Research Policy 41 (2012) 770–779

Contents lists available at SciVerse ScienceDirect

Research Policy
journal homepage: www.elsevier.com/locate/respol

Intra-plant diffusion of new technology: Role of productivity in the study of steel

refining furnaces夽
Tsuyoshi Nakamura a , Hiroshi Ohashi b,∗
a
Department of Economics, Tokyo Keizai University, Japan
b
Department of Economics, University of Tokyo, 7-3-1 Hongo, Tokyo, Japan

a r t i c l e i n f o a b s t r a c t

Article history: This paper examines intra-plant diffusion of new technology in the Japanese steel industry. The introduc-
Received 20 September 2010 tion of the basic oxygen furnace (BOF) was the greatest breakthrough in steel refining in the last century.
Received in revised form 31 August 2011 Using unique panel data, the paper estimates total factor productivity by technology type, and associates
Accepted 22 November 2011
the estimates with intra-plant diffusion. The paper finds that intra-plant diffusion accounts for about a
Available online 30 December 2011
half of the industry productivity growth. Large plants are likely to adopt the new technology earlier, but
retain the old technology longer, than their smaller counterparts.
JEL classification:
© 2011 Elsevier B.V. All rights reserved.
D24
L61
O14
O33

Keywords:
Intra-plant diffusion
Total factor productivity
Innovation
Technological change

1. Introduction rate of intra-firm diffusion—the rate at which a particular firm sub-

Diffusion of new technology has been viewed as the main driv-
ing force of economic growth. An important set of questions often
raised in the literature concerns what factors determine a firm’s
decision to adopt a new technology. While this issue of inter-firm
technology diffusion has been extensively studied, the adoption
of new technology is not in and of itself sufficient for economic
growth.1 For the social benefits of innovation to be realized, the
outcome of an innovation must not only be adopted by a firm,
but also be extensively utilized in economic activities. Productiv-
ity and outputs would not increase in response to the adoption of
new technology, if the utilization of the technology remains low. As
Mansfield (1963: 356) explains, the accurate measurement of the
夽 We thank Shigeru Asaba, Michele Boldrin, Hiroyuki Chuma, and seminar and
conference participants at the BWIE, EARIE, the International Schumpeter Society
Conference, the TEPCO, and the University of Tokyo for their helpful comments. We
also thank useful comments from anonymous referees and the Editor.
∗ Corresponding Author: Tel.: +81 1 3 5841 5511; fax: +81 1 3 5841 5521.
E-mail address: ohashi@e.u-tokyo.ac.jp (H. Ohashi).
1
Although the data used here refer to plants rather than firms, we use the terms
“plant” and “firm” interchangeably, so as to conform to current usage in the litera-
ture.
stitutes a new technology for old in its production process—requires
firm-level data that identify how capital is utilized by technology
type.
Using unique plant-level panel data, this paper analyzes the
role of productivity in intra-plant diffusion, which has received
little attention in previous empirical examinations. In particular,
we focus on the refining furnace technology in the Japanese steel
industry. In the 1950s and 1960s, many integrated steel makers
updated their technology, shifting from the conventional open-
hearth furnace (OHF) to the imported basic oxygen furnace (BOF).
The introduction of the BOF was praised as “unquestionably one of
the greatest technological breakthroughs in the steel industry dur-
ing the twentieth century” (Hogan, 1971: 1543). Interestingly, the
period of the rapid dissemination of BOF technology coincides with
that of the remarkable growth Japan experienced in the wake of the
devastation wreaked by World War II. The steel industry expanded
its production more than fourfold between 1953 and 1964, mak-
ing Japan the world’s largest steel exporter by 1969. As we discuss
in Section 2, intra-plant diffusion played a major role in BOF dif-
fusion, resulting in substantial industry growth in the 1950s and
1960s. Restricting our study to examining refining furnace tech-
nology allows us to abstract from market structure effects in our
study; virtually all steel plants faced the same market for crude

