the model that incorporates epidemic, rank, stock and order effects becomes:
T co
V it = max J {B " (Z it , R it , K t , S T dt - C tT e-'T + J (b 1 (Z it , R it , K t , S T ,/)); "" dt '
T 0 T
11 According to Stoneman (2001), the cost of installation is defined as "the price that has to be paid to acquire new
capital goods that embody the new technology", this is also termed the acquisition cost. I have chosen to follow the
diffusion literature and use installation costs in the estimation of scrubber choice. Therefore, I exclude yearly
operation and management costs in this analysis. Moreover, although operation and maintenance costs are reported,
they do not seem very reliable and many observations are missing from the data. A prediction of these yearly costs
would be somewhat suspect, due to the questionable reliability of actual costs.
12 There is potential for selection bias when predicting installation costs, since the scrubber sample is comprised of
units that may have had lower installation costs than those who did not scrub and are thus not in the sample.
However, I will control for this using a frailty model, which allows me to adjust for those unobservable
characteristics of plants that makes them have lower installation costs and thus more likely to scrub.
E. Frey
13
-------
the individual characteristics of the generating unit, which follows from Kerr and Newell (2003). Some
studies (Stoneman, 2001; Popp, 2006) simplify the role of installation cost by allowing it to vary over
time, but not across firms. Some studies simply omit the effect of cost on diffusion in the model (Rose
and Joskow, 1990; Snyder et al. (2003); Fuglie and Kascak (2001).13 Cost data have often been
unavailable to researchers in the past, which may explain this simplification. In my study, unique cost
data on scrubbers is available, which allows for a detailed look of the effect of installation cost on
diffusion.
Stoneman and Karshenas (1993) point out that there is, in general, a missing link between
empirical and theoretical work in technological diffusion. Many empirical diffusion studies do not
specify theoretical models, for example Rose and Joskow (1990) and Snyder et. al (2003). However, the
basic deterministic model depicted in Equation (1) has been the primary analytical model for those that do
provide theoretical motivations for diffusion. The model that I present emphasizes the importance of the
rank effect of diffusion. The rank effect allows for heterogeneity among potential adopters, so that firms
will get different net benefits from the technology adoption. This explains the pattern of diffusion; not all
firms will adopt at the same time, those with low net benefits may never find it profitable to adopt.
Although the stock and order effects have found to be determinants of technological diffusion, there are
several reasons that I believe these effects are not significant in the case of scrubbers. As explained by
Popp (2006), stock and order effects assume that the decision to adopt a technology involves strategic
interaction among firms. Specifically, there are strategic advantages to those who adopt early. But in the
case of electric power plants, firms are fundamentally organized as local monopolies and do not engage in
strategic interactions. Similarly, since scrubbers are purely abatement technologies and do not improve
the productive efficiency of generation, it is not likely that strategic interaction would be a response, that
is, they would not voluntarily install this technology to compete. Therefore, I believe the rank effect
13 These studies will often use other prices to control for market conditions in the industry if cost data of the actual
technology is not available. For example, Rose and Joskow (1990) use the annual market price of fuel as a proxy for
the cost savings of a more energy efficient technology. Similarly, Snyder et al. (2003) use the annual market price
of chlorine to control for market conditions. However, pure time series effects cannot be measured in the model that
I adopt in this analysis.
E. Frey
14
-------
should dominate the technological diffusion of scrubbers. The econometric model in the next section
presents a framework for testing the rank effect on technological diffusion.
III.B. Empirical Model
Consistent with much of the recent literature on diffusion, I will use survival analysis to estimate
the effects of the stringency and type of environmental regulation on the diffusion of scrubbers.14
Survival models have become very useful for many economic applications that deal with transitions
between events or the time to a particular event (see Kiefer, 1988; Beck, 1999; and Buis, 2006 for an
overview of this class of models). Survival models, also called duration models, arose in applications to
the medical field where the term "failure" or the event that is being modeled is a fatality of a patient. In
this context, the survivor function is the probability that the patient will survive past some point and time,
given that the patient has survived up to this point. Fuglie and Kascak (2001), Popp (2006), Rose and
Joskow (1990), Snyder (2003), Kerr and Newell (2003) and Karshenas and Stoneman (1993) have all
used survival models to investigate technology diffusion. Survival models have been utilized because
they are particularly helpful in measuring epidemic, rank, stock and order effects that have been described
in the previous sections (Stoneman, 2001).
In survival models, the object of interest is the hazard function, /?, (/). In the setting considered
here, the hazard function is defined as the probability of plant manager i adopting the technology at time t.
given that the adoption has not already occurred. More specifically, the hazard rate takes the following
form:
h (t, Xlt, P) = (2)
1 -F(t,Xtt,P)
where f (t, Xit, /?) is the probability density function for adoption, and /¦'(/) is the corresponding
cumulative distribution function. F(t, Xit, /?) = Pr( / < t, Xit, /?) is the probability that the time of
14 Along with the survival models presented above, a logit model for scrubber choice was also contemplated.
Similar results were obtained from the logit model in terms of significance and direction of the variables, but the
hazard approach is more useful when interested in analyzing the speed or the rate of technological change over time.
