SlideShare a Scribd company logo
Does Experience Matter?
 Productivity Effects of ICT in the German Service Sector


                                    Thomas Hempell
            Centre for European Economic Research (ZEW), Mannheim


                                       February 2002



                                           Abstract
        In this paper, it is argued that ICT investment is closely linked with complementary
     innovations and most productive in firms with innovative experience. In an analysis
     based on firm–level panel data covering the period 1994–99, system GMM estimates for
     an extended production function framework reveal significant productivity effects of
     ICT in the German service sector. Moreover, there is strong support for the hypothesis
     that the experience gained from past process innovations is a specific complement that
     makes ICT investment more productive. The results suggest that ICT may have been
     contributing to productivity differentials both between firms and countries.




Keywords: Information and Communication Technologies, Services, Production Func-
tion Estimation, Panel Data

JEL–Classification: C23, D24, O32


Address: Centre for European Economic Research (ZEW)
         Research Group of Information and Communication Technologies
         P.O. Box 10 34 43
         D–68034 Mannheim
         Germany
Phone: +49/621/1235–233
Fax:     +49/621/1235–225
E-Mail: hempell@zew.de
1    Introduction
In spite of the current economic woes and emerging disillusions about a ‘New Economy’,
productivity effects of information and communication technologies (ICT) continue to
play a key role in assessing the prospects and growth potentials of both firms and whole
economies. In fact, the economic downturn currently experienced by some countries
shows that ICT are far from being a panacea that yields permanent growth and the end
of business cycles, as some analysts suggested at the peak of the hype. Rather, there is
growing support for the view that it is the specific economic circumstances that determine
the success of exploiting the productivity potentials induced by new technologies.

     ICT have been compared to other important historical inventions like the steam
engine or the electric motor (David and Wright, 1999). These inventions are interpreted as
‘general purpose technologies’ (GPT) with three main characteristics: they are pervasive,
they carry an inherent potential for technical improvements and they lead to innovational
complementarities and scale economies (Bresnahan and Trajtenberg, 1995). According
to this view, GPT should be understood primarily as ‘enabling technologies’, “opening
up new opportunities rather than offering complete, final solutions” (Bresnahan and
Trajtenberg, 1995:84).

     Various empirical studies seem to support the ‘general purpose’ hypothesis. Microe-
conomic investigations by Brynjolfsson and Hitt (1995; 1996) and Lichtenberg (1995)
found significantly positive contributions of ICT investment to output and productivity
at the firm level. A macroeconomic study by Schreyer (2000) concludes that ICT
investment has fostered output and productivity growth in the OECD countries during
the 90s. Similarly, Colecchia and Schreyer (2001) find evidence that a broad application
of ICT within an economy plays a key role in this context. They conclude that a
rapid diffusion of ICT depends less on the existence of an ICT producing sector but
rather on the flexibility of product and labour markets as well as the business environment.

     The ICT boom at the end of the 90s was accompanied by many exaggerations. The
stock markets soared, the market capitalization of various dot.com firms spectacularly
exceeded the value of several giants of the ‘Old Economy’. Bank credits and venture
capital were given to new firms on the basis of fashionable ideas rather than sound
business plans. In the meanwhile, about two years after the end of the ‘hype’ in 2000,
both managers and economists have become less enthusiastic about the emergence of a
‘New Economy’. On the one hand, numerous of the once highly praised dot.com firms
have declared bankruptcy. Moreover, many firms of the ‘Old Economy’ have not survived
the wave of innovation and increased globalization and competition induced by the new
technologies. On the other hand, some brick–and–mortar firms have been able to use


                                            1
ICT to boost productivity and market power. Obviously, not all firms have been equally
able to benefit from the productivity potentials of ICT.

     More recent studies have tried to explain these differences in the productivity
effects between firms. Brynjolfsson and Hitt (2000) emphasize the role of organizational
changes that are necessary to reap the potential benefits of ICT. Brynjolfsson and Yang
(1999) and Yang and Brynjolfsson (2001) point to the role of intangible assets that
are complementary to the firm’s use of ICT. Bresnahan, Brynjolfsson and Hitt (2001)
find evidence for significant synergies between ICT, workplace reorganization and new
services and products. These studies have an important message in common: in order to
assess the impact of ICT on firm performance it is crucial to investigate the firm–specific
circumstances in which ICT are used. Hence, complementary factors are essential for the
ICT to unfold its productivity impacts as a ‘general purpose technology’.

     The purpose of this paper is to shed more light on the factors that determine
the success of improving a firm’s performance by ICT use. Unlike previous studies
of the topic, the theoretical model proposed in this study stresses the importance of
ICT being part of the innovation process within a firm. With reference to evolutionary
approaches in innovation theory it is argued that firms that have introduced innovations
in the past are better prepared for ICT–induced innovations, like process improvements
and organizational restructuring, than firms without any innovation experience. As a
consequence, the model predicts productivity effects of ICT to be higher in experienced
firms.

     The empirical analysis of this hypothesis is based on a sample of firms from a
representative survey in the German business–related and distribution service sector
covering the period 1994 to 1999. In this regard, the study distinguishes itself from
most previous studies that have concentrated on the manufacturing sector or on samples
of large firms only.1 However, there are several reasons to draw more attention to the
service sector. Firstly, in the service sector, ICT investment is most dynamic and most
extensive in volume (OECD, 2000a; EITO, 2001). Moreover, as pointed out in OECD
(2000b), the service sector — and in particular business–related services — have been the
most important driver of economic growth over the last decades in most industrialized
countries.2 Finally, recent macroeconomic studies — for example by Baily and Lawrence
(2001) and McKinsey Global Institute (2001) — conclude that services have accounted
   1
     See for example Pilat (2001), Brynjolfsson and Yang (1996) and Brynjolfsson and Hitt (2000) for
surveys of the empirical literature.
   2
     In Germany, the fraction of value added contributed by financial services and business services rose
from 24,1% in 1991 to 30,3% in 1999 (Sachverst¨ndigenrat zur Begutachtung der gesamtwirtschaftlichen
                                                a
Entwicklung, 2001).



                                                   2
for a substantial part of productivity acceleration in the U.S. during the 90s. Therefore,
in order to better assess the impact of ICT on the economy as a whole, a more detailed
understanding of the underlying processes in the service sector seems to be particularly
important.

     A further advantageous feature of the sample employed in this study is that it covers
firms of all size classes instead of only large firms. This broader focus is likely to increase
the variation of characteristics and implementation strategies between firms and thereby
helps to identify the role of the firm characteristics. Moreover, the broader focus en-
hances the likelihood of receiving a more representative picture for the economy as a whole.

    Even though the German economy continues to play a key role within the European
Union as well as within the OECD, few attempts have been made to quantitatively assess
the productivity effects of ICT use in German firms. To the knowledge of the author,
the only related study is the paper by Bertschek and Kaiser (2001) who find significant
contributions of ICT investment in a cross–section analysis of business–related services
firms.3 Hence, a further objective of this paper is to shed more light on the impact of
ICT use in the German economy.

     The paper is organized as follows. In section 2, the theoretical background is
discussed and the hypothesis of innovative experience playing a key role in efficient ICT
use is implemented into a production function framework. Section 3 gives an overview
of the data employed, that is the Mannheimer Innovation Panel in the Service Sector
(MIP-S), and describes the methods used to construct different stock values for ICT and
conventional capital separately. Section 4 discusses the econometric issues and presents
the empirical results for both a simple ICT–extended production function framework and
the more specific model about the role of innovative experience. Section 5 concludes with
some comments on the implications of the findings concerning theoretical, methodological
and policy issues.



2    Methodological and Theoretical Framework
Like many other important inventions in history, the rapid technological progress of the
semi-conductor industry and the fast expansion of the internet evoked high expectations
about their impact on productivity and growth. The adoption and diffusion of these
new technologies were expected to boost productivity substantially. However, for a
long time anecdotal evidence of computerized workplaces did not show up in aggregate
  3
    Licht and Moch (1999) use qualitative data to assess the impact of ICT on the performance of German
firms. They find that the main impact of ICT use are quality improvements.


                                                  3
productivity statistics, which led to the proclamation of the “productivity paradox” of
ICT. While investment in ICT started to grow rapidly in the U.S. since the 1970s, labour
productivity growth slowed down substantially and remained at a low level until the
beginning of the 90s (Jorgenson and Stiroh, 1995). The service sector was at the heart
of this paradox. Around 1990, the U.S. service sector accounted for nearly 80 percent of
total IT investment in the U.S. but experienced productivity growth only slightly above
zero (Brynjolfsson and Yang, 1996).

      In this regard, theoretical and empirical approaches that focus on the firm level rather
than aggregate level appear very promising. Theoretically, the productivity impact of ICT
is likely to vary substantially between firms. Like in the case of many other innovation
waves in economic history, some firms are better prepared than others to take productive
advantage of new technologies. For the particular case of ICT, it has been argued that
complementary factors like organizational restructuring and intangible assets play a
key role for ICT to unfold its largest benefits (Brynjolfsson and Hitt, 2000; Yang and
Brynjolfsson, 2001). By treating all firms equally (assuming equal input coefficients in a
production function framework, for example), the true impact of ICT on “well–prepared”
firms may be understated. Therefore, in order to assess the determinants of successfull
ICT use, it may be useful to explicitly model the influence of various firm characteristics
on the efficient use of new technologies.

     A further — rather empirical — source of the productivity paradox is that official
statistics do not adequately reflect quality changes. As pointed out by Brynjolfsson (1994)
and Licht and Moch (1999), quality improvements — in particular improved customer
service — are an especially important goal for ICT investment decisions. Griliches (1994)
shows that the problem of unmeasured quality improvements is especially important
in the case of ‘unmeasurables’ services like trade and F.I.R.E. (finance, insurance, real
estate) where ICT investment has grown most rapidly. As a consequence, the contribution
of ICT to real output growth inferred from aggregate data are likely to be biased towards
zero. For applications to firm–level data, this problem will be less severe. If a firm
invests in ICT in order to improve the quality of a product while its competitors continue
to offer their old products, the innovating firm will try to charge a higher price for its
new product. If the quality improvement is approved by customers, they will in fact
be willing to pay a higher price for the good and the value added of the innovating
firm will increase accordingly.4 Consequently, Brynjolfsson and Hitt (2000) argue that
microeconomic studies will capture this effect and variations in output quality will
contribute to measuring a higher output elasticity of ICT investment. This relationship
   4
     A firm level study by Brynjolfsson and Hitt (1995) did not find any significant differences in IT produc-
tivity between “measurable” and “unmeasurable” sectors, indicating that appropriate quality measurement
is mainly a problem at the aggregate level.