0048-7333/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.respol.2011.11.002
 T. Nakamura, H. Ohashi / Research Policy 41 (2012) 770–779
steel, a homogeneous product manufactured from the refining fur-
naces. The nature of the market, along with the output data by
technology type, allows our analysis to focus on the influence of
other determinants of intra-plant technology diffusion. Our unique
furnace/plant-level data set covers the inputs and outputs from
each furnace type, and the timing and size of new capital instal-
lation. The data permit an estimation of the production function
based on furnace technology and the measurement of the change
in productivity and output growth in the intra-plant diffusion of
new technology.
Our estimation results indicate that intra-plant diffusion makes
a significant contribution to the industry-level productivity growth,
and accounts for more than 70% of the diffusion of the new-
technology in terms of industry production capacity. Furthermore,
the estimates indicate that differences in the productivities of new
and old technologies owned by a plant is negatively correlated
with the rate of intra-plant diffusion; if a new technology is more
productive than an old one within a plant, the plant will shift its
production process from the old to the new technology faster than
it would otherwise, so as to minimize the opportunity cost of retain-
ing the old technology. The paper also observes that large plants are
likely to adopt the new technology earlier than their smaller coun-
terparts. This finding is consistent with those found in the literature
on inter-firm diffusion such as in Rose and Joskow (1990).
In his survey of the literature on new technology diffusion,
Geroski (2000) identifies two leading models: the epidemic and
probit models. The first model, originally proposed by Mansfield
(1963), predicts that the extent of use of a new technology within
a plant increases with the number of years since the first adoption.
Fig. 1 traces the intra-plant diffusion rate of BOF, i.e., the changes
in the share of BOF in a plant’s total capacity size, for each of all
thirteen plants considered in the paper. Note that they are those
that switched from OHF to BOF. Although the BOF share gener-
ally increased over the study period, the epidemic model does not
fully explain the BOF use observed in Fig. 1; the years elapsed since
the first BOF adoption, with the use of a third-order polynomial
of the variable, explain about ten percent of the total variability
of the BOF output share, a finding similar to that of Battisti and
Stoneman (2005). In the empirical implementation on intra-plant
diffusion, along with explanatory variables that are considered as
proxies for the epidemic effect, we incorporate variables that fea-
ture the alternative model—the probit model, which presumes that
differences in the diffusion rates reflect differences in firm and
technology characteristics. Estimation of the model indicates a dif-
ference in the productivity of new and old technologies across
plants, an important determinant of the intra-firm diffusion of new
technology.
The rest of the paper is organized as follows. Section 2 pro-
vides an overview of the Japanese steel market after the Second
World War. It describes several important features of the market
that have a direct bearing on the formulation of empirical strate-
gies and on the interpretation of quantitative results discussed in
the subsequent sections. Section 3 describes our data sources and
presents a method for estimating productivity of furnace tech-
nologies. The panel feature of our dataset enables us to correct
for endogeneity problems when measuring productivity. Using the
obtained productivity estimates, the section evaluates the extent to
which the intra-plant diffusion contributed to productivity growth.
Section 4 quantitatively examines the forces that drive the intra-
plant diffusion pattern observed in Fig. 1. The analysis reveals
that productivity difference between old and new technologies is
an important vehicle in intra-plant diffusion of new technology.
Section 5 discusses the relationship between productivity and both
inter- and intra-plant diffusion patterns of new technology. Section
6 presents our conclusions, followed by the appendices on the data
sources and estimation method used in the paper.
771
2. Overview of the post-war Japanese steel market
In the early 1950s, most Japanese steel was produced by inte-
grated steel manufacturers. Integrated steel works transform raw
materials (iron ore and coking coal) into pig iron in a blast fur-
nace. Pig iron is subsequently transformed into crude steel in a
second furnace by removing carbon and other elements. The preva-
lent technology used in this second or “refining” stage was the OHF,
which blows burning fuel gas over the molten pig iron, which pro-
vides the heat required to purify the pig iron. In the late 1950s,
the OHF began rapidly losing ground to the BOF. This new technol-
ogy blows oxygen to oxidize the iron, making it possible for steel
makers to refine molten iron and scrap charge into steel in approxi-
mately 45 min—a sharp decrease from the 6 h that the OHF normally
required then. Since molten pig iron is a key input for the BOF and
the pig iron is made by blast furnaces, we focus our attention to
plants owning blast furnaces as well as OHFs. At the time the BOF
was invented, there were thirteen plants owned by nine firms, uti-
lizing blast furnaces as well as OHFs, all of which have shifted their
entire steel refining technology from OHF to BOF by early 1970s.
Our data set is derived from these thirteen plants.
Invented in Austria, BOF technology was further developed by
Japanese steel makers after being imported to Japan. The Japanese
have been responsible for the two most important improvements
in the BOF hardware: the multi-hole lance and the oxygen con-
verter gas recovery system (OG system) (Lynn, 1982: 34; Odagiri
and Goto, 1996: 149). The multi-hole lance reduces splashing in the
BOF, thus increasing steel-making yield and improving refractory
life. Over the course of our study period, the BOF lance continu-
ously improved its capability for softer blowing at lower velocities
while achieving higher production rates. The OG system allows the
recovery of gases from the BOF. It controls pollution and helps
reduce energy costs, while contributing to steel-making yield.
These “user-centered technological improvements” (von Hippel,
2005) associated with the BOF are known to have contributed to
the increase in steel-making productivity in Japan. In Section 3, we
observe the effects of these user-side technological innovations on
the process of intra-plant diffusion. 2
Fig. 2 illustrates the diffusion of the new technology observed
from the dataset. Three BOF diffusion paths are plotted in the figure:
overall diffusion is the BOF share in the industry’s total capacity
size; inter-plant diffusion is defined by the percentage of plants that
installed at least one BOF out of the population of 13 plants; and
intra-plant diffusion is the annual average across all thirteen intra-
plant diffusion patterns shown in Fig. 1. The inter-plant diffusion
indicates that all plants represented in the data had adopted the
BOF by 1965, at which time within-plant technology penetration
had reached approximately 30%, and intra-plant diffusion became
the sole driving force of the overall diffusion. Interestingly, it was
between 1965 and 1970 that the Japanese steel industry doubled
its output. The figure thus illustrates the importance of intra-plant
diffusion in the later stages of the diffusion process. This finding
has also been observed with regard to other technologies, including
computer numerically controlled (CNC) machine tools as reported
in Battisti and Stoneman (2004).
Industry circles have recognized that producing steel involves
substantial learning from current and previous production. Hogan
(1971) and Lynn (1982) both noted that it was only through
extensive furnace use that detailed knowledge of furnace oper-
ation was gained. Both OHF and BOF refining furnaces cannot
be operated without skilled workers. It was the experience and
judgment of skilled workers that made it possible for plants to
2
This paper does not consider the electric furnace (EF), because its share in pro-
duction was small during our study period.
772 T. Nakamura, H. Ohashi / Research Policy 41 (2012) 770–779

100%

80%

60%

40%

20%

0%
Years 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971
No. BOF plants 1 2 2 4 7 8 10 11 13 13 13 13 13 13 13
No. OHF plants 13 13 13 13 13 13 13 13 12 10 9 8 6 3 0

Fig. 1. Diffusion of BOF capacity size (13 plants from 1957 to 1971).

adjust the frequency and size of furnace operations, while main-
taining the quality and durability of the crude steel produced. While
Ichniowski and Shaw (1999) found that job rotation was a marked
feature of Japanese steel-finishing lines in the 1980s and 1990s, by
the time of our study period, Mori (2006) discovered in the archives
that rotation was already rarely observed in the operation of refin-
ing furnaces in Japan. This finding underscores the importance of
experience in the refining stage of Japanese steel production.
Theoretical and empirical research informs us that firm size
plays an important role in the diffusion of new technology, and
casual observation of our data indeed reveals a clear relationship
between plant size and intra-plant penetration of the BOF. Fig. 3
plots the year in which the first BOF was adopted and the year in
which the last OHF was terminated from use for each steel refin-
ing plant. The adoption and termination years are sorted by plant
size, as measured by the logarithmic number of workers in 1968.
The number of workers evaluated at a different year of the study
period has little impact to the results discussed here. The figure
contains two important observations. First, a negative correlation
is observed between plant size and the year of new technology
adoption, and that larger plants tended to adopt the BOF earlier.
This observation, which concerns inter-plant technology diffusion,
is well documented in the current body of literature. For exam-
ple, Rose and Joskow (1990) reports that larger firms adopt new
technologies earlier than smaller ones in the electric utility indus-
try. Second, a negative relationship is observed between plant size
and the rate of intra-plant diffusion of the BOF. The figure indi-
cates that the smallest plant needed four years to fully replace the
OHF, whereas the largest plant took twelve years. The correlation
between replacement speed and plant size is large enough to gener-
ate a negative correlation between plant size and the year in which
the OHF ceased to be used.
While the first observation regarding inter-plant diffusion has
been extensively studied, the second one has not received signifi-
cant attention. To address this imbalance, the econometric analysis
presented in Section 3 concentrates on analyzing the second
observation. Note, however, that our empirical analysis uses evi-
dence pertaining to inter-plant diffusion, when controlling for