E. Frey
15
-------
adoption, '/'has happened prior to the current time. I. f3 is a vector of parameters to be estimated. The
vector Xn captures relevant firm characteristics and regulatory information. It includes the expected cost
of adoption or installation cost, CiT , and /'s individual characteristics, Zit, which captures size of the
unit, type of technologies already in place, regional effects, age, type of ownership structure of the firm,
and distance from a major low sulfur coal source, the Powder River Basin (PRB) in this case. In addition,
Xit includes regulatory variables. Rtl, that encapsulates regulatory structure and stringency. In other
words, the vector X„ in the econometric model can be expressed as Xit= {Z((, CiT, Rit}.
The survival model captures the underlying decision process of each plant manager that was
outlined in the theoretical model. That is, every time period t. given that a scrubber has not already been
installed, the manager has the choice of either installing a scrubber or delaying the decision until the next
time period. The estimation of the hazard rate in Equation (2) measures the effect of the explanatory
variables introduced in the theoretical model on the conditional probability of adoption. In this
estimation, the effect of the type and stringency of regulation on the diffusion of scrubbers will be the
central interest. Notice that the benefits before and after adoption, denoted by B° (•) and B1 (•) in the
theoretical model, are not measured. These data are not observed and are thus omitted from the empirical
model. Most studies omit benefits in the empirical model, due to the lack of data including Kerr and
Newell (2003), Karshenas and Stoneman (1993), Fuglie and Kascak (2001), Snyder et al. (2003) and
Popp (2006).
The proportional hazard model (a specific type of survival model) is commonly used in the
diffusion literature and allows for convenient interpretation of the data. This model separates the hazard
function into a baseline hazard rate, h0 (I) and an exponential function, exp(XitP) in which the
covariates multiplicatively shift the baseline hazard. The proportional hazard function is given by the
following equation:
K (t, Xlt ,/?) = /?„ (t)exp(X'tfl) (3).
16
-------
The in the proportional hazard model, the baseline hazard function may be left unspecified (the
semi-parametric approach) or can be assigned a functional form (the parametric approach). With the
parametric approach, the baseline hazard function could be constant, increasing or decreasing over time,
depending on the specific functional form; commonly used functional forms for the baseline hazard
include exponential, Weibull, Gompertz or gamma distributions. Alternatively, with the semi-parametric
Cox proportional hazard model, no assumption is placed on the relationship of the baseline hazard over
time. However, for all proportional hazard models, it is assumed that the shape of the baseline hazard is
the same for all decision makers in the model and the function is shifted by varying unit level
characteristics. The semi-parametric analysis allows more freedom in terms of the shape of the hazard
function, but parametric models produce more efficient coefficients, only if the functional form of the
baseline hazard is specified correctly.
Ideally, theory would lead the researcher to choose one functional form over the other. However,
in this application, there is no reason to assume any particular pattern for the hazard function. In this
estimation, I chose the Cox proportional hazards specification, rather than a parametric model.15 This
decision was based on a comparison of the best fitting parametric model, which used the Weibull
distribution and the Cox model. The Weibull model was the best fitting parametric model, based on the
Akaike information criterion (AIC), which provides a method of comparing different distributional
assumptions for parametric models (Akaike, 1974). To evaluate these two models, I estimated a
smoothed hazard function from the data, which was increasing and then decreasing over analysis time.
The same shaped function is obtained when estimating the hazard function in the Cox proportional
hazards model. This shape differs from the function estimated with a parametric model and presents
evidence that semi-parametric model is a better fit for the data.16
15 Results for the parametric model are available upon request. The results of the empirical estimation were robust
to changes in the model specification.
16 Recall that if the hazard is increasing over time, there is evidence of epidemic effects. Since the hazard rate in the
Cox proportional hazards model does not increase over time, I do not find evidence of epidemic effects in this
model.
17
-------
To generate the partial log-likelihood used in estimating the coefficients, assume first that each
failure time is unique (or that there are no ties of failure times) and there is no right censoring.17 Then,
the sample contains K unique exit times, from Th...,TK and the risk set includes R(IJ individuals at time I,.
which includes all those individuals that have not exited yet or R(tj)={j:ti > Tkj. Then, the probability
that an individual will fail at time Tk, conditional on only one individual failure is
e^i
Pr(7 = | R(l ) = ——. The estimates of ft are a result of maximizing the partial log-
Ze 1
likelihood function InL = , which does not rely on the baseline hazard
function. Of course, this is a simplified equation. In my analysis, the Breslow method forties is used,
since not all failure times are unique and the partial likelihood function is modified slightly to account for
right censoring.
I also consider the shared frailty model, in order to control for unobserved heterogeneity. That is,
units that were chosen to install scrubbers may possess some unobserved characteristic that make them a
lower cost unit and is therefore more likely to adopt. The shared frailty model is a modified Cox
proportional hazard model that controls for unobserved heterogeneity by introducing an error term that
enters multiplicatively into the model and affects how fast a unit will adopt a scrubber. Error terms can
also be equal across groups within the sample (the firm is the group in this case), since units within a firm
have similar risks for adopting a technology; this is similar to random-effects in a linear model.