                                                    4
is independent of the price deflator employed for the corresponding industry. Moreover,
firm–level data sets are particularly suitable for empirical assessment of productivity
effects if — apart from input and output data — they contain information on additional
firm characteristics. These data may allow to identify the determinants of success in ICT
adoption. On the contrary, this identification will be much more difficult with aggregate
data since differences between firms will even out by aggregation.

     For these reasons — both theoretical and empirical —, firm level analysis seems to be
the appropriate approach for a more detailed investigation of productivity effects of ICT
investments. In the following first subsection, it is argued on the basis of evolutionary
approaches of innovation that a distinction between firms according to their innovative
experience may play an important role for the benefits of ICT use in services. In the
second subsection, this distinction is formalized within an extended microeconomic
production function framework that is implemented empirically in the following.



2.1    Innovation in Services: The case of evolutionary approaches
In earlier studies on innovation, the service sector has been characterized as a mere
applier of technological innovations developed in the manufacturing sector (see for
example the influential taxonomy by Pavitt, 1984). Recent studies, however, tend to
confirm a more active role of the service sector in the process of technological change
(see Blind et al., 2000; Sirilli and Evangelista, 1998). Though new technologies of the
manufacturing sector play an important role in initiating innovation processes in the
service sector, the restructuring of processes and improvement of existing services often
leads to the development of completely new services (Barras, 1986). Therefore, like
in the manufacturing sector, differences in technological opportunities, appropriability
conditions, and cumulativeness of innovative capabilities may lead to differences in the
innovation paths between firms.

     As pointed out by Cohen and Levinthal (1990, p. 128), “the ability of a firm to
recognize the value of new, external information, assimilate it, and apply it to commercial
ends” is critical to its innovative capabilities. They argue that this ‘absorptive capacity’ is
largely a function of the level of prior related knowledge.5 Firstly, the absorptive capacity
   5
     Building on studies in cognitive and behavioral sciences, Cohen and Levinthal (1990) state that at
the level of individuals, “learning is cumulative, and learning performance is greatest when the object of
learning is related to what is already known” (131). At the level of organizations, for example firms, the
absorptive capacity is not just the sum of absorptive capacities of its individual employees but also depends
on the transfer of knowledge across and within subunits. As a consequence, there is a trade off between
the aim of highly diversified capacities of the individuals in order to recognize the value of various new
external information on the one hand, and the need of specialization or a “shared language” to ensure the
transfer of knowledge within the organization. Therefore, the effect of specialization on the learning effect


                                                     5
accumulated in a particular area in one period will permit a more efficient accumulation
in the next. Secondly, ‘experienced’ firms will be able to better predict the nature and
commercial potential of technological advances in an uncertain environment. Both aspects
of absorptive capacity — cumulativeness and the impact on expectation formation —
“imply that its development is domain–specific and is path– or history–dependent” (136).

     There are further studies that stress evolutionary aspects of innovation in a similar
way. Mansfield (1968) and Stoneman (1983) argue that a firm’s innovative success
enhances its technological opportunities and thereby makes further success more likely.
This ‘success breeds success’ hypothesis is supported by an empirical study by Flaig and
Stadler (1994). They derive a stochastic dynamic optimization model with intertemporal
spill–overs of innovations within firms. In their empirical application, they find that
firms that have introduced innovations in the past are indeed more likely to innovate in
subsequent years.

     Similar to these arguments and findings, innovative experience may be an explanation
for substantial differences in firms’ ability to boost productivity by the application of
ICT. As emphasized by Bresnahan et al. (2001), the use of ICT is closely linked to
innovations within the firm. Case studies show that organizational changes and process
reengineering play a key role as complementary innovational efforts (Brynjolfsson and
Hitt, 2000). Both the ‘absorptive capacity’ and the ‘success breeds success’ hypotheses
suggest that firms that have introduced such innovations in the past are probably better
prepared than non-innovative firms to reap the potential benefits of ICT.

     There are several more specific reasons supporting this view. Firstly, managers
are likely to have learned from past mistakes.6 They are better prepared to assess the
potentials and limits of introducing major changes within their company, being aware of
possible reactions of their employees and their traditional customers. In short, they have
varied experience concerning the optimal speed of innovation of their firm. Secondly,
innovative firms will have more experience in how to implement innovations. In most
cases, the efficient introduction of ICT requires complementary changes of organizational
structures and processes as pointed out by Brynjolfsson and Hitt (2000). Innovative
firms might be more successful in training and motivating their employees to take part
actively in the innovation process. This argument suggests, in fact, that it will mainly
be experience in process and organizational innovations that increase the ability of
at the level of firms is ambigous.
   6
     The returns to innovations are far from being safe. However, the likelihood of a success may increase
with the experience gathered from past innovations. In this regard, the learning process in introducing
innovations may be compared to the search model proposed by Nelson (1982) for the case of R&D. In-
novation is modelled as a search process in which knowledge helps to lower the search costs by focussing
search on more relevant alternatives.


                                                    6
exploiting the productivity potentials of ICT successfully. Thirdly, firms that are more
innovative may have a higher share of employees that accept an intensive use of ICT at
their workplace. This may be due to both special recruitment strategies and to specific
experience and greater flexiblity of the workforce from former innovation activities.7
Finally, innovative firms may have gained some degree of ‘innovative reputation’ or
branding in new business areas. In particular, it might be easier for them to attract new
customers by exploiting ICT–based sales channels (B2B, B2C) or by introducing new
ICT-based services or products.8 Since innovative reputation can most likely be acquired
by offering innovative products, this argument would imply that past product innovations
raise the ability of benefiting from ICT.

     Taken together, these arguments imply that the returns to ICT investment are
higher in innovative firms than in firms with little or no innovative experience. In the
empirical testing of this hypothesis it will be important to distinguish between different
kinds of innovative experience, in particular between product and process innovations. In
the following formal derivation of the model, however, this distinction is abstracted from
and left for the empirical implementation in section 4.



2.2    The Model
In this part, a theoretical framework of production is proposed focussing on two specific
questions raised in the previous part. Firstly, is innovative experience a crucial prereq-
uisite to implement ICT efficiently? And if so, secondly, is ICT capital ‘special’ in the
sense that other capital inputs are much less dependent on innovative experience?

    In order to assess these questions, the traditional Cobb-Douglas production setup is
used as a starting point:

       Yi = F (Ai , Ki , Li ) = Ai Lγ1 Kiγ2
                                    i                                                                     (1)

where Yi is value added of firm i, Ki is conventional (non-ICT) capital, Li is effective
labour and Ai is the multifactor productivity of firm i.9 ICT capital does not enter
   7
     Nelson and Winter (1982) suggest that much of the knowledge of a firm’s organizational routines and
objectives is tacit. They summarize this observation by claiming that organizations ‘remember by doing’.
As a consequence, innovational capabilities are difficult to raise by hiring new personnel. Also in this
regard, the firm’s innovative history plays a key role (see Cohen and Levinthal, 1990).
   8
     Smith and Brynjolfsson (2001) find that brand is an important determinant of consumer choice in
internet transactions. They find that “consumers use brand as a proxy for retailer credibility in non–
contractible aspects of the product and service bundle, such as shipping reliability” (541). Interestingly,
in the study the strongest brand effect was found for amazon, which has gained reputation as one of the
leading internet retailers.
   9
     F (•) may be such that Y exhibits constant returns to scale in K and L (γ1 +γ2 = 1), but not necessarily.



                                                      7
the production function directly, but rather affects multifactor productivity jointly with
innovative experience as a complement:

       Ai = A(ICTi , Ei , mi )                                                             (2)

     where Ai represents firm’s i (time–invariant) multifactor productivity, ICTi rep-
resents its amount of ICT capital, Ei innovative experience and mi collects other
unobserved efficiency parameters of firm i like management skills, location, branding,
                                                                              ∂Ai
etc. Furthermore, the following properties of function A(•) are imposed: (1) ∂ICTi ≥ 0,
      ∂Ai               ∂ 2 Ai
(2)   ∂Ei   ≥ 0, (3)   ∂ICTi2
                                 ≤ 0. That is, a firm’s productivity depends positively on both
ICT input (1) and innovative experience (2). Marginal contributions to productivity are
assumed to decrease with the amount of ICT employed (3). That is, the productivity
potentials of ICT are limited. Finally — and most importantly — it is predicted
that the potential benefits from ICT are higher for firms with innovative experience.
More specifically, the productivity contributions of ICT are increasing in the degree of
                               ∂2 i
innovative experience: (4) ∂ICTA∂Ei ≥ 0. A very convenient functional form satisfying
                                  i
these properties is the following specification:


                                 x(Ei )
       Ai = C · mi · ICTi                                                                  (3)

     with C as a common scale factor and x(•) being a strictly monotone function of Ei
such that ∂x(•) > 0 and x(•) ∈ [0, 1]. Past innovation activities are considered as a proxy
           ∂Ei
of a firm’s innovative experience such that Ei = 1 if firm i has been an innovator and
Ei = 0 otherwise. Denoting x = x(1) and x(0) = x, it follows that x = x + ∆x with
∆x > 0. Then equation (3) becomes:

       Ai = C · mi · ICT x+∆x·Ei                                                           (4)

    Inserting equation (4) into (1) and taking logs then yields the following extended
production function equation:

       ln Yi = ln C + ln mi + x ln ICTi + ∆x(Ei · ln ICTi ) + γ1 ln Li + γ2 ln Ki          (5)


     Thus, the model corresponds to an ICT–extended Cobb–Douglas function except that
the coefficient of ICT is predicted to be different for experienced firms (x = x + ∆x) and
unexperienced firms (x). One implication of this framework is particularly noteworthy.
The theoretical property that innovative firms exhibit a higher elasticity of output
with respect to ICT (∆x > 0) does not necessarily imply the marginal product of ICT
to be higher in innovating firms since the marginal product depends also on the ICT


                                                   8
intensity (ICTi /Yi ) of the individual firm.10 What a higher elasticity does imply is
that in innovating firms the marginal output contributions of ICT decrease less rapidly
when ICT–input is raised. In the context here, this is equivalent to saying that there
are more potential benefits to be exploited by the use of ICT within innovative firms.
In the extreme case, innovative experience is an essential condition for ICT to yield
benefits at all. If so, the output elasticity of ICT would be zero for non–experienced firms
(x = 0) or even negative11 , whereas experienced firms exhibit positive returns to ICT
capital (∆x > max(0, −x)). Under the assumption that marginal products of ICT (MPI)
are equal across all firms (that is ICT is earning equal returns), the implication is that
                                                                                  x(E)
experienced firms will produce their services with a higher ICT intensity ICT = M P I and
                                                                            Y
a lower intensity in other factors.