(%)
100

80

Inter-plant Diffusion
(Percentage of plants
60
installed BOF)

40
Intra-plant Diffusion
(Average BOF share in the
total capacity across plants
20

Overall Diffusion
(BOF share in the total
capacity at the industry level)
0
1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971

Fig. 2. BOF Diffusions patters, 1957–1971.

 T. Nakamura, H. Ohashi / Research Policy 41 (2012) 770–779
Fig. 3. Relationship between plant size and the rate of intra-firm diffusion.
selection issues. The econometric analysis described in Section 4
reports that productivity differences between furnace technologies
account for the intra-plant diffusion of the BOF. The next section
describes the method used to estimate the productivity of furnace
technology.
3. Measuring productivity
This section presents the method used to estimate total factor
productivity (TFP) of furnace technology of the steel refining pro-
cess at the plant level. We use thirteen plants owned by nine firms in
the estimation, whose BOF capacity sizes are presented in Fig. 1. Our
data set consists of annual data from 1957 to 1970. The Japan Steel
Federation (1955–1970) provides data on the amount of outputs
and inputs at the plant level by technology type. Data sources are
detailed in Appendix B. The OHF and BOF both produce crude steel,
a homogeneous product. Our econometric model of the produc-
tion function describes how efficiently the furnaces completed the
transformation process. We use the following Cobb-Douglas equa-
tion estimating the parameters, ˇx , ˇk , and ˇz (all are in logarithmic
form):
s
yi,t s
= xi,t s
ˇx + ki,t s
ˇk + zi,t ˇz + usi,t ,
s is the annual output (in tons and logarithm) for furnace
where yi,t
s (either OHF or BOF) at plant i in year t. In (1), we use plant as the
unit of analysis, and abstract issues of multi-plant operation.
The production function comprises of a number of input vari-
s includes electricity and labor along with a constant
ables. Vector xi,t
term. All furnaces use electricity as an energy source. The labor vari-
able is in units of man-hours, and we multiply the average hours
worked at the industry level by the number of workers employed
at each plant in the data set.
s . The age of fur-
The capacity size of furnace s is indicated by ki,t
nace s at plant i (i.e., the number of years for which furnace s was
used at plant i) is denoted by zi,ts . The last variable captures two
aspects of capital utilization: on one hand, this variable reflects the
experience level, i.e., the extent to which extensive use of a par-
ticular furnace type leads to more efficient production; and on the
other, it also indicates the degree of capital depreciation, as fur-
nace productivity deteriorates with age. The estimated coefficient
of the variable implies which of the two effects dominates in our
773
application. Considering that the two furnace technologies, OHF
and BOF, exhibit different operational characteristics, we allow for
ˇz to differ in terms of technology type.
Apart from the three factors described in (1), two other impor-
tant influences on steel production are plant-level efficiency of
production management and improved furnace technologies. Such
unmeasured determinants are represented by usi,t . Productivity
unobserved by the econometrician may create endogeneity in input
choice. Endogeneity in input choice arises when producers adjust
the amount of material (electricity and labor in our application)
according to their efficiency differences in usi,t . For example, plants
that are perceived to have higher productivity might use more
of electricity. Our first response to the endogeneity problem is to
use plant-, year-, and technology-specific components in the esti-
mation. Further, we allow the fixed technology dummy to differ
according to the year, as follows: usi,t = i + ts + εsi,t , where εsi,t is
a mean-zero error. The fixed plant component,  i , deals with effi-
ciency differences between plants, which do not change over time.
The inclusion of ts serves to control for the differences in furnace
technologies, which change according to the year. This fixed-effect
specification, however, may appear to be restrictive in that a pro-
(1) ductivity difference known to the firm is constant over time. In
an alternative specification, we follow the generalized method of
moments (GMM) approach developed by Blundell and Bond (1998).
Details of this approach are presented in Appendix A.1.
Table 1 presents three results based on methods without fixed
effects (1-A), those with fixed effects (1-B), and GMM (1-C).
The upper part of the table presents estimates of the regression
coefficients. Our inference is based on heteroskedasticity-robust
standard errors. Note that estimates of constant parameters are not
reported in (1-C), since the estimated equation is quasi-differenced
as noted in Appendix A.1. The reported statistics indicates that
our estimation model performs moderately well: the measures of
adjusted R2 for specifications (1-A) and (1-B) are above 0.8, and
the Sargan test statistic for (1-C) would not allow us to reject the
orthogonality condition between some of the instruments and the
error. Note that Blundell and Bond (1998) propose to use the lagged
explanatory variables for xis , kis and zis as a set of the instruments in
the estimation of (1). The estimation method creates a concern for
endogeneity in the estimates, if εsi is serially correlated. In order
to verify the significance of this concern, we perform the Arellano
774 T. Nakamura, H. Ohashi / Research Policy 41 (2012) 770–779

Table 1
Production function estimates.

(1-A) OLS (1-B) Fixed effects (1-C) GMM

Est. Std. Err. Est. Std. Err. Est. Std. Err.

