The Effect of Industrial Structure on Learning by Doing in

The RAND Corporation
The Effect of Industrial Structure on Learning by Doing in Nuclear Power Plant Operation
Author(s): Richard K. Lester and Mark J. McCabe
Source: The RAND Journal of Economics, Vol. 24, No. 3 (Autumn, 1993), pp. 418-438
Published by: Blackwell Publishing on behalf of The RAND Corporation
Stable URL: http://www.jstor.org/stable/2555966 .
Accessed: 07/05/2011 08:06
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at .
http://www.jstor.org/action/showPublisher?publisherCode=black. .
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].
Blackwell Publishing and The RAND Corporation are collaborating with JSTOR to digitize, preserve and
extend access to The RAND Journal of Economics.
http://www.jstor.org
RAND Journal of Economics
Vol. 24, No. 3, Autumn 1993
The effect of industrialstructureon learningby
doing in nuclear power plant operation
Richard K. Lester*
and
Mark J. McCabe*
*
Learningfrom experiencein the nuclear industryhas had a significant impact on the operating
performance of light water reactor (LWR) power plants. Performance comparisons between
the United States and France indicate that the relationship between experience and performance has been strongly influenced by industrial structure. In the United States, a sizable
operatingperformancepenalty has been paid both as a result of the diffusion of several types
of LWR technology and because of the relative scarcity of multiunit sites caused by the
fragmented structureof the electric utility industry. In France, by contrast, performance has
benefitedfrom the very high degree of plant design standardization and the prevalence of
multiunit siting. These results suggest both short-term and long-term opportunitiesfor improvement in the performance of the American nuclear industry.
1. Introduction
* Recent studies of technological innovation have drawn attention to the cumulative
economic importance of secondary technical refinements, improvements, and adaptations
to products or processes after their introduction to the market (Rosenberg, 1982; Freeman,
1982). Secondary improvements may occur as a result of "exogenous" scientific or engineering advances, or they may emerge from the experience accumulated with the technology
following its initial introduction. The role of experience in bringing about technological
change has been extensively analyzed in the literature on learning by doing. Early articles
on the subject include Arrow (1962), Alchian (1963), and Hirsch (1956).
In this article we investigate the impact of learning on the operating performance of
nuclear power plants of the light water reactor (LWR) type in the United States and France.
Of particular interest is the effect of industry structure on the relationship between learning
and operating performance. We conjecture that two structural factors prevent the benefits
of operating experience from being uniformly distributed throughout national industries.
* Massachusetts Institute of Technology.
* * Antitrust Division, U.S. Department of Justice.
Views expressed herein are not necessarily those of the U.S. Department of Justice. This research has been
supported by grants from the Mellon Foundation and the MIT Center for Energy Policy Research. The authors
would like to thank Paul Joskow, Richard Schmalensee, Jeffrey Wooldridge, Tim Bresnahan, John Solow, and
several referees for their helpful comments.
418
LESTER AND MCCABE
/
419
First, differences in reactor technology may limit the potential for information sharing.
Second, even when reactors have similar technology, the additional costs associated with
sharing information between plant sites and across corporate boundaries may discourage
this activity. Performance in the U.S. nuclear power industry is likely to be affected by both
factors: by the first because U.S. reactors are not standardized, and by the second because
the ownership of reactorsis spreadover many firms. The French nuclear industry, in contrast,
has a single owner, numerous multireactor sites, and a high degree of plant standardization.
Empirical evidence of a "structural"performance penalty in the United States would have
important implications for decisions affecting the choice of technological standards, siting
policy, and firm size in the nuclear industry.
Most previous studies of learning in nuclear power plant operation ignore the role of
information sharing between reactors. Typically, attention is restricted to the relationship
between a reactor'soutput performanceand the experience accumulated at that same reactor.
Estimates of this relationship, based primarily on U.S. data, generally provide support for
a "same-reactor" learning curve (see Komanoff (1976), Joskow and Rozanski (1979),
Easterling (1982), and Krautmann and Solow (1988)). Rothwell (1990) contradicts this
conclusion for most types of U.S. reactors. He attributesthis result to differences in the data
sampling period.1
Roberts and Burwell ( 198 1) is the single exception to the same-reactor approach. Their
results suggest that reactor safety performance improves more quickly at multireactor sites
than at sites with a single reactor.2 However, the effects of technology on learning are not
considered, and the role of intersite information sharing is not adequately addressed.
In contrast to the existing output literature,this article adopts a "multireactor"approach
to learning: reactor output performance is specified to be a function of the experience
accumulated by all reactors operating in a national industry. By taking this approach, we
are able to evaluate the effects of industry structuredescribed above. The article is organized
as follows. Section 2 describes the structure of the nuclear industry in the United States and
France. In Section 3 we introduce some key aspects of nuclear power plant operations and
explain our choice of "equivalent" availability as an appropriate measure of operating performance. In Section 4 we present a conceptual model of learning that incorporates the
potential effects of nuclear industrial structure. Section 5 describes the functional form of
the estimating equation. Sections 6 and 7 discuss the data and the econometric methods
used in estimating the equation. The final three sections contain the results and conclusions.
2. Nuclear industrial structure
* LWR technology is the dominant civilian nuclear reactor technology in use in both the
United States and France. About two-thirds of the operable LWRs in the United States are
of the pressurized water reactor (PWR) type; the remainder are boiling water reactors
(BWRs). All French LWRs are of the PWR type.
Despite the technical similarities, there are major cross-national differencesin the structure of both the electric power and nuclear plant supply industries, which we expect to have
' The studies cited by Rothwell ( 1990) use data from the 1970s, when most reactors in operation were
relatively new; however, his analysis relies on data from the late 1970s and early 1980s. One might expect aging
effectsamong the reactor population to diminish any learning effects. Failing to account for these competing effects
in an empirical model would bias the estimates of learning. Departing from the literature, our model attempts to
account for both effects. The results strongly suggestthe existence of a same-reactorlearning curve during the period
1975-1986.
2 It is sometimes asserted that a reliable, high-output plant is also a safe plant. In the long run there is probably
much truth to this, since the skills and qualities required of the plant staff to enable them to sustain a high level of
output over a prolonged period are similar to those required for safe operation. In the short term, however, output
performanceis not necessarily indicative of safe operation; indeed, it might even have been achieved only because
certain safety margins were sacrificed. See Wilkinson ( 1983).
420
/
THE RAND JOURNAL
OF ECONOMICS
affectedpatterns and rates of learning in the two countries. The United States is characterized
by high levels of disaggregation in both sectors. At the end of 1986 there were 47 U.S.
utilities operating nuclear power plants, of which 24 had only one unit in service. The other
23 firms operated 70 units, 58 of which were located at sites with at most two units. Ten
of these utilities operated reactors at multiple sites. The largest nuclear utility had nine
operatingunits; the second largesthad seven. Seven utilities operated both PWRs and BWRs.
On the supply side, design responsibility has been divided between the reactor vendor
for the nuclear steam supply system (NSSS) and the architect-engineer for the balance of
plant. Four LWR vendors-Westinghouse (41 of the 94 reactors), Babcock & Wilcox (8),
and Combustion Engineering ( 14) for PWRs, and General Electric for BWRs (3 1) -have
been active in the United States; nine independent architect-engineeringfirms have participated in U.S. nuclear projects, and several utilities have acted as their own architect-engineer.
One consequence of this fragmentation has been a high level of design variation, even within
each type of LWR. The federal government, though active in several areas of the fuel cycle,
has played a relatively small role in the development of LWR technology since its initial
commercialization. The government's major technical contribution has been to carry out
safety-related research in support of nuclear regulation.
France lies at the other end of the spectrum of nuclear industrial structures. The stateowned utility, Electricite de France (EdF), is the sole owner of commercial nuclear power
plants. The other major industrial participants are Framatome, the sole NSSS vendor,3
Alsthom Atlantique, the sole nuclear turbine-generatorsupplier; and the government body
Commissariat de l'Energie Atomique (CEA), which has a major industrial presence in all
stages of the fuel cycle as well as a responsibility for basic research and development. EdF
acts as the architect-engineer,
construction
coordinator, and overall project manager for
nuclear power plant projects. Led by EdF, the French industry has implemented a standardization program more far-reachingthan that of any other country, building long series
of almost identical reactor units. These reactors are mostly located in four-unit clusters.
The manager of each four-unit site reports directly to EdF headquarters. There have been
two "preseries" and two major series of 900 MW units, a 1300 MW series, and in recent
years a new 1450 MW series has been started. The design differences between the various
900 MW series and preseries are quite modest. (The database for the present study consists
only of 900 MW units.) The centralized industrial structure and the high level of design
standardization might be expected to increase the effectiveness of information transfer between reactors.
3. Nuclear power plant operations
* In this article we use the annual "equivalent availability factor" to measure plant operating performance. This is defined as the total amount of energy that a plant couldhave
generated during the course of the year had it been called on to operate continuously at full
power, divided by the maximum annual energy output at continuous full-power operation.