         Finally, equation (5) is transformed into the following econometrical model:


          ln Yit = c + ηi + β1 ln Lit + β2 ln ICTit + β3 ln Kit
                     +β4 (ln ICTit · Ei ) + [β5 (ln Kit · Ei )] +   it                                (6)

     with c = ln C, ηi = ln mi , β1 = γ1 , β2 = x, β3 = γ2 , β4 = ∆x and it as a normally
distributed disturbance term. In this framework, the answer to the question whether
innovative experience influences the output contributions of ICT depends crucially on
whether the coefficient β4 turns out to be significantly positive or not. As far as the
second initial hypothesis is concerned, the effect of experience on ICT is contrasted by the
comparable effect on non–ICT capital (see the β5 -term in brackets). If ICT is a ‘special’
capital input by its dependence on past innovations, β5 is expected to be zero and β4 to
be positive in a simultaneous estimation of both coefficients.



3        The Data
In order to implement the production framework empirically, data from the Mannheimer
Innovation Panel in Services (MIP-S) are employed. This survey is conducted by the Cen-
tre of European Economic Research (ZEW) on behalf of the German Federal Ministry for
Education and Research (bmb+f). The data has been being collected annually since 1994
in a representative survey of innovation activities in the German business related-service
and distribution sector and includes information of more than 2000 firms (Janz et al.
    10
    Formally, the marginal returns to ICT (MPI) are the product of the output elasticity of ICT and the
                                                  ∂Yi              Yi
inverse ratio of ICT capital in output: M P Ii = ∂ICTi = x(Ei ) · ICTi . Therefore, the MPI increases with
the output elasticity x(Ei ) but decreases with the share of ICT capital in output ICTi .
                                                                                    Yi
  11
     A negative coefficient would imply that the introduction of ICT leads to efficiency losses, for example
due to high costs of internal restructuring or long lags in ICT causing measurable productivity effects.


                                                    9
2001). It has an (unbalanced) panel structure in important key variables for the years since
1994. The survey methodology is closely related to the guidelines proposed in the Oslo-
Manual on innovation statistics (OECD/Eurostat, 1997). Furthermore, the 1997 wave
of the survey in the service sector formed part of the Community Innovation Survey (CIS).

    For the particular purpose of the empirical analysis, the MIP-S data set contains
annual data on sales, number of employees (full-time equivalents), skill structure,
expenditures for gross investment and for ICT-capital (hardware, software and telecom-
munication technology). Furthermore, firms were asked whether they had introduced a
technologically new or substantially improved products (product innovation) or processes
(process innovation) within the last three years.

     In order to estimate the model equation, some data transformations were necessary.
Most importantly, capital stocks for ICT capital and conventional (non-ICT) capital
were constructed separately. For this purpose, investment on conventional capital was
defined as total investment expenditures minus ICT expenditures.12 In order to derive
the corresponding real investment, conventional investment is deflated by the deflator of
the German Statistical office.13 As far as deflators for ICT goods are concerned, German
official price statistics on ICT goods tend to understate the real price decline of this
product class (Hoffmann, 1998). Therefore, the price deflators for total ICT in the U.S.,
which are calculated on a hedonic basis (see Jorgenson, 2001), are used. Additionally,
these prices were corrected for exchange rate changes, multiplying the U.S. deflators by
the annual averages of the $/DM exchange rates (taken from German Bundesbank, time
series wj5636).

    Given the deflated investments for both types of capital, the perpetual inventory
method with constant, linear depreciation,14 was applied to construct the capital stocks
   12
      Some firms reported investment expenditures in ICT that exceeded total investment (leading to a
negative gross investment in conventional capital). These inconsistencies were most frequent for the years
1994, 1997 and 1998 (6.3%) but almost absent in 1995 and 1996 (0.2%). The most likely explanation seems
to be questionnaire design: for the years 1995 and 1996, the question on total investment was immediately
followed by the question on ICT investment, thereby drawing the respondent’s attention to the consistency
issue. The most likely underlying source for the inconsistent answers seems to be that respondents had not
included ICT investment in total investment expenditures. In particular, expenditures for software may
have been left out. In order to avoid the loss of many observations, the inconsistent cases were interpreted
as having equated total investment with conventional investment only.
   13
      The index of the producer prices, investment goods, from the Statistical Yearbooks with 1996 as the
base year, was employed for this purpose.
   14
      It may be argued that especially for the case of ICT capital it might be more adequate to apply a
vintage model in which computers maintain their productive efficiency over the lifetime of computers (see
for example Jorgenson and Stiroh, 1995). However, the definition of ICT includes a very broad range of
technologies. Furthermore, the length of the time series available is very short. The assumed life cycle
of ICT capital (1/δ2 = 4 years) would exceed the time series available for the vast majority of the firms.



                                                    10
for ICT and non-ICT. Accordingly, capital Kt in period t with investment It−1 is defined
as:

       Kkt = (1 − δk )Kk,t−1 + Ik,t−1                                                                    (7)

     with k = 1 for conventional and k = 2 for ICT capital and investment.

      There are two potential problems in this approach. Firstly, reasonable values for the
depreciation rates of both types of capital have to be found. Secondly, since no information
is available on the level of capital stocks, initial capital stocks have to be constructed for
all individual firms. Therefore, the method proposed by Hall and Mairesse (1995) for the
construction of an R&D stock was followed since the problems in both contexts are very
similar. Under the assumption that investment expenditures in capital good k have grown
at a similar, constant average rate gk in the past for all firms, by backward substitution
equation (7) can be rewritten for period t = 1 (1994) in the following way:15


       Kk1 = Ik0 + (1 − δk )Ik,−1 + (1 − δk )2 Ik,−2 + . . .                                             (8)
                    ∞                             ∞              s
                                                        1 − δk
              =          Ik,−s (1 − δk )s = Ik0
                   s=0                            s=0
                                                        1 + gk
                     Ik1
              =
                   gk + δk

     Constant linear depreciation rates are assumed for conventional capital (δ1 = 0.06)
and ICT capital (δ2 = 0.25) correspondingly. In particular, with δ1 < δ2 it is taken into
account that the fast technological progress in ICT implies more frequent replacement
of ICT inventory than of conventional capital (including buildings and office furniture
among others). In order to derive the initial capital stocks, assumptions about pre-period
growth rates of both type of investments must be made. For non-ICT investment
expenditures, an annual growth rate of approximately 5% (g1 = 0.05) is assumed.16 For
ICT investment, no time series are available for Germany. In order to get a rough idea
of the evolution of ICT investments during the last decades, U.S. data are referred to as
Therefore, there is little gain in trying to exploit the potential advantages of the vintage approach in this
context.
  15
     In fact, the initial value of investment in conventional capital I1,1 was replaced by the average of the
observed values of conventional investment for each firm. With this “smoothing” it was aimed to correct
for cyclical effects which might have affected the estimated capital stock due to different initial years in
the unbalanced panel. The underlying assumption is that long term growth of investment in conventional
capital (g1 = 0.05) is relatively low compared to cyclical variations in this variable. On the contrary, the
first observation on ICT capital was not replaced by the corresponding averages since long-term growth
(g2 = 0.4) rates of ICT investment are more likely to dominate changes that are due to cyclical fluctuations.
  16
     Calculations on capital data provided by M¨ller (1998) show that gross capital stock in German services
                                                  u
has grown on average by 4.8% annually between 1980 and 1991.


                                                        11
a rough guideline. Jorgenson and Stiroh (1995) calculate an average annual growth rate
of 44.3% in real computer investment and of 20.2% for OCAM (office, computing, and
accounting machinery) between 1958 and 1992 for the U.S. Since the share of computers
in OCAM has been steadily rising and is more closely linked to ICT investment as
defined in the context here, an annual pre-period growth rate of — somewhat arbitrarily
— g2 = 0.4 for ICT investment is assumed.17 Since there are time lags between the
installation and productive contribution of capital goods, the capital stock at each
period’s beginning (or at the end of the corresponding fore-period) are taken as measures
for ICT and conventional capital input.

     For labour input, the annual average of the number of the firm’s employees (full-time
equivalences) is used. Moreover, for some firms, data on human capital — based on formal
education — were available. Two particular variables are used to proxy human capital:
the fraction of employees with vocational training (Berufs- or Fachschulabschluss) on the
one hand and the fraction of employees with a university degree including universities of
applied sciences (Hochschul- or Fachhochschulabschluss) on the other. However, there is
a vast amount of item non-responses in these variables, leading to a substantial decrease
(nearly 50%) in the number of firms with sufficient observation for panel analysis.
In the remainder, this data set will be referred to as the “small sample”. Estimates
based on this sample are just used to explore the effects of omitted human capital variables.