III.C. Description of Data
The Energy Information Administration (EIA) gathers a large amount of data for the purpose of
analyzing economic and environmental factors of energy and makes these publicly available. This
research utilizes EIA's Form-767: Steam Electric Plant Operation and Design Report, which is available
from 1985 onward. Data is not available prior to 1985. EIA-767 includes all generating units that have a
17 My data is right censored. That is, at the end of the observed time period, there are units left in the sample that
have not adopted a scrubber.
E. Frey
18
-------
nameplate capacity of 10 MW or more.18 From this annual survey, important unit and plant
characteristics are gathered, including age, nameplate capacity, boiler type, technology usage, type of
ownership and location. Other yearly production specific variables in the survey include heat input,
emissions of major pollutants and characteristics and quantity of fuels used and detailed information on
installed scrubbers. Finally, regulatory information such as the unit's phase, state-level emissions
standards and NSPS status are also gathered.
In my model, 1990 is the beginning of analysis time for all firms, since Title IV was signed into
law in this year. Prior to 1990, firms would only install scrubbers in order to comply with NSPS. In
addition, in the period 1990 through 1996, only Phase I units were at risk for scrubbing, since they were
the only units required to reduce emissions. After 1996, all units were at risk for scrubbing.19 The data
used for the scrubber choice analysis includes all coal-fired units from 1990 to 2002 that are at risk of
installing scrubbers in order to comply with Title IV. Units that were not at risk for scrubbing and are not
included in the sample are primarily units that were regulated under New Source Performance Standards
(NSPS). These units had emissions below 1.2 lbs of S02/mmBtu, which is the rate that was targeted
using permit allocations in Phase II; if a unit had an emissions rate lower than the allocation, there would
be no need to install a scrubber in either Phase I or Phase II. For this reason, coal-fired units that were
built between 1972 and 1977 and had emission levels below 1.2 lbs S02/mmBtu or had installed
scrubbers simultaneously with the construction of the unit were omitted from the sample. All new coal-
fired units that were built after 1977 either had emission levels below 1.2 lbs S02/mmBtu or had installed
scrubbers and are therefore also excluded. Essentially, the units included in the sample are at risk for
18 Units that have nameplate capacity between 10 MW and 100 MW are only required to fill out a portion of the
survey; units that are greater than 100 MW must complete the entire form. The information needed for this study
was required by all units with a capacity of 10 MW or greater, so this does not affect the sample size.
19 Since I am modeling the decision to install a scrubber and not the unit's abatement choice, I will include all units
that are regulated under Title IV and had not been affected by NSPS regulations. Although some units may decide
to buy S02 permits or switch to low sulfur coal in a given year to comply with Title IV, these decisions do not
preclude them from installing a scrubber, since they are not permanent actions.
E. Frey
19
-------
retrofit scrubbing between the periods 1990 to 2002.20 This results in 777 units in my sample, 42 of
which installed retrofit scrubbers over the time period. Again, there were more scrubbers installed during
this period, but only 42 scrubbers were retrofitted to existing generating units (as opposed to new units) in
response to Title IV. Although the number of adoptions is relatively small compared to the sample size,
the results below do show sufficient variation in the data to obtain reasonable parameter estimates.
However, this will limit the degrees of freedom available in the model.
There are some units that have a status of "cold standby", which means that it would take 3-6
months to reactivate them, or a status of "standby", which means the unit is not normally used but is
available for service at any time. Standby units are kept heated so that they can start up quickly and are
placed on standby in case of a failure of a currently operating boiler or in case of unusual periods of high
demand. Units that are on cold standby are deactivated (mothballed) and not kept heated. Although this
is not as common as standby units, a firm might keep a unit on cold standby if it is in poor condition, if it
has uncompetitive heat rates or because abatement costs due to environmental regulations are high. These
types of units (standby and cold standby) can have zero heat input for a year or several years at a time
during the observation period. I assume that during these shutdown periods, units are not at risk for
installing a scrubber, since they are not producing electricity and have no emissions during this period.
Thus, I treat these instances of shutdown as a truncation problem.21 There are also 18 units that
permanently retire, which I assume is four or more consecutive years with zero heat input and no
subsequent positive heat input;22 these units are subsequently dropped from the sample in the first year of
retirement.
20 Besides being under different regulatory regimes, the costs of installing a scrubber with a new unit rather than a
retrofitted scrubber are significantly lower. Therefore, the decision to install a scrubber with a new unit should be
considered separately from the decision to install a retrofit unit.
21 There were 2 units truncated from the left, 10 units have interval truncation or shutdown in the middle of the
sample, 18 units permanently shutdown and 16 units shutdown at the end of the sample and it is not known if it is a
permanent shutdown.
22 This assumption is based on the data; there are no units in the data from 1985 to 2001 that had positive heat input
after shutting down for four or more years.