Several previous studies of operating performance have used the annual "capacity factor"
as the principal performance indicator. This is defined as the energy actuallygenerated by
the plant during the course of the year, divided by the total amount of energy it would have
generated had it operated at its design power rating continuously throughout the year. For
plants operated in baseload mode, the equivalent availabilityfactor is identical to the capacity
factor. This is the case for virtually all U.S. nuclear plants. In some circumstances, however,
nuclear plants may be operated in a "load following" mode. This will occur if, for example,
the total generating capacity available to the operating utility exceeds the load. In France,
where nuclear plants now account for about half the total generating capacity, nuclear plant
3Framatome's
LWR technology was initially licensed from Westinghouse.
LESTER AND McCABE
/
421
load following occurs during part of the year. In these circumstances, the annual capacity
factor would understate the performance of the plant and its staW hence our choice of the
equivalent availability factor (hereafter referredto simply as the availability factor).
The maximum theoreticalavailabilityfactor for LWRs is less than 100%because reactors
of this type must be shut down periodically for a number of weeks for refueling. Such
outages are planned many months in advance. During the refueling outage, a fraction of
the core-typically a fourth or a third-is replaced with fresh fuel. Maintenance work that
cannot be done while the plant is in service is also carried out at this time.4 Apart from
these scheduled refueling and maintenance outages, unplanned outages or deratings may
also occur as a result of operator error, equipment malfunction, or the violation of safety
specifications.5
Availability is a composite indicator of how well a plant has been designed, built,
operated, and maintained. A common assumption is that both suppliers and operators seek
to maximize plant availability subject to the technical constraints imposed by refueling
requirements. While this is likely to have been true during the period covered in this study
(i.e., through 1986), it may not continue to be true indefinitely, at least insofar as operators
are concerned. Over time, nuclear plants, like any item of capital equipment, will wear out,
and the cost of maintaining them will increase. Utility planners must balance these costs
against the increased cost of purchasing or generating replacement energy during nuclear
plant outages.6 Eventually, the economically optimal course of action may be to settle for
lower nuclear plant availabilities. Since this decision may already have been made by operators of older units, we allow for this possibility in our model.
4. Learning
in nuclear
power
plant operations
* It is useful to distinguish between two different forms of learning occurring in nuclear
plant operations. First, experience gained from plant operations may result in a clearer
understanding of design-performancerelationships, leading in turn to design improvements
in subsequent generations of the technology ("embodied" learning). Alternatively, learning
may result in the adoption of new operating or maintenance practices that improve performance without requiring design modifications ("disembodied" learning).
With respect to disembodied learning, two mechanisms might be expected to lead to
improved performance at a particular unit as it matures (Joskow and Rozanski, 1979).
First, despite the best efforts of the commissioning team, there will be specific technical
bugs-improperly installed equipment, defective parts, incorrectly written software, and so
on-that will become evident during the early stages of operation and that will be corrected.
Second, as operations and maintenance personnel gain familiarity with the plant, they will
develop progressivelygreater facility in their tasks; fewer errorswill be made over time, and
improvements will continually be made to plant procedures.
The opportunities for learning will grow scarcer over time, and there will be a gradual
decline in the rate of improvement. Moreover, despite the best efforts at preventive maintenance, individual plant components will begin to wear out. Eventually, performance will
be allowed to decline as the marginal cost of maintaining high levels of availability exceeds
4Most reactorswere originallyoperatedon a 12-month fuel cycle. Many utilitiesare now findingit economically
attractive to switch to longer cycles (e.g., 18 months, or in some cases even two years).
On occasion, an unplanned outage at a particularunit dictates a response with an industrywide impact. The
response of U.S. regulatory authorities to the 1979 accident at the Three Mile Island plant is a case in point.
6 Since utilities are subject to economic regulation at the state level, this balancing of costs will also reflect
the relevant regulatoryincentives. Further researchis needed that explicitly examines the relationship between such
factorsas age, operations and maintenance (O&M) expenditures, replacement power costs, regulatory policy, and
plant performance.(Reports by the U.S. Energy Information Administration(1988, 1991) analyze the determinants
of O&M costs but do not consider overall plant performance.)
422
/
THE RAND JOURNAL
OF ECONOMICS
that of alternative sources of power. Though the nominal design lifetime of nuclear power
plants is 30 years, some deterioration can be expected well before plants reach the age
of 3O.'
What is learned at a particularunit is presumablyalso applicableto other units, provided
that there are sufficient similarities in design and operating procedures. We conjecture,
however, that this body of knowledge will not be symmetrically transferred to older and
newer units, since a fraction of it is likely to be age-specific-in the sense of pertaining only
to the general stage of a unit's life at which it was originally learned. We therefore expect
that information derived from operations at a given plant will tend, on average, to be more
valuable to younger units than to older ones.
Furthermore, it is likely that interreactorlearning will be influenced by industry structure. Since it is likely that the cost of information sharing is larger for intersite learning, we
expect that the potential for learning will be exploited more when units are located at the
same site. (Indeed, some pooling of operations and maintenance personnel is typical at
multiunit sites.) When information is transferred between sites, moreover, we conjecture
that the impact of learning will be greater when the plants involved are owned by the same
utility; ownership boundaries in the United States may obstruct the communication of
operating experience. It is also expected that the value of communication will be greater
for units that are more closely related technically to the source units. As noted earlier, LWR
technology includes two distinct design types, PWRs and BWRs. Although France's installed
nuclear capacity consists of only PWRs, the U.S. industry operates both types. In the United
States it is also reasonable to classify reactors according to vendor.8 Hence, we expect that
learning in the U.S. industry will exhibit a dichotomous pattern, with the most benefit
arising from information transfers within a given type of unit or, perhaps, among reactors
supplied by the same vendor.
Embodied improvements will become possible as designers learn more about the operating problems experienced by units that are already in service. In some cases these design
modifications may be made at the direct request of the user; in others they will be initiated
by the designer.9 (The designer will, of course, make some improvements based on "exogenous" scientific or engineering advances, i.e., improvements that are independent of
prior operating experience.)
5. The model
* The model proposed in this work is the log-linear availability regression shown in equation ( 1). This multiplicative specification assumes that performance is a separable function
of the various factors (an approach employed in the previously cited performance studies).
The basic version of the U.S. model is as follows: 10
In UNAVil = A + f1A GEit + /2AGE~it+ f3PREINTRASITEi + /4INTRASITE]it
+ /5INTRASITE2it + f6PREINTRAFIRMi + /7INTRAFIRM]it
PWR steam generators are a case in point. Contrary to early expectations, corrosion and related problems
have necessitated major overhauls and in some cases even replacement of steam generatorsseveral years after plant
startup.
8 Rothwell ( 1990) adopts this approach.
Despite proprietaryrestrictions, designers probably also derive some benefit from studying the performance
of plants supplied by their rivals. But the much greater access to their own units-it is not unusual for engineers
from the supplier firm to remain in residence at the plant site after operations begin-leads us to expect that this
will be the dominant communication channel for embodied learning.
10As discussed in Sections 8 and 9, we also estimate several more general versions of the U.S. model that
consider the possibility that information transfers between different classes of reactors are not as effective as those
between the same class. Two classifications are used: design type (PWR and BWR) and reactor vendor (PWRs:
Babcock & Wilcox, Combustion Engineering, and Westinghouse; BWRs: General Electric).
LESTER AND MCCABE
+
38INTRAFIRM2it
+
+ 311INTERFIRM2it +
39PREINTERFIRM,
012
+ f15TMI1 +? 16TMI2 +
In SIZEi +
+
/
423
310INTERFIRMlit
313VENDORi
+ 314VINTAGE,
sit
(i
=
1, .. ., 76)
(t = 1975, . . ., 1986).
(1)
UNAVit is the reactor unavailability (defined as (1 - annual availability) for unit i in
year t). Factors expected to improve (erode) performance should therefore be associated
with reduced (increased) unavailability. In recognition of the fact that refueling outages are
an unavoidable feature of LWR operation rather than the outcome of avoidable operator
problems, and that there is some minimum refueling period that cannot be further reduced
through learning, reported annual availability losses for years in which refueling occurred
were adjusted downward by subtracting an estimated minimum availability loss associated
with refueling. Details of this procedure are given in the Appendix.
The equation intercept is A. The next variable on the right-hand side of ( 1 ) constitutes
the learning curve for an individual reactor. AGEit (the time since reactor i entered commercial operation, expressed in years) is used as the measure of operating experience. Since
this variable also measures the opposing effect of plant aging, evidence that unavailability
initially declines (increases) with time indicates that the effect of learning(aging) is dominant.
A quadratic term is included to allow for the possibility that the net impact of these two
factors changes over time."
The next nine variables in equation ( 1 ) account for the effects of information transfers
from other reactors. The model allows for three different sources of information with respect
to ownership and location: (1) reactors owned by the same utility and located at the same
site (-INTRASITE-); (2) reactors owned and operated by the same utility but located at
other sites (-INTRAFIRM-); and (3) reactors owned and operated by different utilities
(-INTERFIRM-).