     As far as output is concerned, a firm’s value added, deflated by prices at the industry
level, would serve as an appropriate measure for output. However, the empirical analysis
had to rely on information on the firm’s undeflated total sales only since the survey does
not contain further information on intermediate inputs, which would allow to derive a
firm’s value added. Furthermore, there are no official deflators available for output in the
German service sector. Finally, there is no information on factor utilization available. In
the next chapter, it is shown that under some quite weak assumptions it is still possible to
control for these influences econometrically. In order to make results comparable between
different types of econometric techniques, only firms for which consistent information
on at least three consequent periods were available were included in the sample. The
resulting unbalanced sample consists of 1246 firms with a total of 5355 observations, that
is with an average of 4.3 observed periods per firm and is referred to as the “full sample”
in the remainder.

    Some firms reported a share of ICT investment in total investment expenditures
equal to zero for all the periods surveyed. Since the econometric specification is in logs,
  17
    In fact, later results in the production function estimates turned out to be robust to variations in both
g and d.



                                                     12
these firms must be excluded from the full sample. However, there are reasons to assume
that ICT investment in these firms is not zero, in fact, but rather very low and rounded to
zero by the respondents. In order to prevent potential biases in the results (in particular
in those regressions discriminating between experienced and not experienced firms) the
ICT stock per worker in these firms was assumed to be equal to the corresponding
industry minimum and the corresponding values were imputed. This sample is referred
to as the “extended sample” in the remainder.

      Several variables of the MIP-S data set make it possible to distinguish between firms
with more and less innovative experience. In each wave of the survey, firms were asked
whether they had introduced new or significantly improved services on the one hand and
new processes on the other. These questions refer to the past 3-year-period of each wave.
These variables were used to construct two different classifications of firms in order to
proxy for its innovative experience. According to the first (broader) distinction, a firm
is classified as a “panel product innovator” (PPD) or “panel process innovator” (PPC)
if it had declared itself an innovator in one of the periods surveyed. On the contrary,
according to the second, narrower distinction, a firm is classified as an “experienced
product innovator” or an “experienced process innovator” (EPD or EPC respectively)
only if it has declared itself an innovator in the first period surveyed.18 By this more
restrictive definition, it is intended to focus more strongly on innovation experience as
a history–oriented concept since the first definition also includes firms that have been
introducing innovations in the course of implementing ICT technologies. Therefore, the
broader definition can be viewed rather as the firm’s characteristic of how ICT capital
goods are implemented. By comparing the results for both of these classifications, it can
therefore be concluded if it is really experience that matters for ICT output contribution
or rather the way the new technologies are implemented.

    The statistics of the sample are summarized in the tables in the appendix. Table
4 shows the (pooled) summary statistics of the logs of the variables on input, output
and innovation that are employed in the regressions, and Table 5 gives an overview over
the shares of innovating firms according to the different classifications. Tables 6 and 7
show that the sample reflects industry and size structure of the German distribution
and business–related services fairly well.19 Finally, in table 8, the medians20 of the firms
averages of capital and output intensity are displayed for the full sample.21 The figures
  18
    Note that an “experienced innovator” must also be a “panel innovator”.
  19
    Retail trade is slightly oversampled whereas traffic and postal services as well as software and telecom-
munication are slightly oversampled. As far as firm size is concerned, large firms are oversampled in their
mere number and undersampled in their respective share in sales (see last two columns of table 7)
 20
    The corresponding mean values are substantially higher, since some firms — in particular real estate
— display very high values for both inputs and output per employee.
 21
    The corresponding intensities for the other samples (not reported) are very similar.


                                                    13
indicate that at the median, workplace in services is equipped with ICT capital worth
slightly more than DM 3000, and with non–ICT capital worth about DM 55,000. The
median turnover per employee is DM 200,000. These figures show that — as suggested
in Griliches (1994) — the small share of ICT input (less than 6% of the value of other
capital goods) together with measurement errors may make it difficult to determine the
productivity effects of ICT use. However, the columns 2 and 3 of table 4 indicate that,
in fact, the distinction according to firms’ innovative experience may help to identify
the productivity effects. In fact, the median of the per capita value of ICT stock in
experienced firms (defined according to its narrowest definition) is more than twice as
high than among non–experienced firms whereas sales and conventional capital hardly
differ. These simple summary statistics coincide with the outlined hypotheses: if ICT use
is more productive in experienced firms, these firms will tend to spend a higher amount
on ICT per worker than other firms.



4         Empirical Results
Before the full model from equation 6 is implemented empirically, the data problems
described in the previous section have to be addressed methodologically. Besides,
potential biases from various sources are to be addressed by using different econometric
approaches. These investigations may also shed some light on the potential sources of the
‘productivity paradox’.

     To keep things simple, this empirical section is organized in the following way.
Firstly, econometric issues arising from data characteristics are discussed. In the second
subsection, estimation results abstracting from firms’ innovative experience are discussed
in order to investigate the effects of various potential sources of biases on the estimates.
Finally, the initial hypotheses about the role of ICT for a firm’s exploiting its “experience
benefits” are analyzed in more detail.



4.1         Econometric Issues
One of the main drawbacks in the MIP-S data set is that it contains only information
on undeflated nominal sales as a measure of firms’ output instead of deflated value
added (see section 3). If the deflators of output and the share of value added in
sales did not vary substantially, this lack of data would not cause any substantial
bias: both the deflator and the common share of value added in sales would enter the
constant term in the logarithmic specification.22 However, the share of value added in
    22
         If, furthermore, these factors were constant over time, they could be “differenced away”.


                                                        14
sales varies substantially across industries. Industries that are typically at the end of
the value chain (like wholesale and trade) will exhibit much lower shares than other
service industries that are much less dependent on intermediate goods in quantitative
terms. Moreover, prices have evolved quite differently between industries in Germany
in the period concerned. Prices of telecommunication, for example, are likely to have
declined to a larger extent than prices in other services have. Finally, different cyclical
effects between industries are very likely to have induced different evolvements of in-
tensity in factor utilization between industries over the observed period from 1994 to 1999.

     To control for these potential distortions econometrically, it is assumed that the most
substantial variation in the share of value added, price deflators, and factor utilization
is due to differences (both in levels and temporal change) between industries.23 More
specifically, firms are classified with respect to 7 industries (j = 1 . . . 7)24 , such that
nominal sales Yijt of firm i from industry j at time t can be expressed as:

                Pjt Ujt
       Yijt =           Vit                                                                              (9)
                  Sj

     where Pjt is the price deflator for industry j in year t, Ujt is the corresponding index
of factor utilization, Sj the industry’s average share of value added in sales (equal to one
minus the share of intermediate goods and materials in sales), and Vijt is the value added of
firm i belonging to industry j at time t. By taking logs, the industry- and time-dependent
factors can be separated from V :

                                                             Pjt Ujt
       ln Yijt = Djt + ln Vit        with       Djt = ln                                                (10)
                                                               Sj

     Therefore, by introducing interacted industry and time dummies Djt , the potential
biases induced by prices, business cycles and differences in the share of value added
in sales can be controlled for. In fact, the allowance for interacted dummies might be
particularly important for analyzing the output contribution of ICT. ICT use tends to
have increased especially strongly in industries for which substantial price declines during
the past decades can be observed, such as telecommunications and software. If the price
  23
      Similar, but more restrictive assumptions were made by Lichtenberg (1995). He corrected for differences
in the share of value added in sales between industries by introducing industry dummies on the one hand and
controlled for price movements by introducing common time dummies on the other. However, differences
in price movements and in business cycles between industries cannot be captured with this specification.
In the discussion of the empirical results, the empirical relevance of including interacted dummies will be
considered separately.
   24
      These are (with the corresponding nace-codes in brackets): wholesale trade (51), retail trade (50,
52), transport and postal services (60-63, 64.1), electronic processing and telecommunications (72, 64.2),
consultancies (74.1, 74.4), technical services (73, 74.2, 74.3), and other business-related services (70, 71,
745-748, 90). Since there are no output data available for banking and insurance (only the balance sheet
total and insurance premiums respectively), these industries must be excluded from the analysis.


                                                     15
declines in these industries are not taken into account, the increase in real output for
firms of these industries will be understated. As a consequence, the output contribution
of ICT in general will be underestimated. Furthermore, variations in factor utilization
due to cyclical effects may add substantial ‘noise’ to the residual variation in output.
Measuring the output contributions of ICT, however, has been compared to looking for
”the needle in the haystack” (Griliches, 1994). Even though real quantities of ICT used
by firms have grown dramatically over the last decade, ICT still represents a relatively
small portion of overall inputs. In the U.S., where ICT has diffused much more rapidly
than in Europe, the share of ICT services in GDP has risen from 4.4% in 1994 to a
modest 5.5% in 1999 (Jorgenson, 2001). Griliches (1994) points out that the small share
of ICT combined with poor measurement in output and deflators makes it difficult to
distinguish the output contributions of ICT from stochastic events. Thus, by omitting
industry specific cyclical effects and price deflators, the stochastic noise may be too strong
to identify the contributions of ICT econometrically.25



4.2    Evidence from the ICT–augmented production function
In order to explore the effects of different potential biases, the ICT-augmented production
function is estimated firstly in a simple pooled OLS regression.26 In the regression
equation, interacted industry and time dummies as well as a dummy variable for firms
located in East Germany are included. This last variable is expected to be significantly
negative since the transformation process in the Eastern part of Germany is still lagging
behind in both productivity and wages when compared to West Germany. The results for
the pooled regression are summarized in the first column of table 1. The coefficients of
all three inputs are significantly different from zero at the one percent level. The output
elasticity of labour takes the reasonable value of 61%.27 What is most striking in the
results is that the point estimate of the coefficient of ICT capital (24.2%) exceeds the
coefficient of conventional capital (12.6%).28 Given that the share of the average invest-
  25
     However, the costs of including interacted time and industry dummies are substantial as well. There
are n = J ·T −1 = 7·6−1 = 41 additional variables that must be included additionally, with J representing
the number of industries and T the number of years.
  26
     All estimations were computed with the DPD98 programme developed by Arellano and Bond (1998)
running in GAUSS. For all the results, heteroskedasticity–consistent standard errors are reported.
  27
     Under the assumption of constant returns to scale and perfect competition, the income share of labour
in an economy must equal its labour coefficient in the production function. For the German economy as
a whole, the average share of labour payments in national income between 1994 and 1999 amounted to
72.4% (Statistisches Bundesamt, 2001).
  28
     Similar results have been found in cross section regressions by Bertschek and Kaiser (2001) for a sample
of firms in the business-related service sector taken from a different survey. On the contrary, Brynjolfsson
and Hitt (1995) report point estimates of the coefficients of ICT capital (10.9%) that were only about half
of the value for non–ICT capital (20.9%) in the pooled regression for a sample drawn from the Fortune
500 Manufacturing and Service listings. There are two main reasons for these differing results. First,



                                                     16
ment expenditure in conventional capital exceeds the average IT-investment, and given
that the average of the estimated non-ICT stock per worker exceeds the corresponding
value for ICT capital by a factor close to twenty, these results would imply that firms
heavily underinvest in ICT capital goods. However, a more reasonable explanation is that
the results from both pooled and simple cross–section regressions are biased. There are
five potential sources that will be considered step by step in the empirical exploration:
unobserved heterogeneity between firms, simultaneity of the decisions about inputs and
output, measurement errors in the input variables, autocorrelation of exogenous shocks,
and biases from omitted variables.