E. Frey
20
-------
Recall that each unit is not only regulated under Title IV, but also by State Implementation Plans
(SIP), which are usually emission rate standards that the unit cannot exceed. These state regulations can
be binding to units in the sense that it is restricted to the emission rate that the state sets; if the state
emission rate is set tighter than federal regulation, it could effectively force a coal-fired unit to install a
scrubber. These state regulations are reported in EIA-767 in four units of measurement: lbs S02 per
million Btu (lbs S02/mmBtu), lbs of S02 emitted per hour, lbs of sulfur per million Btu and percent of
sulfur content of fuel. To make the regulations comparable, all units were converted to lbs S02/mmBtu,
EIA's preferred measure. Regulation variables in the model include a dummy variable for those units that
were subject to an emission standard equal to or below 1.2 lbs S02/mmBtu and a dummy variable for
units that were regulated under Phase I of Title IV.
One section of the EIA-767 survey provides annual data on those coal-fired units that have
installed scrubbers; these data are also available from 1985 to 2002 and are used to construct the scrubber
cost data set. Data include the age of the scrubber, S02 removal efficiency, type of scrubber, hours
utilized, amount of electricity consumed by the scrubber, operating and management costs each year,
installation cost, scrubber status and the amount of S02 emitted. Of course, the characteristics of the plant
and unit where the scrubber was installed is known as well. There are several categories of scrubber
status listed each year, including scrubbers that are in the planning stage, those that are being constructed
and tested, currently operating, on standby and retired. I only use data from years that the scrubber is
actually operating, since prior to and during construction of a unit, costs that are reported are estimated,
not actual costs.23 Moreover, units on standby or retired are not being used, and should not be compared
to units that are fully utilized.
These data are an unbalanced panel, since many scrubbers were built after 1990, due to Title IV.
The cost data set includes 192 scrubbers. The initial data set had 241 scrubbers but for various reasons,
231 only dropped the years that a scrubber was being planned or under construction under between the years of
1985-2002. Therefore, only three scrubbers were dropped entirely from the sample because they were in the
planning stages in 2001 or 2002 and had not begun to operate before the end of 2002.
E. Frey
21
-------
49 were dropped from the sample.24 Because of the limited number of existing scrubbers, this sample
does not exclude scrubbers that were constructed at the same time as the unit itself. This means that many
of these generating units were not in the risk set for installing scrubbers under Title IV; only 42 of the
retrofit scrubbers were installed to comply with Title IV, the remaining scrubbers were installed due to
regulations set by NSPS.25 Following Keohane (2005), the installation cost data were deflated using the
Handy-Whitman Electric Cost Index publicly provided by the U.S. Census Bureau and are expressed in
1996 dollars.
IV. Empirical Results
The covariates used to estimate the semi-parametric model include individual generating unit
characteristics and regulatory variables. Unit characteristics include the natural log of nameplate capacity
in megawatts (MW), age of the unit as of 2002, the distance from the plant to the Powder River Basin
(PRB), the natural log of expected scrubber installation costs in dollars per lbs of S02 abated if
installation occurs and an indicator variable for units with very low expected installation costs (those in
the lowest quartile). I use a natural log transformation for size and installation cost because these
variables are skewed heavily to the right. Regulatory variables include a binary indicator for those units
that are regulated under Phase I and an indicator variable for those units that have very strict state
regulations (equal to or below 1.2 lbs S02/mmBtu).
Descriptive statistics for these variables are provided in Table 1 and are stratified by Phase I and
Phase II units. The descriptive statistics show that there are important differences between Phase I and
Phase II units. Phase I units tend to be larger and slightly younger than Phase II units. Moreover, Phase I
24 Eleven scrubbers were installed on units that burned fuels other than coal as their primary fuel (wood, oil and
refuse) and would presumably have different operating and maintenance costs and technological designs than coal-
fired scrubbers. Seventeen scrubbers were dropped because there were no years in the sample where the scrubber
was operating (i.e. the scrubbers status was standby, under construction, retired or planned between 1985 and 2002).
An additional 14 scrubbers were dropped because capital cost data were missing. Note that average size of the units
that were dropped was very close to the average size of units that were in the sample, so there was no relationship
between those units that were dropped and size. Finally, seven more units were dropped because there was no
matching plant and unit found in the EIA-767 production data.
25 Although I include all scrubbers in the sample, when I estimate the equation for installation costs, I use a dummy
variable to control for those units that are retrofit scrubbers.
E. Frey
22
-------
units have considerably higher pre-Title IV emission rates but tend to also have higher emission standards
than Phase II units (i.e. Phase I units have less stringent state standards than Phase II units). There is a
positive correlation between pre-Title IV emission rates and emission standards imposed on the unit.
This pattern may have emerged because states were protecting their local coal markets. For example, in
Ohio, there are large supplies of high sulfur coal. To encourage plants to buy local coal, a non-binding
emissions rate is set so that plants continue to burn Ohio coal.26 Finally, the estimated installation cost of
scrubbers is lower for Phase I units than Phase II units. Since this cost estimate is expressed in terms of
dollars per ton of S02 abated, this shows that larger, dirtier firms generally have lower expected scrubber
installation costs; that is, there seems to be some economies of scale for scrubbers.