For each of these three source categories, the model furtherdifferentiatesbetween three
different information channels: ( 1) information transfers from older reactors before reactor i enters service (PREINTRASITEi,
PREINTRAFIRMi,
PREINTERFIRMi);
(2)
transfers from these older units after reactor i has started up (INTRASITEl it,
INTRAFIRMlit, INTERFIRMlit); and (3) information transfersfrom newer units occurring
after they have started up (INTRASITE2it, INTRAFIRM2it, INTERFIRM21t) .12Each variable is computed by summing the relevant number of reactor-yearsof service over all reactors
in that category. The implicit assumption is that the value of a year of service (in a learning
sense) is invariant across reactors and across time within each category of information
transfer. Note also that no quadratic term is needed for the interreactor learning effects
because they, unlike the reactor learning curves themselves, do not have an associated
degradation component.
SIZEi (the net power rating of reactor i in electrical megawatts) accounts for the effect
of scale on performance. It has been suggested that smaller plants, because of their generally
simpler designs, are easier to operate and maintain.
VENDORi (the total number of reactor-yearsof operating experience accumulated by
the earlier units supplied by reactor i's NSSS vendor before the construction start of reactor i) and VINTAGEi (reactor i's construction start date minus the construction start date
" The inability to separate the effects of learning and plant aging is a significant limitation of the model, as
one referee has noted. Resolution of this problem may well require a detailed study of the causes of outages at
individual plants.
12 "Newer units" are those reactors that began operation in the same month as reactor i or later. "Older
units" are those reactorsthat firstbegan operation in the months (and years) priorto reactori's month of commercial
operation.
424
/
THE RAND JOURNAL
TABLE 1
OF ECONOMICS
United States Data Summary (By NSSS vendor)
Variable
Mean
Standard
Deviation
Minimum
Maximum
Westinghouse (total reactor-yearobservations: 303)
UNAVAILABILITY
AGE
SIZE
VENDORS(207)
VINTAGE
TMIJ
TMI2
PREINTRASITE' (90)
INTRASITEl a (89)
INTRASITE2' (89)
PREINTRAFIRM(S) (14)
INTRAFIRM1(S) a (13)
INTRAFIRM2(S) (20)
PREINTRAFIRM(D) (74)
INTRAFIRM1(D) a (74)
INTRAFIRM2(D)' (36)
PREINTERFIRM(S)
INTERFIRM1(S)
INTERFIRM2(S)
PREINTERFIRM(D)
INTERFIRM](D)
INTERFIRM2(D)
.32
6.45
796
4.64
53.5
.33
.12
1.60
5.38
5.38
12.7
7.33
4.85
8.43
12.1
5.92
53.4
59.1
59.1
68.5
82.1
100
(.20)
(4.16)
(234)
(7.26)
(25.8)
(.47)
(.32)
(1.14)
(3.60)
(3.60)
(2.65)
(4.31)
(4.39)
(7.65)
(12.5)
(4.86)
(59.5)
(45.1)
(65.4)
(87.2)
(68.5)
(106)
.05
0
430
.83
0
0
0
.33
.08
.08
10.5
1.17
.25
1.17
.08
.17
0
0
0
0
0
0
1
18
1180
60.5
145
1
1
4.25
13.3
13.3
15.7
15.2
12.6
29.4
49.0
19.0
283
171
299
400
291
440
General Electric (total reactor-yearobservations: 246)
UNA VAILABILITY
AGE
SIZE
VENDORS (59)
VINTAGE
TMI1
TMI2
PREINTRASITE' (74)
INTRASITElJ (74)
INTRASITE2' (75)
PREINTRAFIRM(S)a (29)
INTRAFIRMAI(S)a(28)
INTRAFIRM2(S)a (28)
PREINTRAFIRM(D)a (25)
INTRAFIRM](D)a (24)
INTRAFIRM2(D)a (87)
PREINTERFIRM(S)
INTERFIRM1(S)
INTERFIRM2(S)
PREINTERFIRM(D)
INTERFIRM1(D)
INTERFIRM2(D)
.41
6.70
808
4.93
40.1
.36
.04
1.45
6.48
6.41
11.6
13.3
13.3
8.26
4.54
9.87
34.9
56.6
52.1
55.3
88.6
115
(.20)
(3.93)
(184)
(7.37)
(24.4)
(.48)
(.20)
(1.52)
(4.06)
(4.37)
(17.3)
(7.56)
(8.44)
(6.16)
(2.98)
(7.20)
(43.4)
(43.2)
(54.8)
(78.07)
(70.5)
(108)
.08
0
514
.25
0
0
0
.08
.08
.08
3.83
.67
1.17
4.67
.16
.25
0
0
0
3.83
0
0
1
16.1
1095
21.4
107
1
1
4.58
17.7
19.7
50.8
25.8
28.9
21.3
10.2
23.8
243
180
257
427
275
448
Combustion Engineering (total reactor-yearobservations: 74)
UNAVAILABILITY
AGE
SIZE
VENDORa (8)
VINTAGE
TMI1
TMI2
.30
5.76
790
7.98
28.9
.32
.11
(.17)
(3.71)
(159)
(6.99)
(27.29)
(.47)
(.31)
.04
0
478
2.58
0
0
0
.88
14.1
1100
16.41
122
1
1
LESTER AND McCABE
TABLE 1
/
425
Continued
Variable
Mean
Standard
Deviation
Minimum
Maximum
Combustion Engineering (total reactor-yearobservations: 74)
PREINTRASITE' (14)
INTRASITE] a (14)
INTRASITE2- (14)
PREINTRAFIRM(S)
INTRAFIRMA(S)
INTRAFIRM2A(S)
PREINTRAFIRM(D)a (18)
INTRAFIRM](D) a(16)
INTRAFIRM2(D) a
PREINTERFIRM(S)
INTERFIRM](S)
INTERFIRM2(S)
PREINTERFIRM(D)
INTERFIRM](D)
INTERFIRM2(D)
2.75
3.52
3.52
(2.18)
(2.78)
(2.78)
.58
.33
.33
6.67
8.75
8.75
12.1
6.44
(5.58)
(5.86)
7.42
.33
20.8
18.0
12.3
13.3
16.0
127
151
117
(18.7)
(12.0)
(18.4)
(145)
(102)
(128)
0
0
0
17.9
7.42
0
69.5
45.7
75.5
535
413
494
Babcock & Wilcox (total reactor-yearobservations: 68)
UNAVAILABILITY
AGE
SIZE
VENDOR
VINTAGE
TMI]
TMI2
PREINTRASITE- (24)
INTRASITE]P (24)
INTRASITE2- (24)
PREINTRAFIRM(S)
INTRAFIRM](S)
INTRAFIRM2(S)
PREINTRAFIRM(D)
INTRAFIRM](D)
INTRAFIRM2(D) a (15)
PREINTERFIRM(S)
INTERFIRM](S)
INTERFIRM2(S)
PREINTERFIRM(D)
INTERFIRM](D)
INTERFIRM2(D)
.39
5.39
886
(.20)
(3.71)
(24.4)
.06
.08
819
1
12.5
917
8.16
.38
0
1.42
8.50
8.50
(12.6)
(.49)
0
(.26)
(6.20)
(6.32)
0
0
0
1.17
.16
.08
40
1
0
1.67
22.16
22.42
2.73
5.49
11.8
12.1
101
176
88.4
(2.35)
(8.46)
(14.6)
(12.9)
(59.9)
(109)
(89.2)
.08
0
0
0
41.4
2.58
.17
6.50
24.3
46.8
42.3
231
397
369
Notes: A blank cell denotes the absence of any observationswith nonzero values. Interreactor
learningvariablesfollowed by (S) and (D) are computed using reactorsof the same and different
vendor class, respectively. To save space, we have omitted the summary statistics for these
variables when reactors are classified according to design type.
a Because so many of the observations for these variables assume a zero value, we have
calculatedthe correspondingstatisticsusing only observationswith positive values. The number
of positive observations is in parentheses.
of the earliest reactor supplied by the same NSSS vendor, expressed in months) account
for the effects of embodied learning and exogenous design improvements, respectively. Note
that although these variablesestimate averageratherthan firm-specificeffects,the assumption
underlying this specification is that the design improvements are firm-specific, e.g., in the
case of embodied learning, designers only learn from operating experience obtained from
units that they themselves have previously designed.
426
/
THE RAND JOURNAL
OF ECONOMICS
The next two variables in (1) account specifically for the reactions of regulatory authorities and utilities to the Three Mile Island (TMI) accident in March 1979. TMI1 (equal
to 1 for t ? 1979 and 0 for later years) is a dummy variable specified to capture the expected
negative impact of the accident on reactorsthat began operations before the accident. TMI2
(equal to 1 for reactors that began commercial operation after 1979 and 0 otherwise) is a
dummy variable specified to investigate whether the designers and operators of post-TMI
reactors were better able to respond to changes in the regulatory environment.13 These
changes affected both design (backfits) and operating and maintenance practices.