     Unobserved heterogeneity may bias the results if the investment strategies of highly
productive firms are systematically different from their less productive competitors within
the same industry.29 In particular, if highly productive firms tend to invest more in
new technologies than firms with low productivity do, the ICT coefficient will be biased
upwards in a pooled or cross section OLS regression.30 Brynjolfsson and Hitt (1995) find
that unobserved heterogeneity may explain as much as half of the productivity effects
attributed to ICT in their pooled regressions.

     In order to control for the firm–specific effects, the within–estimator was used.31 In
the second column of table 1, the corresponding results are summarized. The figures
indicate that once unobserved heterogeneity is controlled for, the output contributions of
both types of capital are no longer significantly different from zero whereas the labour
coefficient rises slightly (67.6%).32 The figures show that all the output contributions
assigned to both types of capital in the pooled regression were in fact due to unobserved
heterogeneity. These results coincide with very similar findings by Black and Lynch (2001)
and Wolf and Zwick (2002) for production function estimates with one type of capital only.

Brynjolfsson and Hitt (1995) referred to a very different sample of large firms only. They drew their
sample from the Fortune 500 listings of both the manufacturing and the service sector which consists of
large firms only. Second, they constructed the estimate of the firms’ ICT capital stock in a very different
manner. While the ICT capital in the study of Bertschek and Kaiser (2001) and this paper ultimately
builds upon ICT investment data, Brynjolfsson and Hitt (1995) used data on the stock of mainframes and
PCs and converted these data into corresponding estimates of their market values. In particular, their
measure of ICT capital is much narrower than the measure employed in this study in that it excludes
factors such as telecommunication hardware, peripherals and software. With this background, the much
higher ICT coefficient in the results presented here comes at no surprise.
  29
     Productivity differences between different industries are captured by the industry dummies.
  30
     In fact, the highly significant first– and second–order autocorrelation in the errors of the pooled
regression indicate that fixed effects exert a strong effect on the results.
  31
     More specifically, the orthogonal deviations from the corresponding firm’s mean of the variables were
employed.
  32
     Since there is no variation in the East dummy over time, this variable is excluded from the within
estimation.



                                                  17
Table 1: Results for the ICT-augmented production function


                                                             Dep. Variable: Sales
 production inputs      pooled       within             GMM        SYS–GMM SYS–GMM (2)        SYS–GMM (3)
 constant              1.230***       -0.005            -0.025         0.615         0.456        0.551*
                        (0.136)      (0.022)           (0.028)        (0.299)       (0.293)      (0.294)
 labour                0.610***     0.676***          0.549***       0.665***      0.675***     0.664***
                        (0.019)      (0.063)           (0.099)        (0.066)       (0.066)      (0.066)
 ICT capital           0.242***       -0.016             0.040       0.075**         0.050        0.041*
                        (0.018)      (0.021)           (0.053)        (0.036)       (0.037)      (0.023)
 non-ICT capital      0.146***        -0.002             0.102       0.233***      0.230***     0.232***
                        (0.014)      (0.072)           (0.234)        (0.043)       (0.044)      (0.042)
 East                 -0.101***         —                  —        -0.396***     -0.397***     -0.404***
                        (0.041)                                       (0.046)       (0.047)      (0.045)
 observations:           5355        4109                4109          5355          5355          5529
 number of firms:         1246        1246                1246          1246          1246          1292
 R-square                0.848       0.315               0.211         0.851         0.848         0.840
 joint significance,
 Wald-test [df]:      6906[4]***   125.5[3]***    35.1[3]***    567.8[4]***    573.1[4]***     601.0[4]***
 Sargan (p–values):       —             —           0.043          0.091          0.073           0.046
 AR(1) of errors        0.000         0.000         0.002          0.001          0.001           0.001
 (p–values)
 AR(2) of errors        0.000        0.362             0.400       0.291         0.279           0.271
 (p–values)

***,**,*=significant on the 1,5 and 10 per cent level
All regressions except SYS-GMM (2) contain industry dummy variables interacted with year dummy
variables. Heteroscedasticity consistent standard errors reported.


     The unplausibly low estimates of the capital coefficients in the within estimates may
be due to a downward bias of the point estimates due to measurement errors as argued
by Griliches and Hausman (1986). Measurement errors, however, are very likely to be
substantial in both types of capital stocks. First, since there is no information available
about the share of expansion investment in total investment expenditures, common
depreciation rates were assumed for all firms. This may have induced a significant
(though presumably not systematic) measurement error into the construction of the
capital stocks. Second, there was no initial value of the capital stocks available in the
data employed here. For the calculation of approximate values for the initial stocks,
however, both the depreciation and the pre–sample growth rates of the capital stocks had
to be assumed equal across firms. Again, deviations from this assumption are very likely
and will add much noise to the calculated values of ICT– and non–ICT stocks. On the
contrary, the measurement errors for labour input will be less severe, even though the
transformation of part–time workers into full–time equivalents may — apart from the
well–known problem of overtime accounting — add some measurement error here as well.


                                                 18
On the other hand, the downward bias due to measurement error may be offset
by a second type of error which is simultaneity. If input and output are determined
simultaneously, exogenous demand shocks result in an increase of both input and output.
In this case, the output elasticities of the corresponding factors would be biased upwards.33

     In order to correct for these two potential biases, the GMM estimator is applied
for the production function in first differences. Similar to the proceeding proposed in
Mairesse and Hall (1996), all three inputs are assumed to be predetermined and the
(log) levels of their lagged values xt−2 , xt−3 , ..., x0 are used to instrument the input in
differences ∆xt = xt − xt−1 .34 The corresponding results in column 3 of table 1 show
that in this specification the point estimates for the capital coefficients increase while the
labour elasticity decreases as compared to the within estimator. These findings indicate
that the measurement error bias in the capital coefficients exceeds the counteracting
simultaneity bias.35 However, the capital coefficients remain insignificant from zero in
this specification. Furthermore, the low p–value of the Sargan test rejects the validity of
the instruments employed in this specification.

     A further estimation issue related to the simultaneity issue is the potential au-
tocorrelation of the errors. If the exogenous shocks (demand shocks, cyclical effects
etc.) are autocorrelated and this effect is not taken into account, the estimates will
not be consistent. Therefore, the p–values are reported for the corresponding AR(1)–
and AR(2)–tests of the errors in the corresponding specification. In the specification in
first differences of the variables, however, the first–order correlation of the errors will be
induced by the data transformation.36 Therefore, the relevant test for equations in first
differences is whether the corresponding errors are AR(2) or not. As shown in table 1,
autocorrelation of the errors can be rejected for all specifications except simple pooled
regression.

     A possible reason for the insignificant capital coefficients found in the GMM
regressions is the small power of the instruments used.37 Blundell and Bond (1998b)
   33
      The simultaneity bias might apply in particular to those factors that can be adjusted rapidly which
is not so much the case for capital stocks. For a simple formal derivation of the origin of the simultaneity
bias, see Griliches and Mairesse (1995).
   34
      This means that the firms’ corresponding fixed effect are eliminated by explaining output growth by
the growth rates of the inputs.
   35
      These findings coincide with similar results in Black and Lynch (2001) for estimates of the production
function with one type of capital only.
   36
      It is easy to see that if the errors it are i.i.d. with variance σ 2 their corresponding first differences
will be AR(1): E(∆ it · ∆ i,t−1 ) = E(( it − i,t−1 )( i,t−1 − i,t−2 )) = −σ 2 .
   37
      Since capital stocks within firms are highly persistent over time, the correlation of the first differences
with the second lag in level is close to zero. Formally, this can be illustrated by assuming Kt being


                                                     19
show that this may result in substantial finite–sample biases when the GMM estimator
in first differences is used. In a specific application to production function estimation,
Blundell and Bond (1998a) argue that the weak instruments will bias the differenced
GMM estimates in the direction of the within group estimation, that is towards zero
in the case of the capital coefficients. They use an (extended) SYS–GMM estimator
in which both the equation in differences is instrumeted by suitably lagged differences
(like in the simple GMM–estimation) and the equation in levels is instrumented by
suitably lagged differences additionally. These two specifications are then estimated
simultaneously. This SYS–GMM estimator was originally proposed by Arellano and
Bover (1995). The additional moment conditions required for the equation in levels are
not very restrictive. As shown by Blundell and Bond (1998b), only weak assumptions
about the initial distribution of the variables used are necessary. In particular, the
joint stationarity of the dependent and the independent variables is a sufficient, yet
not necessary condition for the validity of the moment conditions for the equation in levels.