Labor 0.633a 0.077 0.647a 0.120 0.897a 0.085

Electricity 0.169a 0.034 0.132 0.098 0.141b 0.067
Capacity 0.387a 0.040 0.198c 0.100 0.212b 0.094
OHF age 0.158a 0.040 −0.075 0.352 0.060 0.060
BOF age 0.024 0.059 0.060 0.072 −0.074 0.094
Constant 3.598b 1.494
OHF dummy 1.563a 0.605
BOF dummy 2.852a 0.700

Sargan statistic 117.50(109)

R-squared measure 0.9996 0.8631 –
No. observations 229 229 229

Arellano-bond test for zero autocorrelation

Statistics p-value

Order 1 −2.661 0.008

Order 2 1.329 0.184
Order 3 −1.170 0.242

Notes: The data contain thirteen plants considered in the data. Heteroskedasticity-robust standard errors are used in the table. The variables of labor, electricity, capacity,
OHF age and BOF age are all in logarithm. The estimated coefficients on technology- and year-specific dummies are not reported in the table.
a
Signfincance at the 99% confidence level.
b
Signfincance at the 95% confidence level.
c
Signfincance at the 90% confidence level.

and Bond (1991) test for the existence of serial correlation in the
estimated error. As reported in the lower part of Table 1, the corre-
lations are estimated to be statistically insignificant up to the third
order. This result may be understandable in that we already control
for year-specific component by technology type in the estimation.
The table shows that the input coefficients are estimated to be
positive and many of them are statistically different from zero.
While the estimates found in (1-B) are similar to those in (1-A),
we are concerned that endogeneity in input choice may lead to
a correlation between the inputs (labor and electricity) and the
unobserved productivity error that varies over time. If input con-
sumption was readily adjusted to productivity, the resulting bias
in the inputs coefficients could be severe. The GMM estimator
reported in (1-C) accounts for this bias. While both estimates of the
input coefficients in (1-C) are not statistically different from those
found in (1-B), their mean values are estimated to be approximately
40% larger than those in (1-B). Although it is generally impossible
in a multivariate context to sign the bias of the fixed-effects esti-
mates when simultaneity exists and there are many inputs, the
finding from the inputs estimates are consistent with the hypothe-
sis that input’s correlation with productivity shock is much smaller
than either the capacity’s or furnace-age’s correlations with the
productivity.3 The coefficient of the capacity-size variable is less
than one, which may indicate the existence of decreasing returns to
scale. The furnace-age variable is found to be statistically insignifi-
cant.
Based on the estimation results shown in Table 1, we first
aggregate industry productivity for the 1957–1968 period. Our
productivity measure comprises the contribution of furnace age
(represented by zi,t s ) and of disembodied technical progress
s
(represented by ui,t ). Fig. 4 shows annual changes in the aggre-
gated industry productivity. Productivity at the industry level is
3
This hypothesis is corroborated by the observation that the plant-level labor
and electricity inputs did not vary much over time. The positive correlation between
the inputs and capital size (or furnace age) provides an additional support for the
downward bias of the FE inputs estimates.
calculated annually as the share-weighted average of furnace and
plant productivity, defined as:
  
s s s  s 
i,t wi,t yi,t − xi,t ˇx − ki,t ˇk , (2)
i s
s is the output share of technology s (either OHF or BOF)
in which i,t
at plant i at year t, and wi,t is plant i’s share in the total industry
BOF takes the value of zero
output of crude steel at year t. Note that i,t
before year t when plant i adopts BOF. Fig. 4 uses the GMM estimates
under (1-C) in Table 1. The productivity increased at an annual rate
of 14.5% until 1965 when all thirteen plants in the data installed at
least one BOF. Thereafter, the productivity grew by 11.6%.
The figure decomposes the productivity growth into two fac-
tors; one is attributed to inter-plant diffusion, and the other to
intra-plant diffusion. We calculate the productivity contribution of
BOF
inter-plant diffusion by using (2) under the assumption that i,t
BOF for t ≥ t , where t is the year when plant
takes the value of i,t
i
i
i
i adopted BOF. The difference between (2) and the productivity
contribution of inter-plant diffusion calculated above is assumed
to come from intra-plant diffusion. Note that the assumptions
made here are likely to understate (or overstate) the productiv-
ity contribution of intra-plant (or inter-plant) diffusion, because
the intra-plant diffusion is assumed to be absent before the inter-
plant diffusion is completed, in contrast to the observations made
in Fig. 2. Fig. 4 shows that the productivity growth after 1964
when the inter-plant diffusion is completed can be solely attributed
to the intra-plant diffusion. The industry productivity associated
with intra-plant diffusion is 13.1%, higher than that of the indus-
try growth. Moreover, this intra-plant diffusion contributes more
than half of the industry-level productivity at the end of our study
period.
One of key determinants of intra-plant diffusion pattern is the
productivity difference between OHF and BOF at the time of adop-
tion. Fig. 5 presents estimated productivity by furnace type, tBOF
and tOHF , over the period from 1957 to 1968. Note that ts is esti-
mated by both technology- and year-dummy variables. The figure
uses the GMM estimates under (1-C), but the other estimates shown
in Table 1 have qualitatively the same features. The TFP estimates
 T. Nakamura, H. Ohashi / Research Policy 41 (2012) 770–779 775

1.6

1.4

1.2

Aributed to Intra-
plant Diffusion
1.0
T 1957
llog TFPt - log TFP

0.8

0.6

Aributed to Inter-
plant Diffusion
0.4

0.2

0.0
1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968

Fig. 4. Decomposition of TFP growth.