Any attempt to measure more general trends in safety regulation faces the intractable
problem of overcoming collinearity with the industrywide learning variables (-INTERFIRM-) and VINTAGE.Since regulatoryrequirementsbecame increasinglystringentduring
the sample period-and the available evidence suggests that this had an adverse effect on
plant availability (see U.S. Office of Technology Assessment (1984) and Beckjord et al.
(1987)), it is likely that our model underestimates the contribution of industrywide learning
and/or vintage effects. Finally, ej, is the error term, whose properties we consider in the
methods section.
The model we estimate using the French data has fewer variablesbecause of the simpler
industrial structure (one utility, one NSSS vendor), the lack of variation in reactor size (all
900 MW reactors) and the few observations prior to TMI (only two French reactors were
operating in 1979). Furthermore, because only one LWR technology is observed in France
(the PWR), we estimate one basic model. The French model, with variables defined as
before, is as follows:
In UNAV,1= A + 31AGEit+ 32AGEit+ (3PREINTRASITEi + 34INTRASITEJit
+ 35INTRASITE2it+ 36PREINTRAFIRMi + 37INTRAFIRMJit
+ 38INTRAFIRM2it+ 39VENDORi+ 310VINTAGEi+ eit
(i = 1, ...,
28)
(t = 1979, ...,
1986).
(2)
6. The data
* The data sample consists of 104 LWRs, 76 from the United States and 28 from France.
The sample includes virtually all the LWRs in the two countries larger than 300 MW that
had begun commercial operation by the beginning of 1985.14,15Annual availability factor
data for U.S. reactors were obtained for each year from 1975 to 1986, where applicable; for
reactors entering service after the beginning of 1975, the data series begins with the first
complete calendar year after the in-service date. The French data series begins in 1979 and
is also unbalanced. The sources of data are given in the Appendix.
Tables 1 and 2 give the means, standard deviations, and minimum and maximum
values for each variable in equations (1) and (2), respectively, as well as the total number
of observations.
13 We thank Paul Joskow for suggesting that TMI2 be included to distinguish between pre- and post-TMI
reactors.
14
The exceptions here are the U.S. units Millstone 1 and 2 (one GE BWR and one CE PWR, respectively)
and Arkansas Nuclear 1 and 2 (one B&W PWR and one CE PWR). Besides the San Onofre plant (see below),
these are the only units in either country located at the same site and belonging to different reactor classes. More
observationsthan this would be requiredto permit confident estimation of the effect of reactorclass on -INTRASITElearning. Note, however, that these units' operating experience is incorporated into the industry-level variables
(-INTERFIRM-).
The San Onofre units, consisting of one 436 MW CE PWR and two 1100 MW Westinghouse PWRs, are
included in the sample because we treat the plant as two separate sites. This is reasonable because of significant
technologicaland timing differences:the small CE PWR began operations fifteen years before the large Westinghouse
units.
15 We imposed the 1985 cutoff because our estimation technique, which includes the within-unit estimator,
requires at least two years of observations. Our most recent observations occur in 1986.
LESTER AND McCABE
/
427
French Data Summary
TABLE 2
Variable
Mean
Standard
Deviation
Minimum
Maximum
UNAVAILABILITY
AGE
VENDOR
VINTAGE
PREINTRASITE
INTRASITE]
INTRASITE2
PREINTRAFIRM
INTRAFIRM]
INTRAFIRM2
SITEDUMMY]
.19
2.77
.05
40.43
.69
2.80
3.55
19.37
18.87
20.28
.55
(.13)
(1.92)
(.34)
(20.72)
(.9 1)
(3.93)
(4.58)
(21.71)
(16.89)
(24.42)
(.50)
.01
.08
0
0
0
0
0
0
0
0
0
.76
8.08
2.67
90
2.75
18.25
19.58
84.42
69.41
109.30
1
Total reactor-yearobservations: 137.
7. Econometric
methods
* Annual observations for individual reactors are pooled to estimate a linear model whose
error structure is specified to include unit-specific components.16 By specifying ei1 in this
fashion, we can write equation (1) or (2) as follows,
+ it
(t= 1,...,
(i = 1,.. .,N)
(3)
ln Yit= Xi,*f+ai
Ti)
where fiis a (L X 1) vector of coefficients associated with the ( 1 X L) vector of independent
variables (Xit) of equation (1) or (2). The disturbance lit is drawn from an identically and
independently distributed distribution, lit - N(O, 2), and is uncorrelated with the (ai)
and the columns of (Xit). The unobservable individual effect, ai, is assumed to be a timeinvariant random variable, distributed independently across individual units, with vari-
ance
ara.
If the (ai) are uncorrelated with the columns of (X), consistent and efficient estimates
can be obtained using generalized least squares (GLS). However, if the (ai) are correlated
with the columns of (X), GLS estimation is biased and inconsistent. In this case, other
methods of estimation are available.17 For each of the models examined in this study for
which the estimated value of a 2 is nonnegative, the specification test presented by Hausman
(1978) fails to reject the null hypothesis that the (ai) are uncorrelated with the (X), at the
10% level of significance. Hence, GLS is the appropriate estimation method. The GLS
procedure and the specification test used in our analysis are straightforwardgeneralizations
of the techniques described in Hausman (1978) to the case of unbalanced panel data.18
Note that if a 2 < 0, OLS estimation is appropriate.
8. Results:
United States
* Homogeneous technology. The simplest specification of the model treats all U.S. reactors
as drawn from a homogeneous technology, as in equation ( 1). This specification assumes
16
In the case at hand, these unit-specific effects correspond to the idiosyncratic design characteristics and
operating practices associated with individual reactors.
17 Fixed-effects estimation is a simple alternative that produces consistent though inefficient estimates of F.
But use of this technique precludesestimation of time-invariantfactors,e.g., the impact of reactorsize on performance,
vendor learning, etc. To obtain consistent and efficient estimates of both time-varying and time-invariant factors,
the GLS/IV methodology proposed by Hausman and Taylor (1981) can be applied. Joskow and Schmalensee
(1987) employ Hausman and Taylor's approach in their analysis of the performance of coal-burning electric
generating units in the United States.
18
This generalized approach is outlined in Joskow and Schmalensee (1987).
428
/
THE RAND JOURNAL
OF ECONOMICS
TABLE 3
GLS Results for the United States:
Homogeneous Technology
Variable
Estimate
Standard Error
A
TMIJ
TMI2
AGE
AGE2
PREINTRASITE
INTRASITEl
INTRASITE2
PREINTRAFIRM
INTRAFIRM]
INTRAFIRM2
PREINTERFIRM
INTERFIRM]
INTERFIRM2
InSIZE
VENDOR
VINTAGE
-6.9543***
-.3604***
-.3745
-.0310
.0059**
-.0825**
.0051
-.0135
.0050
-.0066
.0007
.0000
-.001 5**
-.0017
.9499***
-.0189**
.0006
1.4050
.0757
.2894
.0423
.0025
.0380
.0104
.0093
.0048
.0051
.0052
.0008
.0007
.0011
.2043
.0091
.0024
* Two-tailed t-test indicates significanceat the 10%level.
** Two-tailed t-test indicates significance at the 5%level.
*
Two-tailed t-test indicates significance at the 1%level.
that the hypothesized relationships determining reactor unavailability are the same across
reactor classes, and in particular, that information transferredbetween units has no greater
value within a class of reactor than between classes. The results for this model are shown
in Table 3. Note that all the variables that are significant (at the 10%level or better) have
the expected sign.
AGE and AGE2 have the expected magnitude and sign, but only AGE2 is significant."9
The point estimates indicate that performance improvements based on experience gained
from operation of an individual reactor continue to occur for only 2.63 years, after which
equipment-aging effects take over and unit unavailability increases; that is, AGE* = 2.63
minimizes unavailability for the same-reactor learning curve. (The 95%confidence interval
for AGE* is [ -4.50, 9.77 ] .)20
The effect of preoperational learning from older reactors located at the same site
(PREINTRASITE) has the expected sign and is significant at the 10%level. The coefficient
estimate suggests that five years of such experience reduces initial unit unavailability by
some 30%. In contrast, PREINTRAFIRM, the coefficient associated with preoperational
learning from reactors owned by the same utility but located at different sites, has the wrong
sign, is estimated with less precision, and is an order of magnitude smaller.
The four coefficients for postoperational learning from both older and younger units
operated by the same utility at the same site and at other sites are relatively small and
insignificant; two of these have the incorrect sign.
The results for interutility learning suggest that the entire population of U.S. reactors
benefited from the accumulation of industrywide experience during the sample period.