     The corresponding results for the SYS–GMM estimation are given in column 4 of
table 1. In this specification, all three factor inputs are significantly positive. The output
elasticity of labour amounts to two thirds which is consistent with the share of income
from labour in the aggregate statistics (see footnote 27). The coefficients of ICT and
non–ICT capital amount to 7.5% and 23.3% respectively, which coincides with the income
share from total capital goods of roughly one third. The null–hypothesis of constant
returns to scale (CRS) cannot be rejected at the 1%–level (not reported). A further
very robust result is that East–German firms in services are significantly less productive
than their West–German counterparts. The coefficient of the East–Dummy (roughly
-0.4) implies that the multifactor productivity in East–German firms is still only about
two–thirds of the West–German level. This finding coincides with aggregate statistics on
productivity differentials in Germany. The corresponding Sargan–statistic (p = 0.093)
does not reject the validity of the instruments at the 5%–level. These robust results in-
dicate that there are substantial output contributions of ICT in the German service sector.

    In order to further investigate the sources of potential biases in assessing the
productivity effects of ICT, the effect of allowing for different business cycles and
inflation rates between industries by including interacted time and industry dummies
has also been analyzed. In order to illustrate the importance of this procedure, the
SYS–GMM estimation with simple (not–interacted) time and industry dummies were
AR(1): Kt = ρKt−1 + t with ∼ i.i.d. If Kt is weakly autocorrelated (|ρ|               1 and ρ = 0), the past
levels are correlated with the contemporaneous levels. For the first available instrument Kt−2 , this is:
E(∆Kt · Kt−2 ) = E((Kt − Kt−1 ) · Kt−2 ) = E(Kt · Kt−2 ) − E(Kt−1 · Kt−2 ) = ρ2 − ρ. However, if the
evolution of Kt resembles a random walk (ρ ≈ 1), the correlation between the variable in differences and
its past values in levels will disappear (ρ2 − ρ ≈ 0) and the instruments will therefore turn out to be weak.



                                                     20
repeated. The corresponding results reported in column 5 (“SYS–GMM [2]”) of table 1
show that the coefficient of ICT capital is substantially affected by this change in the
estimation specification. The corresponding point estimate reduces to roughly 5% and,
more importantly, is not even significant at the 10% level. In contrast, the coefficients of
the other explanatory variables do not exhibit any remarkable changes.38 These results
suggest that the inclusion of interacted industry and time dummies is an important
prerequisite for assessing the contributions of ICT correctly. In fact, the impossibility
of controlling for these effects might be an important methodological reason for the
insignificant output contributions of ICT and the ‘productivity paradox’ found in earlier
studies.

     Another source of distortion is considered in the last column of table 1 (SYS–GMM
[3]). The corresponding results are based on exactly the same estimation method as
column 4 (SYS–GMM). However, the sample was extended by those 46 firms that have
reported zero ICT investment for all years observed (‘extended sample’, see section
3). The inclusion of these firms substantially lowers the point estimate for ICT (4.1%)
compared to the values reported for the original sample (SYS–GMM). Moreover, the
ICT coefficient is significantly positive at the 10% level only. These results appear very
reasonable if one considers that firms may differ in their output elasticities. Those firms
with a low output elasticity of ICT are maximizing profits with a lower share of ICT
capital in output; excluding these firms might therefore overstate the ICT coefficient due
to sample selection bias.39

     Finally, an important issue of estimating the productivity effects of ICT in the
production function framework is the potential bias in the estimates from omitted
variables that are potentially complementary to the firm’s use of ICT. In particular,
recent studies find that human capital plays an important role in this regard (Bresnahan
et al., 2001). Furthermore, Brynjolfsson and Yang (1999) argue that the use of ICT
is strongly complementary to intangible assets. On the one hand, ignoring these
complementary factors might lead to an overestimation of the true impacts of ICT
on production if the output contributions of these factors are wrongly assigned to the
mere use of ICT. On the other hand, a firm’s human capital and intangible assets
are likely to be quite persistent. If these complementary assets hardly vary over time,
their effect will not be distinguishable from other factors like management skills etc.
which are controlled for as unobserved heterogeneity between firms. In this case, no dis-
tortions are to be expected from the omission of these variables in the estimation equation.

  38
    Again, the Sargan statistic does not reject the validity of the instruments (p = 0.073).
  39
    The Sargan statistic of the extended sample, however, indicates that the validity of the instruments is
rejected at the 5%–level (p = 0.046).


                                                    21
Table 2: The effects of including human capital

                                                Dep. Variable: log sales
             production inputs             full         small     small w. skill
             constant                    0.615          0.426           0.637
                                        (0.299)        (0.418)         (0.303)
             labour (log)              0.665***       0.653***        0.584***
                                        (0.066)        (0.101)         (0.061)
             ICT capital (log)          0.075**         0.019         0.087**
                                        (0.036)        (0.045)         (0.039)
             non-ICT capital (log)     0.233***       0.244***        0.172***
                                        (0.043)        (0.070)         (0.051)
             East-Germany              -0.396***     -0.364***       -0.363***
                                        (0.046)        (0.063)         (0.058)
             % university                   –             –          0.472***
                                                                       (0.170)
             % vocational                   –             –           0.318***
                                                                       (0.106)
             observations:                5355          2060            2060
             number of firms:              1246           639             639
             R-square                    0.851          0.826           0.835
             joint significance,
             Wald-test [df]:          567.8[4]***   196.6[4]***     364.9[6]***
             Sargan (p-values):          0.091         0.679           0.206

***,**,* = significant at the 1, 5 and 10 per cent level
All regressions are based on SYS-GMM and contain industry dummy variables interacted with year
dummy variables. Heteroscedasticity consistent standard errors reported.


     In order to assess the potential biases from omitting the probably most important
source, human capital, the list of independent variables was extended by the share of
employees with vocational training and with university degree or equivalent correspond-
ingly. As discussed in section 3, the resulting ‘small sample’ consists of only 639 firms.
The results of the corresponding regression are surveyed in table 2. The first column
replicates the “SYS–GMM” results obtained from the large sample (with interacted time
and industry dummies, fourth column of table 1) to facilitate comparisons while the
second column displays the corresponding results for the small sample. The most striking
result is that while the coefficients of labour and non–ICT remain more or less unaffected
by the sample reduction, the point estimate of the ICT coefficient decreases substantially
and becomes insignificant. This is likely to be the result of the information loss due to
the much smaller number of observations. The effect of including the proxies for human
capital in the regression becomes obvious from the results displayed in the third column


                                             22
of table 2. Including the human capital variables reduces the coefficients of both labour
and non–ICT capital but leads to a substantial increase in the estimated ICT elasticity.
These quite astonishing results seem to imply a complementary relationship between
human capital and non–ICT capital rather than ICT input. Most importantly for the
explorative analysis here, however, the omission of human capital obviously induces an
underestimation rather than an overestimation of the productivity impacts of ICT.

     Taken together, the findings of this subsection indicate that unobserved hetero-
geneity, measurement errors and industry–specific time effects may lead to substantial
distortions in assessing the productivity impacts of ICT. Ignoring heterogeneity may
lead to a substantial overestimation of these impacts while measurement error and the
omission of industry–specific cyclical effects and price deflators work in the opposite
direction. The potential upward bias from simultaneity issues is – at least compared to
the measurement error bias – relatively low. When all these effects are controlled for in
an adequate SYS–GMM estimation (column 4 of table 1), reasonable results are obtained
that are consistent with the rough income shares of the different inputs in aggregate
statistics. A further explorative analysis shows that the omission of human capital does
not lead to an over– but rather an underestimation of the productivity impacts of ICT.



4.3    The role of innovative experience
After assessing the methodological robustness of the SYS–GMM estimator for the
production function framework, this estimator is applied to the regression equation 6. In
order to investigate the role of innovative experience for the success in implementing new
technologies, ICT capital interacted with innovator dummies is included in the regression
as an additional variable (see section 2.2). A potential bias in favour of the hypotheses
to be tested may arise by the mere fact that innovators will be more productive than
other firms due to the returns to their (intangible) R&D or innovation capital that is
likely to be higher for firms with innovative experience. In order to control for this effect,
the corresponding innovator dummy was included as a further variable that captures the
direct productivity contribution of innovations.40 The results are based on the extended
sample in order not to exclude firms with potentially low output elasticities of ICT (see
comments in the previous subsection).

     In the first column, of table 3, the results are replicated for the specification in
which a ‘surplus’ of ICT contributions (∆x) is allowed for in firms with experience in
introducing process innovations according to the narrow definition (epz). The most
  40
     Both product and process innovations may lead to an increase in productivity by either raising the
sales due to improved service quality or by lowering input costs due to more efficient processes.



                                                  23
Table 3: The role of innovative experience


                                                         Dep. Variable: sales
 production inputs          (epc)         (epd)          (ppc)         (epc)           (epd)     epc w/o imp
 constant                0.717***        0.495*         0.497*      0.749***          0.616**      0.794***
                          (0.238)        (0.266)        (0.257)      (0.233)          (0.257)       (0.254)
 labour                  0.607***       0.658***       0.670***     0.582***         0.641***      0.611***
                          (0.238)        (0.063)        (0.060)      (0.051)          (0.057)       (0.057)
 ICT capital               0.032          0.012          0.027       0.051**           0.038         0.066
                          (0.022)        (0.025)        (0.029)      (0.022)          (0.026)       (0.037)
 non-ICT capital         0.237***       0.246***       0.248***     0.226***        0.240***       0.230***
                          (0.041)        (0.042)        (0.039)      (0.055)          (0.059)       (0.043)
 innovator                 0.188*         0.087          0.006         0.296           -0.007        0.127
                          (0.098)        (0.101)        (0.124)      (0.223)          (0.135)       (0.110)
 ICT capital              0.087**        0.061*          0.015       0.060*            0.012         0.074
 (innovator)              (0.039)        (0.033)        (0.035)      (0.034)          (0.042)       (0.047)
 non-ICT capital              —             —              —          -0.008           0.018           —
 (innovator)                                                         (0.074)          (0.078)
 East-Germany           -0.390***      -0.396***      -0.403***    -0.368***        -0.397***      -0.383***
                         (0.043)        (0.042)         (0.043)      (0.045)          (0.042)       (0.044)

 observations:             5529           5529           5529          5529           5529            5355
 number of firms:           1292           1292           1292          1292           1292            1246
 R-square                  0.842          0.842          0.841         0.834          0.842           0.852
 joint significance,
 Wald [df]:             1135[6]***    919.4[6]***    901.0[6]***     1098[7]***    1102[7]***      1076[6]***
 Sargan (p-values):       0.085          0.069          0.124          0.046         0.123           0.194