Table 2 gives the results for four separate semi-parametric hazard models. The first model tests
all 777 units that are at risk for installing scrubbers when Title IV was enacted in 1990. The second
model is a frailty model, where the frailty is gamma distributed and enters multiplicatively on the hazard
function to control for unobserved heterogeneity. Frailties for all units within a firm are assumed to be
equal; that is, units owned by the same firm face the same risk of scrubbing. The third model uses a
restricted sample where those plants that are in three NERC (North American Reliability Council) regions
that have no scrubbing activity are omitted.27 The three regions omitted include 64 subjects in the
Electric Reliability Council of Texas (ERCOT), the Mid-Continent Area Power Pool (MAPP) and the
Southwest Power Pool (SPP). The units in these regions are omitted because they are potentially not in
the risk set. That is, these plants face certain regional and regulatory conditions that would make
scrubbing an unattractive option. These omitted regions include the west north central and west south
central census regions of the U.S. and are close to the Powder River Basin area. It is plausible to think
that they are not at risk for scrubbing because of their proximity to a large source of low sulfur coal.
261 use the term "non-binding" emission rate loosely in this context. I am referring the fact that often a very high,
lenient emission rate will be set for the unit and even with a very high-sulfur coal, would be able to meet the
emission rate standard.
27 Prior to 2006, plants in the risk set spanned 10 NERC regions. There have since been several major changes in
NERC region boundaries and names. I will use the pre-2006 definitions, since this resembles conditions when
plants were at risk for scrubbing.
E. Frey
23
-------
Strictly speaking, the restricted sample is biased in terms of representation of the entire population of
generating units since it excludes three entire regions. However, estimates may be closer to the true
population values, under the assumption that units in these regions were never at risk for scrubbing.
Since no scrubbing activity is observed within these regions, omitting these 64 observations does not
create biased estimates and there is no loss of explanatory power. The fourth model is a frailty model,
using this restricted sample.
The four Cox proportional hazards models in Table 2 show similar results. In all models, age,
phase, state regulations and low expected installation costs are statistically significant and are the major
factors that drive scrubber adoption. The effect of age on the likelihood of scrubbing is negative, similar
to findings by Popp (2006). Other studies, for example, Fuglie and Kascak (2001), Snyder et al. (2003)
and Rose and Joskow (1990) do not include age in the estimation. Karshenas and Stoneman (1993) and
Dunne (1994) find that the coefficient on age is not statistically significant. With my data, the negative
sign on age is expected because all generating units in this sample were installed before 1977. The oldest
units in the sample are approaching retirement age; the expected economic benefits to install a costly
scrubber on a unit that will soon be retired are negative.
The size of the unit, or the natural log of nameplate capacity (in MW), does not have a
statistically significant effect on the hazard but has a positive coefficient. However, this may be due to a
high negative correlation between age and size. When capacity is used without the age variable, it is
statistically significant and still has a positive effect on the hazard rate.28 This is consistent with the
prevailing literature; a number of studies report a significant positive effect due to firm size (Rose and
Joskow, 1990; Popp, 2006; Snyder et al., 2003; Kerr and Newell, 2003; Dunne, 1994; and Fuglie and
Kascak, 2001). Karshenas and Stoneman (1993) find that the coefficient on size is positive but not
statistically significant. They argue that size should have a net positive effect on the hazard as a result of
three effects on adoption behavior. First, large firms are more likely to adopt because there are positive
scale effects of adoption and can adopt new technology at a lower cost. Also, large firms are willing to
28 When the estimation excludes the age variable, there are no significant changes in the coefficient values.
E. Frey 24
-------
undertake more risk than smaller firms, and are therefore more likely to adopt. In the opposite direction,
larger firms may have more rigid managerial approaches, and may actually be less likely to adapt to
changes, relative to smaller firms. In my study, I do find evidence of positive scale effects, since larger
units do have lower expected scrubber installation costs, but I have no data on risk or managerial
preferences.
The regulatory structure and stringency facing a unit is found to be very influential on the
decision to adopt scrubbers in all models. Units that are regulated under Phase I are more likely to install
scrubbers than units regulated under Phase II. This is an expected result; Phase I units are generally
larger and have higher emission rates, lower costs per ton of abatement and thus are more likely to install
scrubbers. The state regulation, which is expressed as a dummy variable for units with an average
emission rate standard below 1.2 lbs S02/mmBtu, is positive.29 Moreover, the magnitude of the
coefficient is very high which suggests that units that face strict state regulations are very likely to install
scrubbers. In my sample, 13 units out of the 42 units that installed a scrubber also had state emissions
rate restrictions at 1.2 lbs S02/mmBtu or below. Again, this is an expected result and is consistent with
findings from Popp (2006), Kerr and Newell (2003) and Snyder et al. (2003). That is, strict emission
standards induce firms to adopt cleaner technologies. I also explored an alternative regulation variable,
the average emission rate (in lbs S02/mmBtu) imposed by the state between 1990 and 2001. This
variable was negative and statistically significant as expected, which shows that increasing the stringency
of the command-and-control variable increases the hazard rate. This variable was not used in Table 2,
because various specification tests showed that the dummy variable was a better fit for the model. Given
the strong influence of command-and-control regulations on the adoption of scrubbers, it is not
unreasonable to wonder how many scrubbers there would have been if all of the states imposed very strict
regulation on plant emissions. Using the estimated coefficients, it is possible to predict the number of
29 Recall that in Phase II, units are allocated enough S02 permits to emit 1.2 lbs S02/mmBtu. If emissions exceed
this amount, the unit must either abate or buy more permits (although all units face the same permit price at any
given time). Therefore, I consider a state regulation that requires emission rates to be below 1.21bs S02/mmBtu to
be strict.