2.0

1.5

BOF
1.0

0.5

OHF
0.0

-0.5
1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968

Fig. 5. Productivity by furnace type, averaged over plants.

confirm that BOF was more efficient than OHF. The figure also
indicates that the TFP measures of the two technologies diverged
over time: the productivity of BOF more than tripled over the
study period, while the productivity of OHF was roughly constant.
The productivity increase of BOF could be due to user-centered
innovations (von Hippel, 2005), including the multi-hole lance and
the OG system mentioned in the previous section. It could also be
due to a feature of inter-plant diffusion process: as experience in
the use of the BOF accumulated in adopting firms, some, if not all, of
this experience would spread among non-adopting firms by word-
of-mouth or knowledge spillover. In either case, the late adopters
would benefit from the knowledge transferred from early adopters,
and thus enjoy higher initial productivity when adopting the BOF.
The productivity decline of the old furnace during the later years
in the study period may be primarily attributed to capital depre-
ciation: smaller plants spent less time and effort maintaining and
repairing the OHF prior to adopting the BOF.4 Although the knowl-
edge spillover also possibly affected OHF operation, the figure
appears to indicate that the depreciation effect dominates. Section
5 discusses the role of productivity in diffusion patterns in further
detail.
While identifying the sources of furnace productivity requires
further data collection, the measured productivity presented in this
section implies a negative relationship between plant size and the
4
Data regarding furnace maintenance time and frequency are available for only
one plant in Yawata, then the largest steel maker in Japan. We observed the four OHFs
owned by the plant, and noted that maintenance time and the OHF sizes were clearly
negatively correlated. Since smaller plants tend to own OHFs of smaller capacity
size, this observation is in line with our finding regarding changes in measured OHF
productivity.
776 T. Nakamura, H. Ohashi / Research Policy 41 (2012) 770–779

Table 2
Determinants of rate of intra-firm BOF diffusion.

(2-A) (2-B) (2-C) (2-D)

Explanatory variables (Est.) (Std. Err.) (Est.) (Std. Err.) (Est.) (Std. Err.) (Est.) (Std. Err.)

Plant size −0.427a 0.089 −0.354a 0.088 −0.410a 0.085

Plant size when the size is over 10,000 −0.065 0.264
Plant size when the size lies between 5000 and 10,000 −0.028 0.286
Plant size when the size is below 5000 0.004 0.312
ln(TFPBOF) − ln(TFPOHF) 0.593a 0.157 0.580a 0.157 0.758a 0.232
ln(BOF age) −0.073 0.244 −0.165 0.227 −0.344 0.259 −0.174 0.221
ln(BOF age) squared 0.213c 0.109 0.275b 0.106 0.345a 0.112 0.284a 0.105
Inverse mills ratio −0.320 0.317
Constant 2.185a 0.776 0.938 0.854 1.606c 0.871 −2.090 2.728
Pseudo R-squared 0.189 0.255 0.262 0.274

The number of observations is 101.

Notes: The data contain thirteen plants considered in the data. Tobit Model is used in the estimation. The independent variable is the share of BOF in production capacity (in
ton) by plant. Heteroskedasticity-robust standard errors are used in the table. The variable of BOF age is in logarithm.
a
Signfincance at the 99% confidence level.
b
Signfincance at the 95% confidence level.
c
Signfincance at the 90% confidence level.