19Collinearity between AGE, INTERFIRM] and INTERFIRM2 is a likely explanation for AGE's lack of
significance.Using the procedure suggestedby Belsley, Kuh, and Welsch ( 1980), we found a high condition number
(>30) and associated high variance-decomposition proportions (>0.5) for these three regression coefficients, thus
satisfying that study's double condition for diagnosing "degrading collinearity." This collinearity problem arises
whenever INTERFIRM variables are included in the estimation; when they are excluded, both the magnitude and
significanceof AGE can increase considerably. In a later section we present results for such a truncated model.
20 This interval is constructed using the Delta method described in Goldberger ( 1991 ).
LESTER AND McCABE
/
429
Although PREINTERFIRM (preoperational learning) is small and insignificant, both
INTERFIRM] and INTERFIRM2 are much largerin magnitude and are significant at the
5% and (almost) 10%levels, respectively. However, a comparison of these variables with
PREINTRASITE and INTRASITE2 suggests that, as expected, interutility learning is less
effective than learning that involves units at the same site.
The estimated effect of the TMI accident (TMII) indicates that reactor availability
losses were on average 30%smaller in 1975-1979 than in the subsequent seven-year period,
offsetting to some extent the benefits of the accumulated industrywide experience.21 In
contrast, units that began operation after TMI may not have sufferedthe same consequences:
TMI2 is comparable in size to TMI1, thus neutralizing the impact of this accident, but it
is estimated imprecisely.
Significant supplier effects were observed only for the embodied vendor learning function, VENDOR. The coefficient for VENDOR is significant at the 5%level and is reasonably
large in size. For example, a reactor operating with an NSSS supplied by a vendor whose
earlier projects had accumulated five years of operating experience prior to the start of the
unit's construction, would enjoy, other things being equal, a 10%decrease in unavailability
over its operating life.
The effect of unit size on unavailability is given by the coefficient of lnSIZE. The
coefficient is significant at the 1%level and, consistent with previous studies, the effect is
large and positive. By way of illustration, a 50% increase in unit size leads to, other things
being equal, a 47% increase in lifetime unavailability.
o Heterogeneous technology, equal learning coefficients. Because reactor technology in
the U.S. sample can be classified in a number of ways, e.g., by design type or by NSSS
vendor, it is reasonable to consider the possibility that some or all of the hypothesized
relationships may differ across reactor classes. Here we relax somewhat the homogeneity
condition of the previous section by allowing for differences in learning based on whether
the information transferoccurs within or between classes of reactors.22This is accomplished
by allowing a subset of the model variablesto vary accordingto the class of reactorgenerating
the operatingexperience. We indicate this distinction by appending the letter S (same reactor
class) or D (different reactor class) to the affected variables (-INTRAFIRM- and -INTERFIRM-). The expectation is that information transfer within a class of reactors is more
beneficial than exchanges between reactorsof differentclasses. Two technology classifications
schemes are considered: (1) design type (PWR vs. BWR) and (2) reactor vendor.23 Employing a chi-square test, 24 we are able to reject the null hypothesis that differences in the
learning coefficients do not exist for the first classification at the 1%level of significance; in
the second case, the significancelevel is slightly above 10%.The results for the two alternative
specifications are shown in Table 4.
The subset of coefficients whose variables are defined as before show, for the most part,
little change in size, magnitude, or significance. Estimates for AGE and AGE2 are essentially
unchanged. PREINTRASITE has the same magnitude and significance level in both spec2' To get some sense of the magnitudes involved here, consider the case of a utility operating a single PWR
that began service in January 1975. Ignoring individual reactor learning but taking into account learning from the
population of older and younger units during the subsequent years, the interfirm estimates suggest that not until
1984 was this unit's level of performance on par with its pre-TMI (1979) level.
22 Recall that this distinction is relevant only for the intersite/intrautility and interutility learning variables;
for the units in the sample, all reactors at any given site belong to the same design type or vendor group.
23 Since there are three PWR vendors in our sample and only one BWR vendor, these two schemes differ
only with respect to how variables for PWRs are calculated; for BWRs, variable calculations are unaffected by the
type of classification.
24 To apply this test correctly, the GLS transformation associated with the unrestricted model is used to
calculateboth the unrestricted and restricted error-sum-of-squares.
430
/
THE RAND JOURNAL
TABLE 4
OF ECONOMICS
GLS Results for the United States: Heterogeneous
Technology, Equal Learning Coefficients
Variable
A
TMI]
TMI2
AGE
AGE 2
PREINTRASITE
INTRASI TEI
INTRASITE2
PREINTRAFIRM(S)
INTRAFIRMI (S)
INTRAFIRM2(S)
PREINTRAFIRM(D)
INTRAFIRMI(D)
INTRAFIRM2(D)
PREINTERFIRM(S)
INTERFIRM] (S)
INTERFIRM2(S)
PREINTERFIRM(D)
INTERFIRMI(D)
INTERFIRM2(D)
lnSIZE
VENDOR
VINTAGE
Vendor
Design Type
-6.4674***
-.3579***
-.3135
-.0424
.0057**
-.0837**
-.0017
-.0183**
.0081
-.0203**
-.0006
.0018
.0044
.0015
.0001
-.0034***
-.0021*
.0004
.0012
-.0009
.8849***
-.0164*
-.0010
(1.2862)
(.0755)
(.2820)
(.0396)
(.0024)
(.0359)
(.0108)
(.0095)
(.0067)
(.0102)
(.0109)
(.0112)
(.0062)
(.0096)
(.0012)
(.0010)
(.0012)
(.00 1)
(.0012)
(.0011)
(.1867)
(.0086)
(.0022)
-6.3688***
-.3710***
-.5286*
-.0362
.0070***
-.0853**
.0085
-.0127
.0070
-.0230**
-.0101
.0076
-.0063
.0082
.0014
.0009
-.0041**
.0003
-.0027***
-.0012
.8752***
-.0185*
-.0024
(1.4592)
(.0759)
(.3118)
(.0429)
(.0027)
(.0382)
(.0105)
(.0094)
(.0067)
(.0117)
(.0112)
(.0098)
(.0059)
(.0095)
(.0017)
(.0013)
(.0020)
(.0009)
(.0009)
(.0012)
(.2188)
(.0098)
(.0028)
Standarderrors in parentheses.
* Two-tailed t-test indicates significance at the 10%level.
** Two-tailed t-test indicates significance at the 5% level.
* Two-tailed t-test indicates significance at the 1%level.
ifications. INTRASITE2, which measures the effects of postoperational experience generated
by younger units located at the same site, is unchanged in the vendor specification but is
larger and now significant at the 5% level in the design-type case; it is still appreciably smaller
than PREINTRASITE, a result consistent with the expectation that experience with older
units is more valuable to operators of recipient units.
The two VENDOR estimates show little change in size but are now only significant at
the 10% level. VINTAGE, which accounts for exogenous design improvements, now has
the expected sign in both specifications; it remains imprecisely estimated.
The results for the technology-specific intrafirm and interfirm learning variables are
mixed. For the design type specification, four of the six coefficients associated with samereactor class (S) learning have the expected negative sign, and three out of the four are
significant. In contrast, only one of the six different-class (D) coefficients is negative, and
all are imprecisely estimated. The results for the vendor specification are somewhat less
clear-cut. Only three of the six (S) coefficients have the expected negative sign; two of the
three are significant. Three of the six (D) coefficients are also negative, including one,
INTERFIRM] (D), that is significant at the 1% level. In contrast, INTERFIRMI(S) is
relatively small, positive, and imprecisely estimated.25
25
A potential explanation for this result is that values for INTERFIRMI (S) are much smaller on average
for B&W and CE reactors than for Westinghouse and GE units (see Table 1); in contrast, values for the
INTERFIRM] (D) variable are less divergent. When aggregatedto estimate a single coefficient, the effect is to shift
the impact of same-class interfirm learning to the INTERFIRM] (D) coefficient. In the next section, where we
allow coefficients to vary by class, this anomaly disappears.
LESTER AND McCABE
/
431
In general, results for intrasite, intrafirm and interfirm learning suggest that reactor
performance can be improved by applying the lessons learned from reactors of the same
class, with the most benefit deriving from knowledge generated by other reactors at the
same site and from older rather than younger units. With one notable exception, this conclusion holds for both reactor technology classification schemes.
o Heterogeneous technology, all coefficients differing.Here we permit all of the coefficients
in the previous subsection to differ according to reactor class. When we classify the reactors
by design type, it involves estimating twice as many parameters as in the previous section,
one subset corresponding to the PWR design and one to the BWR design. When we classify
reactors by vendor, it leads to a quadrupling of parameters. However, because two of the
vendors-Babcock & Wilcox (B&W) and Combustion Engineering (CE)-supplied a relatively small number of reactors (distributed across several owners), the full parameter set
cannot be estimated for each of these vendors. One problem is that no utility in the sample
operates B&W or CE units at more than one site, precluding estimation of -INTRAFIRMvariables. Another difficulty is perfect collinearity among some of the remaining regressors.
We address these problems in two ways. First, we estimate a full model based on the vendor
classification but using data for Westinghouse and General Electric reactors only. Second,
we use data for reactors from all four vendors to estimate a truncated version of the
full model.