E. Frey
25
-------
scrubbers, given that all units had strict state regulations. Using Model 1 in Table 2, this prediction
resulted in over ten times more units adopting scrubbers. This indicates that if all states had adopted strict
regulations, there would be considerably more scrubbers adopted, ceteris paribus.
The variables used to represent expected scrubber installation cost in the model includes the
natural log of expected installation cost per ton of S02 abated (if a scrubber was installed) and a dummy
variable that equals one if this expected installation cost was in the lowest quartile of the sample. The
estimated coefficient on the expected installation cost has a negative effect on the hazard rate and is only
statistically significant in Model 2. The coefficient on the low cost dummy variable is statistically
significant and positive for all models. This suggests that units that have very low expected installation
costs are more likely to install scrubbers than those with higher expected costs. This result is in
accordance with the few empirical and theoretical diffusion studies that have used installation costs,
including Kerr and Newell (2003), Keohane (2005), Stoneman (2001) and Jaffe and Stavins (1995).
Finally, the distance to the Powder River Basin (PRB) has a near zero effect on the hazard rate. Since the
PRB is a major source of low sulfur coal, those plants that are closer to the PRB would have lower
transportation costs of low sulfur coal, which would presumably make scrubbing less attractive.
Nonetheless, in this estimation, distance to PRB is not a determinant in the decision to install a scrubber.
Several diagnostic tests were performed to assess the model specifications and assumptions.
First, I performed an omitted variables test (the link test) and found that the correct variables are
incorporated in the model. To check the proportional hazard model assumption, I interacted all variables
with time and tested the null hypothesis that the coefficients on the interacted variables were equal to
zero. The null hypothesis could not be rejected at the 0.10 level, which means that there is no evidence of
violating the proportional hazards assumption. An alternative test, based on the Schoenfeld residuals
confirms this finding for all variables except the natural log of expected cost and the distance to PRB.
Finally, I tried two graphical tests that assess the proportionality of hazards for discrete covariates shows
that the phase dummy and the low cost dummy may have violated the proportional hazard assumption for
one of these tests. But, since these variables have passed other tests of the proportional hazards
26
-------
assumption, I do not put great importance on this finding alone. Overall, I conclude that the Cox
proportional hazard model is a reasonable specification for the data.
I have also tried to use a number of other covariates in this model including the type of ownership
structure that the plant faces, NERC region of the plant and variables that capture the quality of coal in the
region surrounding the plant. The inclusion of these variables had little, if any, effect on the magnitude
and statistical significance of the coefficients presented in Table 2. No notable insights were obtained
from the inclusion of these variables and were thus excluded from the final results.
Shared frailty models in this context seem particularly appropriate. As previously explained,
shared frailty models control for unobserved heterogeneity by introducing an error term in the hazard
function that enters multiplicatively. Under the shared specification, observations in the same group (i.e.
units in the same firm) are correlated, which is analogous to a random-effects model. Examining the
shared frailty estimates in Table 2, the parameter theta is the estimated frailty variance in the model. The
null hypothesis that theta equals zero is tested, using a likelihood-ratio test. In both frailty models, the
null hypothesis is rejected. Thus, I conclude that there is correlation between units that are owned by the
same firm. Also, examining the data in more detail, out of the 21 firms that had adopted at least one
scrubber, only six of these firms adopted only one scrubber. Therefore, it seems appropriate to control for
this unobserved heterogeneity using the shared frailty model.
In Table 2,1 present estimates for the entire population of generating units that were at risk for
scrubbing (Model 1 and Model 2) and a subset of the population that omitted two NERC regions that had
no scrubbing activity (Model 3 and Model 4). Choosing between these two samples is a judgment call.
That is, the subset of generating units should be used only if the probability of the omitted units installing
a scrubber over its lifetime is zero. Otherwise, if units had a positive probability of adopting, omitting
them will result in biased estimates of the coefficients. In this case, although it seems highly unlikely that
units in the omitted NERC regions (ERCOT, SPP and MAPP) would adopt, I also see no reason why they
would not scrub, if the cost of scrubbing was low enough. Notice that all three of these regions are in the
mid-west region on the U.S., which has enjoyed less stringent state regulations (none of the plants in
27
-------
these regions have emission rate restrictions below 1.2 lbs S02/mmBtu) or simply have very few at-risk
coal fired units in the region (ERCOT only has 4 generating units at risk). Moreover, these three regions
are in close proximity to cheap, low-sulfur coal, which may have deterred them from scrubbing in this
period. Given that the coefficients from both models are very close and there is no change in statistical
significance, I believe that either sample is appropriate.