rate of intra-plant diffusion. Because early generation BOFs exhibit
lower productivity than later models, it takes more years for early
BOF adopters to replace the old technology. In the next section, we
statistically analyze the role of the measured productivity differ-
ences in intra-plant diffusion.
4. Econometric analysis of intra-plant diffusion
This section investigates economic determinants of intra-plant
diffusion of BOF observed in Fig. 3, considering the relationship
between plant size and the number of years a plant took to replace
the old with the new technologies. For this purpose, we use plant-
level panel data that identify technology type as either OHF or BOF.
We employ, as the indicator of the extent of intra-plant diffusion,
the BOF share in the total capacity size for each plant, presented
logarithmically. Thus, the rate of intra-plant diffusion is analyzed
using the following diffusion equation presented in Battisti and
Stoneman (2005):
 BOF  the right-hand side of (3) is the error term,  i,t , and the parameters
Ki,t
ln = W ln(Wi,t ) + u ui,t + i,t . (3)
OHF + K BOF
Ki,t i,t
s ≡ exp(ks ) (where s is
As defined in the previous section, Ki,t i,t
either OHF or BOF) takes the value of zero when furnace s at plant
i ceases to be used at year t. Battisti and Stoneman (2005) summa-
rize the basis upon which recent theoretical advances in diffusion
analysis have been built using a neo-classical dynamic optimization
procedure.5
We estimate (3) using a Tobit model and deal with possible
selection bias when Ki,t s is zero. The vector, W , includes two
i,t
variables. Plant size is a commonly explored variable in the lit-
erature of technology diffusion. There are a number of channels
through which plant size may influence the likelihood of technol-
ogy adoption: larger plants are considered to be more capable of
and less risk-averse to adopting a new technology before substan-
tial experience has been gained by using it; and the presence of
scale economies may also enable larger plants to reduce costs at a
faster pace through learning by doing in production. To capture the
5
In the analysis of the diffusion of CNC, Battisti and Stoneman (2005) explicitly
include relative user costs of capital for new and old technologies as a determinant
of (3). We have no access to the data on user costs on steel refining furnaces; how-
ever, notice that steel-making furnaces were installed far less frequently than was
CNC. Thus, annual changes in user costs may have had little impact on the diffusion
pattern of steel furnaces.
plant-size effect, we use the number of workers at the plant level in
year t. Note that plant size is highly collinear with the number of old
furnaces owned by the plant, with a correlation coefficient of 0.8.
We thus do not use, as an explanatory variable, the number of OHF
owned by plant, and assume that the plant-size variable captures
the feature that a plant that owned more (or fewer) old furnaces
may be slower (or faster) in installing the same number of new fur-
naces. Finally, we add a plant-level variable in (3) to account for
the possibility of technology leapfrogging. The BOF-age variable is
included to assess how a plant’s experience with the BOF affects the
extent of intra-plant diffusion. We include a squared term of this
age variable, so as to capture the S-shaped BOF diffusion patterns
observed in Fig. 1.
The second term on the right-hand side of (3), ui,t , reflects
difference in unobserved productivity between old and new tech-
nologies shown in Fig. 5. In the estimation, we use the productivity
estimates obtained from (1-C) in Table 1, but using these other esti-
mates has little qualitative impact to our results. The last term on
to be estimated are  W and  u . In the intra-plant diffusion analy-
sis, we employ data from plants that operated both OHF and BOF,
so that the value of the left-hand side of (3) lies within the range
(−∞, 0). We correct for this selectivity of furnace technology by
adopting the Heckit method, the details of which are discussed in
Appendix A.2.
We pool the data from the thirteen plants. Table 2 presents four
estimation results based on the Tobit method. Our inferences are
based on heteroskedasticity-robust standard errors. Specification
(2-A) indicates that a one-percent increase in the number of plant
workers decreases the relative BOF share by less than half a percent.
With the other factors held at the sample averages, the intra-plant
BOF diffusion rate is 48% higher for the smallest plant than for the
largest. The diffusion rate decreases to 38% in (2-B), in which the
variable of productivity difference between the old and new tech-
nologies, ui,t , is included in (3) as an explanatory variable. The
elasticity of the diffusion indicator with respect to the productivity
difference between BOF and OHF is found to be 0.59. Since the larger
plants were subject to smaller productivity differences, the sign of
the estimate is consistent with findings concerning the plant-size
estimate. BOF-age variable estimates indicate that it takes the aver-
age sample plant about 10 years to achieve the intra-plant diffusion
rate of 50%, and 14 years for the BOF to fully penetrate throughout
all plants.
Specification (2-C) corrects for selectivity in technology choice.
In the intra-plant diffusion analysis, we need to consider plants
 T. Nakamura, H. Ohashi / Research Policy 41 (2012) 770–779
Table 3
Probit estimation on BOF operation.
Explanatory variables Est.
Plant size 0.778
Blast furnace size 0.675
Plant age 13.651c
Plant age squared −1.967c
OHF age −7.697
OHF age squared 1.441
#OHFsothers −0.104b
Year trend 1.132a
Year trend squared −0.067a
Constant −27.957a
Pseudo R-squared 0.
The number of observations is 169.
Notes: The data contain thirteen plants considered in the data. This estimation is
used to control for sample selectivity in the estimation of Eq. (3). The variables of
plant size, blast furnace size, plant age, and OHF age are in logarithm.
a
Signfincance at the 99% confidence level.
b
Signfincance at the 95% confidence level.
c
Signfincance at the 90% confidence level.
that simultaneously operated both OHFs and BOFs. This sampling
method, although necessary in our analysis, could generate biased
estimates if there existed a persistent relationship between the dif-
fusion rate and the choice of plants in the sample. This concern
would make the number of years of technology use correlate with
the error in the equation. The probit results obtained in Table 3 pro-
vide an estimate of the expected value of the error in the intra-plant
diffusion,  i,t . We have applied the Heckit correction procedure in
the sample selection, and included the inverse Mills ratio. Includ-
ing this variable and assuming normality in the distribution of the
latent variable, the estimates in (2-C) will be consistent even if
the selected sample is endogenous. The results under (2-C) do not
indicate the problem in the sample selection. The magnitude of dif-
ferences in the estimates between the results from (2-B) and (2-C)
are not significantly different from zero. Thus, we conclude that the
selection problem is not severe, probably because the termination
of OHF use or adoption of BOF is not related to the intra-plant dif-
fusion process. This finding that intra-plant diffusion is statistically
independent from inter-plant diffusion bear a resemblance to those
reported in Battisti and Stoneman (2005) for CNC and Battisti et al.
(2009) for e-business.
The three specifications discussed above assume that the effect
of the number of plant workers on the diffusion rate is the same
for all plant sizes. Specification (2-D) relaxes this assumption, and
allows for the plant-size coefficient to differ by size category. We
include three class size-specific variables: plant size of over ten
thousand workers, between five and ten thousand workers, and
the rest of the plants. The three size variables are all estimated
to be insignificant, and would not reject the linearity assumption
regarding the plant size coefficient that we made in the prior spec-
ifications.
5. Discussion on diffusion pattern
Firm size (or plant size in the paper’s application) is a commonly
explored variable in the empirical literature on technology diffu-
sion. Firm size is relatively easy to observe, and is typically taken
as a proxy for factors such as productivity, ability to assimilate and
exploit new process or economies of scale. In the empirical analy-
sis above, we have focused our attention on the role of productivity
by furnace type at the plant level, and have noted three observa-
tions regarding the diffusion of new technology: (i) large plants
are likely to adopt the new technology earlier than their smaller
counterparts; (ii) the productivity of the old and new technologies
measured at the time of adoption diverged over time; and (iii) an
777
early adopter takes more years to replace the old technology than
a later adopter.
Std. Err. In theory, diffusion paths are generated under the assumption
0.546 that adoption cost or uncertainty of the innovation declines over
0.506 time. For example, Jovanovic and Lach (1989) show that S-shaped
7.777 diffusion arises naturally in an environment in which homogeneous
1.101 agents face the prospect of learning by doing, which is an important
5.464
feature of steel-refining technology discussed in Section 2. Owing to
0.916
0.040 learning by doing, costs decrease as time progresses. These reduc-
0.219 tions in costs are not appropriated by the incumbents but spill over
0.013 to potential adopters of later generation technology. In the model
8.025
described by Jovanovic and Lach (1989), since early adopters have
576
the advantage of higher revenues per unit of output, delaying the
adoption of new technology entails both benefits and costs.
Large plants were likely to adopt the new technology earlier
than their smaller counterparts, partly because they were more
productive plants to begin with. Indeed the largest plant exhibits
28% higher productivity than did the smallest in 1957. This gap in
the productivity between the largest and smallest plants increas-
ingly widened up to 43% by the end of the study period. This
observed adoption pattern has been supported by Caselli (1999),
who argued that skilled biased technology tends to be adopted by
plants with high human capital levels, because skill and technology
are complementary under strong learning-by-doing conditions.
Since plants with more skilled workers are more productive, the
theory implies that productive plants are more likely to adopt the
BOF technology.
Indeed, in his study of Japanese steel plants in the 1950s and
1960s, Yoneyama (1978) reported that skilled workers tend to be
allocated to the BOF operation, while unskilled workers were like-
lier to be assigned to the OHF operation. Since a considerable degree
of knowledge regarding steel production was embodied in the
skilled labor force, it is reasonable to consider that skilled workers
may have found it easier to accustom themselves to the operation
of new furnace technology. The observation made by Yoneyama
(1978) evidences the importance of human capital levels in the
adoption and diffusion of BOF in Japan.
While the estimation results reported in Section 4 indicates that
the productivity difference between the old and new technolo-
gies play an important role in intra-plant diffusion, the presence
of human capital provides another interesting viewpoint in deep-
ening the understanding of intra-plant diffusion pattern. A large
plant is assumed to have a wider range of workers both skilled and
unskilled. Therefore, it must have been easier for a large plant to
promptly adopt BOF to take advantage of its stock of high-level
human capital. Indeed, the productivity of BOF grew at an annual
rate of 7.4% for large plants, and only 4.6% for smaller ones. The
difference in the growth rates of BOF productivity may have been
due to large plants having more highly skilled workers than did
smaller plants. A limitation of this study is that our data did not
include indicators of human capital levels, and as such we were
not able to test the role of worker knowledge in skilled workers
in the intra-plant diffusion patterns. This is an area that could be
examined further by future research.
6. Conclusion
In the Japanese steel industry, the share of output produced
using the new technology was limited even several years after dif-
fusion had taken place. While inter-plant diffusion was the main
driver early in the overall diffusion of BOF, intra-plant diffusion
became the main contributor in later years, accounting for at least
half of industrial productivity growth from 1957 to 1968. This paper
further analyzes the intra-plant diffusion pattern of the new tech-
nology, a topic that has been relatively neglected in the diffusion
literature.
778 T. Nakamura, H. Ohashi / Research Policy 41 (2012) 770–779