For each estimation involving the full model we cannot reject the null hypothesis that
the respective parameter subsets (i.e., for the different vendors or the different design types)
are identical, at the 10%level. Therefore, the results of the previous section (i.e., heterogeneous technology, equal learning coefficients) are best in a statistical sense.26Nonetheless,
we should report that INTERFIRMI (S) and INTERFIRMI (D) in the vendor specification
are negative and positive, respectively, for both Westinghouse and GE reactors. Apparently,
the anomalous results for these variables in the previous section are an artifact of data
aggregation(see footnote 25). Also, some individual parametersdo vary considerably across
reactor classes. Since estimates for these parameters-in particular, SIZE and the
-INTRASITE- coefficients-are similar in the truncated model, we refrain from any further
discussion of the full model.
The truncated model omits -INTRAFIRM- and -INTERFIRM- variables, and thus it
makes no distinction between learning from the same or a different class of reactors. Thus,
the appropriate null hypothesis for testing whether vendor-specific differences exist involves
a truncated version of the model in the homogeneous-technology case. The results from
this test indicate rejection at the 1%level.
Results for the truncated model are given in Table 5. Vendor-specific coefficients are
estimated for all four vendors. In the case of B&W and CE, additional variableswere omitted
to avoid perfect or near-perfect collinearity.27
Compared to our earlier results, the most notable changes involve the intercept, A,
and lnSIZE. For Westinghouse reactors, both coefficients are now larger in magnitude and
remain significant at the 1%level. For CE and GE units, however, A and lnSIZE are much
smaller in magnitude and are estimated with little precision. These results suggest that the
adverse effect of reactor size on performance is restrictedto Westinghouse reactors. Rothwell
(1990) reports a similar result. Why this asymmetry should exist is unclear and deserves
26 This is only approximately true for the vendor specification. Recall that the test to reject the null hypothesis
for the vendor classification was not quite significant at the 10%level.
27
Only one regressorwas omitted due to "near perfect" collinearity: lnSIZE for B&W. When it is included
in the model, the estimates for A and ln SIZE had negative and positive signs, respectively, but they were several
times the size of the corresponding
estimates for Westinghouse
and insignificant.
432
/
THE RAND JOURNAL
TABLE 5
OF ECONOMICS
GLS Results for the United States: Truncated
Specification, Different Coefficients
Variable
Westinghouse
A
TMI]
TMI2
AGE
AGE 2
PREINTRASITE
INTRASI TEI
INTRASITE2
lnSIZE
VENDOR
VINTAGE
-7.1571*** (1.5042)
-.4220*** (.1074)
(.2043)
-.0943
-.1496*** (.0325)
.0056*** (.0016)
-. 1528** (.0638)
(.0175)
.0193
.0111
(.0166)
1.0990*** (.2311)
(.0109)
-.0038
-.0117*** (.0041)
Variable
B&W
A
TMI]
TMI2
AGE
AGE 2
PREINTRASITE
INTRASITEI
INTRASITE2
lnSIZE
VENDOR
VINTAGE
GE
-1.2980
-.3123***
-.2388
-..1398***
.0066***
-.0477
.0344**
.0130
.1335
-.0065
-.0004
(2.4252)
(.1232)
(.5838)
(.0412)
(.0021)
(.0741)
(.0173)
(.0153)
(.3542)
(.0312)
(.0038)
CE
-.6935
-.2796
(.3346)
(.2209)
-.0256
.0035
.0720
-.0495**
-.0634***
(.0840)
(.0063)
(.2167)
(.0241)
(.0211)
-1.6236
-.3699
(3.6672)
(.1995)
-.0361
(.0750)
-.0011
(.0048)
-.0544
(.1784)
(.06167)
-.1054*
-.1 109** (.0470)
.1238
(.5325)
-.0347
(.0730)
A blank cell denotes an omitted variable.
* Two-tailed t-test indicates significance at the 10%level.
** Two-tailed t-test indicates significance at the 5%level.
* Two-tailed 1-testindicates significance at the 1%level.
additional study.28The point estimates for A and lnSIZE in Table 5 imply, holding all else
equal, that Westinghouse reactors perform worse than GE or CE units for any observed
reactor size.
For Westinghouse and GE reactors,AGE and AGE2 are now significant at the 1%level.
The point estimates indicate that for Westinghouse reactors, the unavailability is minimized
when AGE = 13.26 years (the 95%confidence interval is [ 9.19, 17.34 ] ); for General Electric
units, this value is 10.59 years ([7.45, 13.74]). Furthermore, improvements due to learning
are initially quite large: for example, in the case of Westinghouse, after five years, holding
all else equal, unavailability declines by 46%. Of course, by omitting the -INTERFIRMvariables, these estimates for same-reactor learning are biased upward. Nonetheless, this
positive evidence of learning contradicts Rothwell's (1990) finding. Results for the other
two PWR vendors indicate that same-reactor learning is less important. Although the point
estimates for AGE and AGE2 suggeqtthat performance improves with experience, the coefficients are relatively small and insignificant.29
28 A possible explanation is Westinghouse's design decision to increase the number of cooling loops from two
to three and sometimes four for its largerunits. The other PWR vendors preservedthe two-loop configuration even
for their largest reactors. Increases in the number of loops and associated components could have contributed to
increasedunreliability.Evidence consistent with this hypothesisis provided by the serious corrosion-relatedproblems
that have occurred in Westinghouse steam generators. (Each cooling loop has its own steam generator.) These
steam generator problems have been a significant contributor to outages in Westinghouse units.
29 Application of the Belsley, Kuh, and Welsch ( 1980) diagnosticindicatesthat collinearityamong the regressors
does not account for this lack of statistical precision.
LESTER AND McCABE
/
433
Significant intrasite learning coefficients, with the expected negative sign, are observed
for three of the four vendors; the one exception is the GE INTRASITE] parameter, which
is both positive and precisely estimated. The differences observed among the three PWR
vendors suggests that the PWR design type classification may not be appropriate. TMI1 is
negative and of similar magnitude across vendors; precise estimates are obtained in the case
of Westinghouse and GE. Although the Three Mile Island accident involved a B&W reactor,
these results suggest that the response to this event was not vendor-specific. This contradicts
Rothwell's ( 1990) finding that the impact of the accident was limited to B&W units. In the
cases where VENDOR and VINTAGE are estimated, the expected negative sign is observed;
VINTAGE is now significant for Westinghouse reactors.
In sum, the results for the truncated model indicate that important differences exist
among vendors and that aggregatingthe data obscuresthese distinctions. The most important
differences include the effects on unit performance of reactor size, same-reactor learning,
and interreactor learning.
9. Results:
France
* The models estimated for the French industry are simplified by the fact that there is
only one utility operating a single type of reactor and that all the units in the sample are
the same size. In addition to equation (2) we also consider a truncated model similar to
the one estimated with the U.S. data. In our attempt to apply GLS to the French data, the
value of a2 was found to be negative, suggesting that OLS should be used for estimation.
The OLS results for the full and truncated models are given in Tables 6 and 8, respectively.
To facilitate comparisons with the United States, we also estimate a vendor specification
for the American industry based on equation (2) using data for Westinghouse and GE
reactors (the -INTRA FIRM- (D) variables are omitted as well as all of the -INTERFIRMregressors). The other two vendors are excluded because no American utilities operate these
reactors at multiple sites. These "full-model" (GLS) results are presented in Table 7. Since
we can reject at the 10%level the null hypothesis that the respective parameter subsets (i.e.,
TABLE 6
OLS Results: France, Full
Model (Equation 2)
A
TMI]
TMI2
AGE
AGE 2
PREINTRASITE
INTRASITEI
INTRASITE2
PREINTRAFIRM
INTRAFIRM]
INTRAFIRM2
InSIZE
VENDOR
VINTAGE
-.9530*** (.2960)
-.1993
.0087
-.0380
-.0214
-.0078
-.0053
-.0113**
-.0029
(.1487)
(.0283)
(.0866)
(.0305)
(.0238)
(.0049)
(.0057)
(.0127)
-.0597
-.0021
(.1730)
(.0067)
A blank cell denotes an omitted variable.
* Two-tailed t-test indicates significanceat
the 10% level.
** Two-tailed t-test indicates significanceat
the 5%level.
* Two-tailed t-test indicates significanceat
the 1%level.