V. Conclusion
In this study, I estimate the effects Title IV of the Clean Air Act on the diffusion of scrubbers.
Regulations imposed on electric generating units are the most important determinants for installing
scrubbers. With this particular technology, Title IV has directly caused 42 scrubbers to be installed as an
abatement strategy. However, units in Phase I were more likely to have scrubbers than units under Phase
II. Moreover, those units that faced strict state regulations were much more likely to have scrubbers
installed than those that had less stringent regulations. Therefore, state and federal regulations had a large
influence on scrubbing behavior.
Two unit characteristics were also important factors in the decision to install scrubbers. First,
those units that faced relatively low expected installation costs were more likely to have scrubbers
installed. Secondly, units in the sample that were relatively older were less likely to have scrubbers
installed. This is a result of the long life of scrubber technologies; it would not make sense to install an
expensive scrubber on a generating unit that is close to retirement. Although the importance of other unit
characteristics was tested, none had a statistically significant effect on the decision to install a scrubber.
Evaluating market-based policies, such as Title IV, and understanding the resulting diffusion
patterns will be particularly useful to policy makers when considering future climate change policy. In
this case, although Title IV required units to reduce S02 emissions, having very strict state regulations
was a strong determinant of whether or not a scrubber would actually be installed. Thus, the interaction
between state and federal regulations substantially influenced firm abatement decisions. Although theory
suggests that tradable permit systems give firms more flexibility in choosing abatement technologies, pre-
E. Frey 28
-------
existing command-and-control regulations can undermine that flexibility, leading to possible
inefficiencies in the diffusion of technology.
E. Frey
-------
Table 1: Descriptive Statistics
Variable Description
Phase I Units
Phase II Units
Total Sample
Mean
St Dev
Mean
St Dev
Mean
St Dev
Dependent Variable
Scrub Indicator
0.09
0.28
0.03
0.17
0.05
0.23
Ave. Year Scrub (if
1995
2
2000
2
1996
3
installed)
Unit Characteristics
Nameplate Capacity
289.26
247.73
200.90
213.23
238.32
232.46
ln(Nameplate Cap.)
5.32
0.85
4.88
0.90
5.07
0.91
Age (in 2002)
41.18
9.22
43.70
9.33
42.63
9.36
Distance to PRB (in
1134.48
579.42
1189.52
649.96
1166.21
621.28
miles) *
Expected Installation
108.27
128.69
145.31
232.32
129.63
196.03
Cost (in dollars/ton of S02
abated)
ln(Exp. Installation
4.52
0.66
4.73
0.76
4.64
0.73
Cost)
Low Cost Dummy
0.29
0.46
0.22
0.41
0.25
0.43
Regulatory Variables
Ave. State Standard (in
4.63
2.05
2.80
1.53
3.57
1.98
lbs S02/mmBtu)
Strict State Reg.
0.03
0.16
0.10
0.30
0.07
0.25
Dummy
Number of Units
329
448
111
Number of Plants
n/a
n/a
293
Number of Firms
n/a
n/a
121
Number of Scrubbers
28
14
42
Source: EIA-767, 1985-2002
'These data were provided by Nathaniel Keohane.
E. Frey
-------
Table 2: Parametric Model Results
Variables
Cox Model
Cox with Shared
Cox Limited
Cox Limited
(1)
Frailty (2)
Sample (3)
Sample and
Frailty (4)
Age
-0.130***
-0.151***
-0.133***
-0.154***
(0.305)
(0.435)
(0.030)
(0.043)
ln(Nameplate
0.212
0.367
0.174
0.327
Capacity)
(0.224)
(0.392)
(0.226)
(0.394)
Distance to PRB
0.000
0.001
0.000
0.001
(0.000)
(0.001)
(0.000)
(0.001)
ln(Expected
-0.230
-1.192*
-0.230
-1.327
Installation Cost)
(0.476)
(0.717)
(0.484)
(0.793)
Low Cost Dummy
0.930*
1.390*
0.921*
1.327*
(0.528)
(0.797)
(0.533)
(0.793)
Phase I Dummy
0.912**
1 374***
0.863**
1.802***
(0.374)
(0.697)
(0.379)
(0.670)
Strict State Reg.
2.582***
4 674***
2.487***
4 437***
Dummy
(0.398)
(1.041)
(0.402)
(1.045)
Log Likelihood
220.021
207.284
218.719
206.709
Number of
777
777
713
713
subjects
Theta3
n/a
3.887***
n/a
3.736***
(1.687)
(1.659)
Figures in parentheses are the standard errors of the coefficient estimates.
*** Significance at the 1% level
** Significance at the 5% level
* Significance at the 10% level
aTheta is the estimated frailty variance in the shared frailty model. The likelihood-ratio tests the null
hypothesis that theta equals zero.
E. Frey
-------
References
Akaike, H (1974). "A New Look at the Statistical Model Identification." IEEE Transaction and
Automatic Control AC-19: 716-723.
Beck, Nathaniel (1999). "Modeling Space and Time: The Event History Approach". Research Strategies
in Social Science, Oxford University Press.