By making use of available panel data that capture the adop- s

instruments xi,t−r s
, ki,t−r s , where r = 2 and 3, and 3 < t, and for
and zi,t
tion and use of old and new technologies used in the Japanese s s s
the level equation xi,t−1 − xi,t−2 , ki,t−1 s
− ki,t−2 s , where 3 < t.
and zi,t
steel refining stage, this paper observed that large plants were We also employ year dummy variables and the dummy variables
likely to adopt new technology earlier, but retain old technology that represents ts as valid instruments. The estimation results are
longer, than their smaller counterparts. This finding, not previously shown under (1-C) in Table 1.
remarked in the literature, implies a negative relationship between
plant size and the rate of intra-plant diffusion of the BOF. The esti-
A.2. Correcting for selectivity in diffusion equation
mation results indicate that productivity difference between new
and old technologies play an important role in the pattern of intra-
Let di,t be a binary response for plant i at year t. It takes the
plant diffusion. If the new technology is more productive than the OHF and 0 < K BOF are both
value of 1 when the conditions of 0 < Ki,t i,t
old one, the plant will shift its production process from the old
satisfied; and takes the value of 0 otherwise. Each plant chooses
technology to the new faster than it would otherwise, to minimize
di,t so as to maximize the discounted stream of profits, the reduced
the opportunity cost of retaining the old technology. The paper also
form of which is parameterized as:
reports that larger plants are estimated to be more productive, as
they might have had higher human capital levels. i,t =
i,t ˛ + i,t , (5)
In addition to the above contributions, this paper observed some
other important findings regarding the intra-plant diffusion of the where
i,t contains the constant term and a vector of plant- and
BOF. The results of the regression of intra-plant diffusion (3) indi- year-specific observed characteristics that affect the profitability
cate the importance of learning by doing in the operation of furnace of plant i’s technology adoption at time t. We will discuss these
technology. The estimation results are robust to the presence of variables in
i,t shortly in this appendix. The unobserved factors
sample selection and endogeneity because of the existence of firm- not captured by
i,t are denoted by  i,t , the mean-zero error and ˛
specific uncertainty. represents a vector of the parameters to be estimated. We assume
This study’s findings on intra-plant diffusion have important that the response di,t is binary and that each plant bases its adoption
public policy implications. Analyses of diffusion policy require
knowledge of whether a firm’s realized intra-plant diffusion per-
formance differs from the optimal performance, and of whether
policy interventions addressing the diffusion path actually improve
social welfare (Stoneman, 2001). The paper’s analysis suggests
that diffusion policies could be justified on the grounds that firms
have insufficient information regarding the use of new technol-
i,t , along with the plant-size variable already introduced
ogy. Our estimation results indicated that experience in furnace
operation was an important determinant of intra-plant diffusion of
BOF. Indeed, our analysis showed that approximately 30 percent
of the variation in BOF diffusion could be explained by operational
experience. If this operational experience exhibits externalities that
cannot be fully accounted for by the firms themselves, there must
be a need for public policy regarding intra-plant diffusion. Measur-
ing the magnitude of the externalities that arise from the adoption
and use of BOF would be the next step to understanding the need
i,t the number of old
for public policy addressing technology diffusion.
Appendix A. Estimation methods
This appendix discusses the methods used to estimate the pro-
duction function (1) and the diffusion equation (3), which appear
in Sections 3 and 4, respectively.
A.1. GMM estimation of production function
In the estimation of production function, we employed the
method proposed by Blundell and Bond (1998) for specifi-
cation (1-C). Remember that usi,t = i + i,t + ts + msi,t , where
i,t =
i,t−1 + ei,t , and that mi,t and ei,t are serially uncorrelated and
are the mean zero errors. We quasi-differenced the production
function and estimate the following dynamic representation:
s s s s s s
yi,t =
yi,t−1 + (xi,t −
xi,t−1 )ˇx + (ki,t −
ki,t−1 )ˇk
s s
+ (zi,t −
zi,t−1 )ˇz + (1 −
)i + (ts −
t−1
s
) + si,t ,
where s represents either OHF or BOF, and si,t ≡ ei,t + msi,t −
msi,t−1 . Estimating (4) takes care of the firm-specific component.
Blundell and Bond (1998) suggests that the production function be
estimated by the GMM as a system combining the first differenced
and level equations. For the differenced equation, we use as a set of
decision on the latent variable i,t using a probit model.
The maximum likelihood estimates of (5) are presented in
Table 3 and are used in the intra-plant diffusion estimation when
controlling for endogeneity in technology adoption. Based on the
literature of inter-plant technology adoption (see Geroski, 2000, for
example), we employ the following four types of plant-specific vari-
ables for
in (3).
We include the variable of blast furnace size (in ton, logarithm)
to test whether the presence of a blast furnace affects the propen-
sity of the plant to adopt the BOF. We also include the variables
of plant age and OHF age (in logarithmic forms). The two age vari-
ables may affect the technology choice because older (or younger)
plants have a greater (or lower) likelihood of substituting the BOF
for the old facilities. We add the squared terms of the age variables
in the estimation. Lastly, we incorporate in
furnaces owned by the other plants in the same company at t − 1. A
company may face difficulty in adopting a new technology, if more
workers are familiar with the OHF. Along with these plant vari-
ables, a yearly trend variable and its squared term are included in
the estimation to control for the industry-wide aggregate shocks.