***,**,* = significant at the 1, 5 and 10 per cent level. All regressions are based on SYS-GMM and
contain industry dummy variables interacted with year dummy variables. All variables in logs, except
innovator and East-Germany dummies. Heteroscedasticity consistent standard errors reported. The
definition of innovating firms (‘innovator’) varies between columns according to the abbreviation given in
the top row of each column (see text).


striking result is that the ICT coefficient, in fact, differs significantly between experienced
and not experienced firms. For epz–firms an implicit output elasticity of x = 11.9%
is found.41 that is significantly higher than the coefficient of the non–experienced
(p–value of the difference: 0.024). On the contrary, the corresponding ICT coefficient
for not experienced firms is quite low (3.2%) and insignificant. Interestingly, the dummy
for process innovation experience is highly significant as well. Obviously, beyond the
importance for ICT use there are direct benefits from new processes introduced in the
past as well. The significant coefficient of 0.188 of the epz–dummy implies that, on
  41
     Note that the estimated ICT coefficient for innovators represents the difference between the elasticities
of experienced and non–experienced firms ∆x (see equations 5 and 6). The implicit ICT elasticity of the
experienced firms x = x + ∆x is just the sum of the two ICT coefficients in each regression.



                                                    24
average, firms with process innovation experience are about 20% more productive than
other firms.42

     In the second column of table 3, the results are reported for the corresponding
distinction of firms according to their experience of product innovations (epd). The results
are very similar to those in column 1. However, the coefficient marking the difference in
ICT contributions is significant only at the 10% level and the implicit output elasticity
for epd–firms (7.3%) is considerably smaller than in the regression with the classification
according to process innovation experience (epc).43 Given that more than 70% of the
product innovators are also experienced in process innovation (see table 5), this difference
may be a consequence of the simultaneous process innovation experience rather than the
experience collected in the turn of product innovations. As a first preliminary result,
it may be concluded that past process innovations have the biggest impact on a firm’s
productive use of ICT.

     In column 3, the results for the less restrictive differentiation between firms with some
process innovation (“panel process innovators”) are replicated. As discussed in section 3,
this wider definition captures both firms with innovative experience and firms that have in-
troduced innovations at a later point in time during the period observed. Therefore, apart
from ‘experience’, this classification variable also includes implementation strategies. In
the application of this definition, there is no significant difference between innovators and
other firms regarding the productivity of ICT use. This finding suggests that it is indeed
the experience generated by a firm’s innovation history that facilitates an efficient ICT use.

     Obviously, ICT productivity is higher in firms with experience in process innovation.
But is this a special feature of ICT as opposed to conventional capital? In order to
address this second question, non–ICT capital interacted with the dummy for experienced
process innovators has been added to the regressors in a further specification (col. 4
in table 3). The results show that, in fact, there is a remarkable difference between
both types of capital. While ICT capital — like in column 1 — continues to be more
productive in experienced firms (at the 10%–level however), no such difference can be
observed in the case of conventional capital. In the corresponding specification for product
innovators (column 6), however, there is no significant difference between both types of
firms for either kind of capital. Taken together, these findings indicate that the effects
of innovation experience are obviously very specific with respect to two aspects: it is
mainly experience from process innovations that matters, and it is the productivity of
  42
      Compared to the SYS–GMM results in table 1, the other coefficients are hardly affected by the addi-
tional variables included in the regression, which points to the robustness of the results.
   43
      A further difference to the preceding results is that there are no significant direct productivity effects
resulting from past product innovations.



                                                     25
ICT capital that is essentially affected.

     Finally, the robustness of the findings of the first specification (process innovation
experience) was investigated with respect to changes in the sample. If firms that have
reported zero ICT investments are excluded from the regression (that is if the ‘full’
instead of the ‘extended’ sample is used; see column 6 of table 3), the difference in ICT
productivity between experienced and non–experienced firms becomes only marginally
significant (p=0.110, not reported). However, the point estimates of the coefficients do
not change greatly, indicating that the exclusion of firms with a potentially low ICT
productivity mainly affects the precision of the distinction between the two groups of
firms analyzed.44

     To sum up, the results for the empirical model specified in equation 6 deliver strong
support in favour of the initial hypotheses. Firstly, innovative experience is found to
significantly influence the productivity potentials of ICT use in services. Secondly, this
dependence on a firm’s innovative history apparently is a feature that distinguishes ICT
from other capital goods. Beyond this, the findings suggest that it is experience collected
from past process innovations which is particularly worthwhile for the efficient use of
ICT. This finding is in line with other studies that emphasize the close link between ICT
use on the one hand and organizational change and restructuring of business processes
on the other. Because of this close link, experience from past innovations may reduce
mistakes and will improve expectation formation with regard to the costs and benefits of
ICT–induced changes.




  44
    In some further excercises to validate the robustness of the results, only insignifcant differences between
the ICT coefficients of experienced and not experienced firms were found for the small sample of 639 firms
that includes information on human capital. This result was independent of the actual inclusion of human
capital. Furthermore, in the small sample specification without human capital, the result yielded no
significant output contribution of ICT at all. These findings can be interpreted as an indicator that
the precision loss due to a smaller sample size is an important issue in this empirical framework. Not
surprisingly, the need for a large sample for more precise estimates is most important for the capital inputs
for which measurement error was found to be an important issue.


                                                     26
5    Conclusions
In this paper, the productivity effects of ICT use in the German business–related and
distribution services are analyzed with firm–level data. Using a production function
framework, two types of models are analyzed. Firstly, for the simple Cobb–Douglas
specification with labour and two types of capital as inputs, a highly significant output
elasticity of ICT–capital of 7.5% is found, indicating substantial productivity effects
of ICT in the service sector. These estimates are based on a SYS–GMM estimator
that controls for a variety of potential estimation biases, like unobserved heterogeneity,
simultaneity issues and measurement errors. Secondly, based on a theoretical model, the
production function framework has been extended to allow productivity contributions of
ICT capital to vary between firms. This more detailed analysis reveals that firms that
have introduced process innovations in the past — labelled ‘experienced’ firms — are
especially successful in ICT–use. The output elasticity of ICT in these firms amounts
to about 12% and is significantly higher than for non–experienced firms (3%). On the
contrary, no such difference can be observed for conventional capital. These findings
support the hypotheses developed in this paper which assign ICT the role of a ‘special’
capital input: unlike other capital goods, the productive use of ICT is closely linked
to innovations in general and the re-engineering of processes in particular. Firms with
experience in process innovations are therefore predicted to exploit the potential benefits
of new technologies more successfully than other firms.

     There are several implications of these findings concerning theoretical, methodological
and policy issues. At the theoretical level, the results contribute to a clarification of the
role of ICT as a general purpose technology (GPT). In spite of the diverse uses and the
rapid diffusion of ICT throughout all industries, the productivity effects of ICT are far
from self–enforcing but rather demand an active implementation strategy within firms.
The role of innovative experience found in this paper indicates that the determinants
for the efficient use of ICT are in the range of a firm’s long–term strategies rather than
characteristics that can be changed easiliy in the short term. Innovative experience is
likely to be acquired within years rather than months.

     Furthermore, the role of innovative history found at the micro level may also be
useful to shed more light on the differences of ICT–induced productivity effects found
between countries. In fact, the competitive and innovative business environment in the
U.S. may be one reason that helps explain why the productivity impact of ICT has been
much higher there than in continental Europe. The higher innovation pressure in the
U.S. over the last decades may have led firms to collect much more diverse innovative
experience than the more protected firms in Europe. This may have enabled firms in the
U.S. to recognize the productive value of the wave of ICT–induced innnovations faster


                                            27
and to react more flexibly with an appropriate restructuring and re-engineering of the
internal processes in order to reap high benefits from the use of ICT. In this respect, ICT
may have led to a further widening of the productivity gap between the U.S. and Europe.

     As far as the empirical methodology is concerned, the findings of this paper illustrate
that unobserved heterogeneity, measurement errors and omitted variables (including
differences in price and quality evolutions over time) are important sources of estimation
biases in assessing the impacts of ICT within a production function framework. While
unobserved heterogeneity may result in a substantial overestimation of ICT impacts,
measurement errors forcefully work into the opposite direction and may lead to an
apparent affirmation of the ‘productivity paradox’. Furthermore, as shown in this paper,
particular firm characteristics may play a key role for the potential impact of ICT on
firm performance. The results suggest that firm–level studies form a promising basis of
analysis that may add more detailed insights and further complementarities between ICT
and firm characteristics in future research.

     As far as economic policy is concerned, the findings of this paper point to the
importance of an innovative business environment that is needed to lay the fundamentals
for an efficient use of ICT. New technologies like ICT may be compared to the invention
of a new fertilizer in farming: though its potential uses may be very general and its costs
quite low, a sound climate, a cultivated soil and a gifted farmer will still be needed to
actually increase crop yield. Unlike the case of farming, however, the climate in economics
may be favoured to a large extent by sound policies. The results of this study suggest
that the enhancing of competition and innovation incentives may serve as an important
driver of both the rapid diffusion and a productive use of ICT.




                                            28
References
Arellano, M. and Bond, S. (1998). Dynamic Panel Data Estimation Using DPD98 for
     GAUSS: A Guide for Users. ftp://ftp.cemfi.es/pdf/papers/ma/dpd98.pdf.

Arellano, M. and Bover, O. (1995). Another look at the instrumental variable estimation
     of error-components models, Journal of Econometrics 68: 29–51.

Baily, M. N. and Lawrence, R. Z. (2001). Do We Have a New E-conomy, AEA Papers and
     Proceedings 91(2): 308–312.

Barras, R. (1986). Towards a Theory of Innovation in Services, Research Policy 15: 161–
    173.