Bellas, Allen S. (1998). "Empirical Evidence of Advances in Scrubber Technology". Resource and
Energy Economics 20, 327-343.
Buis, Maarten L. (2006), "An Introduction to Survival Analysis". Working Paper.
David, P. (1969). "A Contribution to the Theory of Diffusion", Center for Research in Economic Growth
Research Memorandum, No. 71, Stanford University.
Downing, Paul B. and Lawrence J. White (1986). "Innovation in Pollution Control". Journal of
Environmental Economics and Management, 13, 18-29.
Dunne, Timothy (1994). "Plant Age and Technology Use in U.S. Manufacturing Industries". The RAND
Journal of Economics 25 (3), 488-499.
Fischer, Carolyn, Ian W. H. Parry and William A. Pizer (2003). "Instrument Choice for Environmental
Protection when Technological Innovation is Endogenous". Journal of Environmental Economics and
Management, 45, 523-545.
Fuglie, Keith O. and Catherine A. Kascak (2001). "Adoption and Diffusion of Natural-Resource-
Conserving Agricultural Technology". Review of Agricultural Economics 23 (2), 386-403.
Geroski, Paul A. (2000). "Models of Technology Diffusion". Research Policy, 29, 603-626.
Gray, Wayne B. and Ronald J. Shadbegian (1998). "Environmental Regulation, Investment Timing and
Technology Choice". Journal of Industrial Economics, 46, 235-256.
Griliches, Zvi (1957). "Hybrid Corn: An Exploration in the Economics of Technological Change".
Econometrica 24 (4), 501-522.
Hall, Bronwyn H. and Beethika Khan (2003). "Adoption of New Technology". New Economy
Handbook, Chapter 10, Academic Press.
Hannan, Timothy H. and John M. McDowell (1987). "Rival Precedence and the Dynamics of
Technology Adoption: An Empirical Analysis". Economica, New Series 54 (214), 155-171.
Ireland, Norman J. and Paul L. Stoneman (1986). "Technological Diffusion, Expectations and Welfare".
Oxford Economic Papers, New Series 38 (2), 283-304.
Jaffe, Adam B., Richard G. Newell and Robert N. Stavins (2002). "Technological Change and the
Environment". Environmental and Resource Economics, 22, 41-69.
E. Frey
32
-------
Jaffe, Adam B. and Robert N. Stavins (1995). "Dynamic Incentives of Environmental Regulations: The
Effects of Alternative Policy Instruments on Technology Diffusion". Journal of Environmental
Economics and Management, 29, S-43-S63.
Jung, Chulho, Kerry Krutilla and Roy Boyd (1996). "Incentives for Advanced Pollution Abatement
Technology at the Industry Level: An Evaluation of Policy Alternatives". Journal of Environmental
Economics and Management, 30, 95-111.
Karshenas, Massoud and Paul L. Stoneman (1993). "Rank, Stock, Order, and Epidemic Effects in the
Diffusion of New Process Technologies: An Empirical Model". The RAND Journal of Economics 24 (4),
503-528.
Keohane, Nathaniel O. (2005). "Environmental Policy and the Choice of Abatement Technique:
Evidence from Coal-Fired Power Plants". Yale University Mimeo.
Keohane, Nathaniel O. (2006). "Environmental Policy Design and Technology Diffusion". Yale School
of Management Working Paper.
Kerr, Suzi and Richard G. Newell (2003). "Policy Induced Technology Adoption: Evidence from the
U.S. Lead Phasedown". Journal of Industrial Economics 51 (3), 317-343.
Kiefer, Nicholas M. (1988). "Economic Duration Data and Hazard Functions". Journal of Economic
Literature 26 (2), 646-679.
Milliman, Scott R. and Raymond Prince (1989). "Firm Incentives to Promote Technological Change in
Pollution Control". Journal of Environmental Economics and Management, 17, 247-265.
Popp, David (2001). "Pollution Control Innovations and The Clean Air Act of 1990". NBER Working
Paper Series, #8593.
Popp, David (2006). "Exploring Links Between Innovation and Diffusion: Adoption of NOx Control
Technologies at U.S. Coal-Fired Power Plants". NBER Working Paper Series #12119.
Rose, Nancy L. and Paul L. Joskow (1990). "The Diffusion of New Technologies: Evidence from the
Electric Utility Industry". The RAND Journal of Economics 21 (3), 354-373.
Schumpeter, J. (1942). Capitalism, Socialism and Democracy. New York: Harper.
Snyder, Lori, Noland Miller and Robert Stavins (2003). "The Effects of Environmental Regulation on
Technology Diffusion: The Case of Chlorine Manufacturing". The American Economic Review 93 (2),
431-435.
Stoneman, Paul L. (2001). The Economics of Technological Diffusion. Blackwell Publishers:
Massachusetts.
Stoneman, Paul L. and Myung-Joong Kwon (1994). "The Diffusion of Multiple Process Technologies".
The Economic Journal, 104, 420-431.
U.S. Department of Energy, Energy Information Administration (EIA). 1985-2001. "Steam-Electric
Plant Operation and Design Report". EIA Form 767.
E. Frey
33
------- |