434
/
THE RAND JOURNAL
TABLE 7
OF ECONOMICS
GLS Results: United States, Full Model (Equation 2),
by Vendor
Variable
Westinghouse
A
TMI]
TMI2
AGE
AGE2
PREINTRASITE
INTRASI TEI
INTRASITE2
PREINTRAFIRM(S)
INTRAFIRM](S)
INTRAFIRM2(S)
lnSIZE
VENDOR
VINTAGE
-7.3399*** (1.5764)
-.4183*** (.1067)
(.2121)
-.1047
-. 1504*** (.0330)
.0057*** (.0016)
-. 1602** (.0662)
.0253
(.0184)
.0159
(.0173)
.0125
(.0236)
-.0148
(.0298)
-.0195
(.0227)
1.127*** (.2431)
(.0114)
-.0031
-.0120*** (.0043)
GE
.0017
-.3329***
-.3675
-.1696***
.0085***
-.0845
.0591 ***
.0293*
.0066
-.0328***
-.0199*
-.0360
-.0005
-.0021
(2.5441)
(.1217)
(.6070)
(.0429)
(.0023)
(.0778)
(.0194)
(.0165)
(.0085)
(.0114)
(.0120)
(.3710)
(.0334)
(.0041)
Two-tailed t-test indicates significance at the 10%level.
*
* Two-tailed t-test indicates significance at the 5% level.
* Two-tailed t-test indicates significance at the 1%level.
for the Westinghouse and GE reactors) are identical, separate coefficients are reported for
the two vendors. Finally, to simplify comparisons with the U.S. truncated model, we report
again the results for the Westinghouse and GE data in Table 8.
For the French reactors, the two same-reactor learning coefficients, AGE and AGE2,
have the expected sign in both tables. In the full model, neither parameter is significant at
standard levels.30The point estimates for AGE and AGE2 are largerin magnitude than the
U.S. values in Tables 3 and 4. As expected, estimates in the truncated case are much larger
in magnitude and very precisely estimated. These point estimates indicate that unavailability
reaches its minimum when the value of AGE is 11.47 years ([-53.24, 76.18]) in the fullmodel case, and 6.35 years ([ 3.90, 8.79 ]) when the -INTRAFIRM- variables are omitted.31
These improvements are largerthan in the United States; e.g., in the truncated model, after
five years unavailability decreases by 72% in France and by 41-46% in the United States,
holding all else equal.
The intrasite coefficients all have the expected negative sign in the full model; in the
truncated case the one exception is INTRASITE2. None of the parameters are precisely
estimated (see footnote 30 regardingINTRASITEI). The magnitudes of the point estimates
suggest that information transferredfrom older reactors is more useful than transfers from
younger reactors.32
Application of the Belsley, Kuh, and Welsch ( 1980) diagnostic indicates that collinearityproblems involving
AGE, AGE2, INTRASITEI and INTRAFIRM2 may explain this lack of statistical precision.
31 Since the oldest reactor in the French sample had been operating for about nine years by the end of 1986,
it is not clear whether meaningful aging effects had yet been observed in the French reactor population.
32 Although the Westinghouse result for PREINTRASITE (see Tables 7 and 8) seems to suggest that intrasite
learning is sometimes more effective in the United States, one should consider the aggregateimpact of both PREINTRASITE and INTRASITEI. When a dummy variable (that equals 1 when PREINTRASITE > 0 and 0 otherwise)
is substituted for PREINTRASITE/INTRASITE] in the full and truncated (U.S. and French) models, the corresponding coefficients are negative, with magnitudes ranging between .20 and .30 for Westinghouse and French
reactors;the GE parameter is negative but much smaller in magnitude. For the Westinghouse and French reactors,
the dummy is precisely estimated in the full models and almost significant at the 10%level in the trucated cases.
Thus, intrasite learning in France is, on average, as good or better than it is in the United States. (This dummy
specification also greatly reduces collinearity between AGE and intrasite variablesin the French data. This problem
arises because all French reactors are located at multiunit sites.)
LESTER AND McCABE
TABLE 8
/
435
OLS Results, France and GLS Results,
United States: Truncated Model
France
A
TMI]
TMI2
AGE
AGE 2
PREINTRASITE
INTRASITEI
INTRASITE2
VENDOR
VINTAGE
Westinghouse
A
TMI]
TMI2
AGE
AGE 2
PREINTRASITE
INTRASITEI
INTRASITE2
InSIZE
VENDOR
VINTAGE
-7.1571*** (1.5042)
-.4220*** (.1074)
-.0943
(.2043)
-. 1496*** (.0325)
.0056*** (.0016)
(.0638)
-.1528**
.0193
(.0175)
.0111
(.0166)
1.0990*** (.2311)
(.0109)
-.0038
-.0117*** (.0041)
-.4657**
(.2093)
-.4254***
.0335***
-.0190
-.0013
.0086
-.0183
-.0149***
(.1177)
(.0130)
(.0822)
(.0256)
(.0208)
(.1602)
(.0033)
GE
-1.2980
-.3123***
-.2388
-. 1398***
.0066***
-.0477
.0344**
.0130
.1335
-.0065
-.0004
(2.4252)
(.1232)
(.5838)
(.0412)
(.0021)
(.0741)
(.0173)
(.0153)
(.3542)
(.0312)
(.0038)
A blank cell denotes an omitted variable.
* Two-tailed t-test indicates significance at the 10%level.
** Two-tailed t-test indicates significance at the 5% level.
*** Two-tailed t-test indicates significance at the 1%level.
In the full model, two of the intrafirm learning coefficients (PREINTRAFIRM and
INTRAFIRMI) have the expected sign. Although PREINTRAFIRM is relatively small and
insignificant, INTRAFIRMI, the coefficient for postoperational learning from older reactors
located at other sites, is larger and significant at the 5%level. INTRAFIRM2 has the wrong
sign but is small and insignificant. Together, these results translate into a large improvement
in the performance of newer units during the period 1979-1986.33
None of the corresponding Westinghouse intrafirm estimates are precisely estimated.
Because PREINTRAFIRM and INTRAFIRMI have similar magnitudes but opposite signs,
there is no benefit (on average) derived from experience accumulated at older plant sites.
In contrast, the GE results imply that intersite learning from older and newer units improves
performance.34These results suggest that intersite information sharing in France is relatively
effective, though it does seem primarily to benefit newer units. Such an interpretation is
consistent with the expected pattern of learning in an industry with highly standardized
reactor designs and a strong central mechanism for coordination and technical support of
site operations and maintenance. The French practice of rotating teams of startup engineers
3 Consider a hypothetical unit that began service in January 1982. Ignoring individual reactor learning but
taking into account the population of older and younger units operating in France, the intrafirm results indicate
that this unit's unavailability losses would have decreased aproximately 60% by 1986.
3 An estimate of the joint effect of PREINTRAFIRM and INTRAFIRM] in the case of GE and France,
using the point estimates evaluated at their sample means, indicates that the benefits of intersite learning from older
reactorsare comparable; the average impact of learning from newer units is greater for GE units.
436
/
THE RAND JOURNAL
OF ECONOMICS
and operators from one site to the next as new plants approach operation is a specific
example of the kind of arrangement that would explain the relatively strong result for
intersite learning.
10. Discussion
and conclusions
* The results for the United States and France generally support our expectation that
industry structureinfluences the relationshipbetween operatingexperience and performance.
In the United States, the diffusion of several types of LWR technology has reduced the
potential benefits from interreactorlearning. Interreactorlearning (both within and among
firms) is observed only when reactors of the same class are involved. In France, where only
standardized PWRs are employed, this problem is avoided. The organization and size of
firms has also affected industry performance. As expected, interreactor learning is greatest
at multiunit sites. However, because reactor ownership in the United States is spread over
many firms, multiunit sites are relatively scarce. In contrast, the existence of a single public
utility in France has made it possible for all its reactors to be located at multiunit sites.
This evidence of a "structural"performance penalty in the United States raises several
questions regardingthe choice of technological standards and the size and organization of
domestic nuclear utilities. First, should we characterize the U.S. outcome as inefficient?
Second, what actions might be taken to improve performanceamong the existing population
of reactors?And finally, what are the lessons for government and industry decision makers
as they consider the possibility of a "second nuclear era"?
The question of efficiency is not as simple as the preceding analysis might suggest. With
regardto standards, the failure of the U.S. industry to select a single reactor type reflects its
pioneering role in the commercialization of nuclear power. Early uncertainty about the
capital and operating costs of differenttechnologies encouraged the adoption of several types
of reactor. By the time this uncertainty was resolved in the 1970s, distinct networks of
reactor operators had emerged. The French decision to opt for Westinghouse-based PWR
technology at this time was informed by the American experience. Thus, although the
existence of multiple networks can be considered inefficient in an ex post sense, it may be
seen as one feasible outcome of a market operating with incomplete public information.35
It is possible that both BWR and PWR designs would still have been adopted in the
United States even if the industry had been more concentrated. Japan and Sweden, which
both have more concentrated electric utility industries, each adopted both types of reactor
system. However, it is clear that interreactorlearning within each class of LWR would have
been enhanced by the increase in multiunit sites associated with a shift to larger operating
organizations.36Why did the industry not anticipate this? One factor often cited is naive
expectations: few observers foresaw the need to redesign institutions to accommodate a
technology that was "just another way to boil water." The utilities, accustomed to declining
costs and predictable demand growth, used their past success with conventional power
sources as a guide for the future.37
Despite the clear importance of these structural problems, the results also suggest a
potential area for improvement among the existing reactor population. Recall that in the
United States the effectiveness of interreactor learning varies across the reactor classes (see
Tables 5 and 7). Since there is no apparent technological explanation, it is more likely that
35 See McCabe ( 1991 ). David and Rothwell ( 1991 ) consider a similar set of issues from a central planner's
perspective.