The binary probit model predicts 148 out of 169, or 88%,
of the observations correctly, suggesting a fair fit to the data.
Table 3 shows that the estimated coefficients for plant and blast-
furnace sizes are both positive, but not significantly different from
zero. While the coefficient of the OHF-age variable is statisti-
cally insignificant, the plant-age and yearly trend estimates are
both statistically and economically significant, indicating that the
plants that were most likely to install the BOF have a plant age
of 32 years (because the plant-age variable has a value of 32.13)
in the year 1965 (because the year-trend variable has a value of
1964.5). Finally, the estimate indicates that a plant with more old
furnaces was less likely to adopt the BOF than one with fewer
OHFs.
Appendix B. Data description
(4)
The data used in the paper are plant-level annual data for 13
steel-making plants from 1957 to 1970. These plants are owned
by 9 Japanese steel firms. The data are disaggregated by plant and
furnace type, and cover approximately 95% of total industry steel
production throughout the study period. Most plants in the data
operated more than one furnace in a given year; however, the input
 T. Nakamura, H. Ohashi / Research Policy 41 (2012) 770–779
and output data are aggregated over the multiple furnaces of the
same technology type within each plant. The output data of crude
steel and the input data are both obtained from the Japan Steel
Federation (1955–1970). The estimation uses two inputs variables;
the amount of electricity and labor. The electricity data are obtained
from the same source as the output data.
Data on labor input are constructed from two variables: the
number of workers at the plant level and by furnace type; and the
average work hours at the industry level. The data on the former
variable are disaggregated by furnace type. The latter variable is
obtained from Tekko Shimbunsha (1955–1970). We multiply the
two variables to obtain the variable of labor input, expressed in
terms of total man hours.
The data pertaining to furnace capacity by plant are obtained
from companies’ semiannual financial reports. They identify the
capacity sizes of each of the furnaces located in each plant. The
data recorded the capacity size at the end of year t, and investment
was made only when a new furnace was built. The capacity size of
furnace js of technology s located at plant i in year t is thus expressed
js = (1 − ı)K js
as: Ki,t i,t−1
, where ı is the depreciation rate. The estima-
tion result reported in Section 3 is based on the assumption that
the value of ı equals zero. To check the robustness to the assump-
tion of no capital depreciation, we set the value of ı to be at 0.05 to
re-estimate the model. We find that this assumption makes little
change to the original results. While we know the capacity size of
js over s to make the construc-
individual furnace, we aggregated Ki,t
tion of the variable consistent with those of the others used in the
paper.
Finally, the data on furnace age by plant are from the Ministry
of International Trade and Industry (MITI, 1958, 1961). The data
source provides the year that an individual furnace was installed
at each plant in our data set. We supplement the data with semi-
annual financial reports published by each company to obtain the
year at which each individual furnace was shut down. We take a
simple average of furnace ages over each plant to construct the
779
variable of the number of years for which furnace s was used at
s .
plant i, namely zi,t
References
Arellano, M., Bond, S., 1991. Some tests of specification for panel data: Monte Carlo
evidence and an application to employment equations. Review of Economic
Studies 58 (2), 277–297.
Battisti, G., Stoneman, P., 2004. Inter- and intra-firm effects in the diffusion of new
process technology. Research Policy 32, 1641–1655.
Battisti, G., Stoneman, P., 2005. The intra-firm diffusion of new process technologies.
International Journal of Industrial Organization 23, 1–22.
Battisti, G., Canepa, A., Stoneman, P., 2009. E-business usage across and within firms
in the UK: profitability, externalities and policy. Research Policy 38, 133–143.
Blundell, R., Bond, S., 1998. Initial conditions and moment restrictions in dynamic
panel data models. Journal of Econometrics 87, 115–143.
Caselli, F., 1999. Technological revolutions. American Economic Review 89, 78–102.
Geroski, P.A., 2000. Models of technology diffusion. Research Policy 29, 603–625.
Hogan, W.T., 1971. Economic history of the iron and steel industry in the United
States. Lexington.
Ichniowski, C., Shaw, K., 1999. The effects of resource management systems on eco-
nomic performance: an international comparison of U.S. and Japanese plants.
Management Science 45, 704–721.
Japan Steel Federation (1955–1970), Reference Material on Steel Making, Tokyo.
Jovanovic, B., Lach, S., 1989. Entry, exit, and diffusion with learning by doing. Amer-
ican Economic Review 79, 690–699.
Lynn, L.H., 1982. How Japan Innovates, A Comparison with the U.S. in the Case of
Oxygen Steelmaking. Westview Press, Boulder, Co.
Mansfield, E., 1963. Intra-firm rates of diffusion of an innovation. Review of Eco-
nomics and Statistics, XLV, 348–359.
MITI, 1958, 1961. Current States of Steel Production Facilities. Japan Steel Federation,
Tokyo (in Japanese).
Mori, T., 2006. How Important was Skill in Japanese Manufacturing Firms, mimeo,
Faculty of Economics. University of Tokyo.
Odagiri, H., Goto, A., 1996. Technology and Industrial Development in Japan: Building
Capabilities by Learning Innovation and Public Policy. Oxford University Press.
Rose, N.L., Joskow, P.L., 1990. The Diffusion of New Technologies: Evidence from the
Electric Utility Industry. Rand Journal of Economics 21 (3), 354–373.
Stoneman, P., 2001. The Economics of Technological Diffusion. Blackwell, Oxford.
Tekko Shimbunsha, 1955–1970. Almanac of Steel. Tekko Shimbunsha, Tokyo (in
Japanese).
von Hippel, E., 2005. Democratizing Innovation. MIT Press, Cambridge.
Yoneyama, K., 1978. Innovation and Workplace Management (in Japanese). Boku-
takusha, Tokyo.