36 Other dimensions of U.S. nuclear industry performance benefited from large operating organizations as
well. For example, McCabe ( 1991 ) shows that learning during plant construction was greater for firms building
several reactors. An earlier discussion of this subject is found in Lester ( 1986).
37 See Bupp and Derian (1978) on this point.
LESTER AND McCABE
/
437
organizational differences among U.S. utilities account for this pattern. If firms differ in the
extent to which interreactor information exchange is promoted, then pooling of the data at
the level of reactor class could produce the observed asymmetries. If this is true, U.S. firms
would benefit from sharing information about the mechanisms they employ to facilitate
learning.38And given our results for interreactor learning in France, consideration should
also be given to organizational practices in that country.
The structural lessons of our analysis are more useful to the architects of a second
nuclear era in the United States. Regarding plant operating performance, two general recommendations are apparent.First, future reactordesigns should be more highly standardized
to enhance interreactor learning. Second, operating organizations should be large enough
to populate one or more multiunit sites (within a reasonable amount of time), since this
type of siting increases the benefits derived from interreactor learning. Efforts should also
be taken to improve intersite communication within and between these organizations; the
French system may be a useful model.
Appendix
*
Calculating the effective unavailability. The measure of operating performance used as a dependent variable
in this article is the reactor "equivalent" unavailability factor. Reactor performance is routinely characterized by
the annual "equivalent" availability factor, defined as the total energy that the reactor could have generated during
the course of the year had it been called on to operate continuously at full power, divided by the energy that the
reactor would have produced if it had operated continuously at full power. Equivalent unavailability may be defined
simply as ( 1 - annual equivalent availability factor). However, since we wish to estimate the effects of learning on
performance, the unavailability figures should exclude any outages that are not susceptible to learning-based reductions. The most important source of such outages for light water reactorsis refueling, which requiresthe reactor
to be shut down for some minimum period, and which may or may not take place each year. Of course, learning
may yield reductions in refueling outages too, but some loss of availability is unavoidable. The lower limit for the
durationof these outages depends on the technical requirementsof the refuelingoperationand may also be constrained
by regulatory requirements. Since these limits could not be determined directly, a minimum refueling outage was
estimated for each country and subtracted from the reported unavailability of each unit in that country in each
year that refueling took place. The minimum refueling outage observed in each nation during the sample period
(approximately 21 days) was the basis for these estimates. (When a refueling outage spanned two consecutive years,
the appropriate fractions of this minimum outage period were used to adjust availability figures for each year.)
Although we assume that this minimum duration outage cannot be reduced further by learning, the net addition
to plant unavailability from a refueling outage of longer duration is included in the dependent variable. Like other
types of outages, this net refueling outage is assumed to be susceptible to learning-basedreductions.
In recent years some utilities, motivated by economic considerations, have begun to shift to fuel cycles longer
than the standard 12-month period. By adjusting the annual unavailabilities as described above, we have corrected
for the non-learning-based performance improvements associated with this shift.
0
Data sources. Availability factor data for the U.S. plants were obtained, in part, from the Operating Plant
Evaluation Code (OPEC-2) database, compiled by the S.M. Stoller Corporation initially for the Electric Power
ResearchInstitute(Koppe, Olson, and LeShay, 1984), and more recentlyfor the Instituteof Nuclear Power Operations
(INPO). The OPEC-2 database was made available to MIT by INPO. French availability factor data were provided,
in part, by Electricite de France.
These data were compared for consistency with, and in some cases supplemented by, the performance analysis
reportspublished annually by the InternationalAtomic EnergyAgency (IAEA), OperatingExperience with Nuclear
Power Stations in Member States (Vienna: IAEA, various years). The IAEA reports also provided information on
the incidence of refueling outages.
Data on reactor size, dates of construction start and commercial operation start, and the reactor vendor and
operator in the U.S. case were obtained from several sources: Nuclear Engineering International, July Supplement,
1985; Nuclear News, "World List of Nuclear Power Plants,"August 1986; the Atomic IndustrialForum's "Historical
Profile of U.S. Nuclear Power Development," January 1, 1986; and the annual IAEA volumes.
38 The utility-sponsored Institute of Nuclear Power Operations (INPO) may serve as a vehicle for improved
interreactorlearning. INPO, which was created after the Three Mile Island accident, facilitates communication
among nuclear utilities on issues related to plant safety and reliability.
438
/
THE RAND JOURNAL
OF ECONOMICS
References
ALCHIAN,A. "Reliability of Progress Curves in Airframe Production." Econometrica, Vol. 31 (1963), pp. 679693.
ARROW,K.J. "The Economic Implications of Learning by Doing." Review of Economic Studies, Vol. 29 (1962),
pp. 155-173.
BECKJORD,E., GOLAY, M., GYTOPOULOS,E., HANSEN, K., LESTER,R., AND WINJE, D.K. International Comparison
ofL WR Performance.Report No. MIT-EL 87-004, MassachusettsInstitute of Technology, Energy Laboratory,
1987.
BELSLEY,D.A., KUH, E., AND WELSCH, R.E. Regression Diagnostics: Identifying Influential Data and Sources of
Collinearity. New York: Wiley, 1980.
Bupp, I.C. ANDDERIAN,J.-C. Light Water:How the Nuclear Dream Dissolved. New York: Basic Books, 1978.
DAVID, P.A. AND ROTHWELL,G.S. Performance-Based
Measures of Nuclear Reactor Standardization.
Report No.
247, Center for Economic Policy Research, 1991.
EASTERLING,
R.G. "Statistical Analysis of U.S. Power Plant Capacity Factors through 1979." Energy, Vol. 7
(1982), pp. 253-258.
FREEMAN,
C. The Economics of Industrial Innovation. 2d ed. Cambridge, Mass.: MIT Press, 1982.
GOLDBERGER,
A.S. A Course in Econometrics. Cambridge, Mass.: Harvard University Press, 1991.
HAUSMAN,J.A. "Specification Tests in Econometrics." Econometrica, Vol. 46 (1978), pp. 1251-1272.
AND TAYLOR,W.E. "Panel Data and Unobservable Individual Effects." Econometrica, Vol. 49 (1981),
1377-1398.
HIRSCH,W.Z. "Firm Progress Ratios." Econometrica, Vol. 24 (1956), pp. 136-143.
JOSKOW, P.L. AND ROZANSKI, G.A. "The Effects of Learning by Doing on Nuclear Plant Operating Reliability."
Review of Economics and Statistics, Vol. 61 (1979), pp. 161-168.
ANDSCHMALENSEE,
R. "The Performance of Coal-Burning Electric Generating Units in the United States:
1960-1980." Journal ofApplied Econometrics, Vol. 2 (1987), pp. 85-109.
C. Power Plant Performance. New York: Council on Economic Priorities, 1976.
KOMANOFF,
KOPPE, R.H., OLSON, E.A.J., AND LESHAY, D.W. Nuclear Unit Operating Experience: 1980 Through 1982 Update.
Report No. EPRI NP-3480, Electric Power Research Institute, Palo Alto, Calif., 1984.
KRAUTMANN, A.C. AND SOLOW, J. "Economies
of Scale in Nuclear Power Generation."
Southern Economic
Journal, Vol. 55 (1988), pp. 70-85.
LESTER,R.K. "Organization, Structure, and Performance in the U.S. Nuclear Power Industry." Energy Systems
and Policy, Vol. 21 (1986), pp. 335-384.
MCCABE,M.J. "IndustrialStructureand Technological Change in the Nuclear Power Industry."Ph.D. thesis, Sloan
School of Management, Massachusetts Institute of Technology, 1991.
ROBERTS, P.C. AND BURWELL,C.C. The Learning Function in Nuclear Reactor Operation and Its Implications for
Siting Policy. Oak Ridge, Tenn.: Institute for Energy Analysis, 1981.
N. Inside the Black Box: Technology and Economics. Cambridge:Cambridge University Press, 1982.
ROSENBERG,
G. "Utilization and Service: Decomposing Nuclear Reactor Capacity Factors." Resources and Energy,
ROTHWELL,
Vol. 12 (1990), pp. 215-229.
U.S. Energy Information Administration. An Analysis of Nuclear Power Plant Operating Costs. Washington, D.C.,
1988.
. An Analysis of Nuclear Plant Operating Costs: 1991 Update. Washington, D.C., May 1991.
U.S. Office of Technology Assessment. Nuclear Power in an Age of Uncertainty.Washington, D.C.: U.S. Office of
Technology Management, 1984.
E.P. "Safety Initiatives and Emergency Preparednessin the Nuclear Utility Industry."In Proceedings,
WILKINSON,
95th Annual Conference, National Association of Regulatory Utility Commissioners, November 1983.