Bertschek, I. and Kaiser, U. (2001). Productivity Effects of Organizational Change: Mi-
    croeconometric Evidence, Discussion paper no. 01-32, Center for European Economic
    Research (ZEW).

Black, S. E. and Lynch, L. M. (2001). How to Compete: The Impact of Workplace
    Practices and Information Technology on Productivity, Review of Economics and
    Statistics 83: 435–445.

Blind, K., Czarnitzki, D., Ebling, G., Gottschalk, S., Janz, N. and Niggemann, H. (2000).
     Quellen f¨er Innovationen im Dienstleistungssektor, in N. Janz (ed.), Quellen f¨er
              u                                                                      u
     Innovationen: Analyse der ZEW–Innovationserhebungen 1999 im Verarbeitenden
     Gewerbe und im Dienstleistungssektor, ZEW Dokumentationen, Centre for European
     Economic Research (ZEW), Mannheim, pp. 27–55.

Blundell, R. and Bond, S. (1998a). GMM Estimation with Persistent Panel Data: an
    Application to Production Functions, Working paper series no. w99/4, Institute for
    Fiscal Studies.

Blundell, R. and Bond, S. (1998b). Initial conditions and moment restrictions in dynamic
    panel data models, Journal of Econometrics 87: 115–143.

Bresnahan, T. F. and Trajtenberg, M. (1995). General purpose technologies ’Engines of
    growth’ ?, Journal of Econometrics 65: 83–108.

Bresnahan, T. F., Brynjolfsson, E. and Hitt, L. M. (2001). Information Technology,
    Workplace Organization, and the Demand for Skilled Labor: Firm-level Evidence,
    Forthcoming in the Quarterly Journal of Economics.

Brynjolfsson, E. (1994). Technology’s true payoff, Informationweek, October 10 pp. 34–36.




                                           29
Brynjolfsson, E. and Hitt, L. (1995). Information Technology as a Factor of Production:
    Thr Role of Differences among Firms, Economics of Innovation and New Technology
    3: 183–199.

Brynjolfsson, E. and Hitt, L. (1996). Paradox Lost? Firm-level Evidence on the Returns
    to Information Systems Spending, Management Science 42/4: 541–558.

Brynjolfsson, E. and Hitt, L. M. (2000). Beyond Computation: Information Technol-
    ogy, Organizational Transformation and Business Performance, Journal of Economic
    Perspectives 14: 23–48.

Brynjolfsson, E. and Yang, S. (1996). Information Technology and Productivity: A Review
    of the Literature, Advances in Computers 43: 179–214.

Brynjolfsson, E. and Yang, S. (1999). The Intangible Costs and Benefits of Computer
    Investments: Evidence from the Financial Markets, Proceedings of the international
    conference on information systems, atlanta, georgia, MIT Sloan School.

Cohen, W. M. and Levinthal, D. A. (1990). Absorptive Capacity: A new Perspective on
    Learning and Innovation, Administrative Science Quartely 35: 128–152.

Colecchia, A. and Schreyer, P. (2001). ICT Investment and Economic Growth in the 1990s:
    Is the United States a Unique Case?, STI Working Papers 2001/7, OECD.

David, P. A. and Wright, G. (1999). General purpose technologies and surges in produc-
    tivity: Historical reflections on the future of the ICT revolution , Discussion papers
    in economics and social history (number31), University of Oxford.

EITO (2001). European Information Technology Observatory 2001, EITO, Frank-
   furt/Main.

Flaig, G. and Stadler, M. (1994). Success Breeds Success.The Dynamics of the Innovation
     Process, Empirical Economics 19: 55–68.

Griliches, Z. (1994). Productivity, R&D, and the Data Constraint, American Economic
     Review 84: 1–23.

Griliches, Z. and Hausman, J. A. (1986). Errors in Variables in Panel Data, Journal of
     Econometrics 31: 93–118.

Griliches, Z. and Mairesse, J. (1995). Production Functions: The Search for Identification,
     NBER Working Paper No. 5067, National Bureau of Economic Research.

Hall, B. H. and Mairesse, J. (1995). Exploring the relationship between R&D and pro-
     ductivity in French Manufacturing firms, Journal of Econometrics 65: 263–293.

                                           30
Hoffmann, J. (1998). Problems of Inflation Measurement in Germany, Discussion paper
   no. 01-98, Economic Research Centre of the Deutsche Bundesbank.

Janz, N., Ebling, G., Gottschalk, S. and Niggemann, H. (2001). The Mannheim Innovation
    Panels (MIP and MIP-S) of the Centre for European Economic Research (ZEW),
    Schmollers Jahrbuch 121: 123–129.

Jorgenson, D. W. (2001). Information Technology and the U.S. Economy, American
    Economic Review 91: 1–32.

Jorgenson, D. W. and Stiroh, K. (1995). Computers and Growth, Economic Innovation
    and New Technology 3: 295–316.

Licht, G. and Moch, D. (1999). Innovation and Information Technology in Services, Cana-
     dian Journal of Economics 32: 363–383.

Lichtenberg, F. R. (1995). The Output Contributions of Computer Equipment and Per-
     sonnel: A Firm-Level Analysis, Economics of Innovation and New Technology 3: 201–
     217.

Mairesse, J. and Hall, B. H. (1996). Estimating the Productivity of Research and De-
    velopment in French and United States Manufacturing Firms: An Exploration of
    Simultaneity Issues with GMM Methods, in K. Wagner and B. van Ark (eds), In-
    ternational productivity differences: Measurement and explanations, Elsevier Science
    Publishers B.V., Amsterdam, pp. 285–315.

Mansfield, E. (1968). Industrial Research and Technological Innovation: An Econometric
   analysis, Norton, New York.

McKinsey Global Institute (2001). US Productivity Growth 1995–2000 — Understandig
   the contribution of information technology relative to other factors, report.

M¨ller, A. A. (1998). Kapitalstock und Produktionspotential im privaten und ¨ffentlichen
 u                                                                          o
    Sektor Deutschlands, Nomos, Baden–Baden.

Nelson, R. R. (1982). The Role of Knowledge in R&D Efficiency, The Quarterly Journal
    of Economics 97/3: 453–470.

Nelson, R. R. and Winter, S. (1982). An Evolutionary Theory of Economic Change,
    Harvard University Press, Cambridge MA.

OECD (2000a). A New Economy? The Changing Role of Innovation and Information
   Technology in Growth, OECD, Paris.

OECD (2000b). The Service Economy, Business and industry policy forum series,
   STI/OECD.

                                          31
Productivity effects of ict in the german service sector
Productivity effects of ict in the german service sector
Productivity effects of ict in the german service sector
Productivity effects of ict in the german service sector

More Related Content

What's hot

Discussant slides final
Discussant slides finalDiscussant slides final
Discussant slides final
Wafi Ahmad
 
bigshift_mediaentertainment_11_09
bigshift_mediaentertainment_11_09bigshift_mediaentertainment_11_09
bigshift_mediaentertainment_11_09
Gyula Kangiszer
 
Debate Brasil e Alemanha - Formas Modernas de Contratação do Trabalho, 21/11/...
Debate Brasil e Alemanha - Formas Modernas de Contratação do Trabalho, 21/11/...Debate Brasil e Alemanha - Formas Modernas de Contratação do Trabalho, 21/11/...
Debate Brasil e Alemanha - Formas Modernas de Contratação do Trabalho, 21/11/...
FecomercioSP
 
BPIT-spez-II-2009-europe[1] - Kopieren
BPIT-spez-II-2009-europe[1] - KopierenBPIT-spez-II-2009-europe[1] - Kopieren
BPIT-spez-II-2009-europe[1] - Kopieren
Ruediger J. Gunther
 
intelligent-workplace-white-paper
intelligent-workplace-white-paperintelligent-workplace-white-paper
intelligent-workplace-white-paper
Jamie Hall
 

What's hot (20)

A review on a thesis, titled The Impact Of Information Technology On Producti...
A review on a thesis, titled The Impact Of Information Technology On Producti...A review on a thesis, titled The Impact Of Information Technology On Producti...
A review on a thesis, titled The Impact Of Information Technology On Producti...
 
Labour market policy more than just finding round pegs for round holes
Labour market policy more than just finding round pegs for round holesLabour market policy more than just finding round pegs for round holes
Labour market policy more than just finding round pegs for round holes
 
F0391051054
F0391051054F0391051054
F0391051054
 
Dan Andrews
Dan AndrewsDan Andrews
Dan Andrews
 
Jaana Remes
Jaana RemesJaana Remes
Jaana Remes
 
The Global Productivity Slowdown, Technology Diffusion and Public Policy
The Global Productivity Slowdown, Technology Diffusion and Public PolicyThe Global Productivity Slowdown, Technology Diffusion and Public Policy
The Global Productivity Slowdown, Technology Diffusion and Public Policy
 
IAOS 2018 - Towards a framework for measuring the digital economy, N. Ahmad
IAOS 2018 - Towards a framework for measuring the digital economy, N. AhmadIAOS 2018 - Towards a framework for measuring the digital economy, N. Ahmad
IAOS 2018 - Towards a framework for measuring the digital economy, N. Ahmad
 
The drivers of productivity: The role of organisational change and other firm...
The drivers of productivity: The role of organisational change and other firm...The drivers of productivity: The role of organisational change and other firm...
The drivers of productivity: The role of organisational change and other firm...
 
Discussant slides final
Discussant slides finalDiscussant slides final
Discussant slides final
 
Germany- ICT Opportunities & Business Analysis
Germany- ICT Opportunities & Business AnalysisGermany- ICT Opportunities & Business Analysis
Germany- ICT Opportunities & Business Analysis
 
Top Cited Articles in 2018 - International Journal of Managing Information T...
 Top Cited Articles in 2018 - International Journal of Managing Information T... Top Cited Articles in 2018 - International Journal of Managing Information T...
Top Cited Articles in 2018 - International Journal of Managing Information T...
 
Session 1 A Ryan Decker
Session 1 A   Ryan DeckerSession 1 A   Ryan Decker
Session 1 A Ryan Decker
 
bigshift_mediaentertainment_11_09
bigshift_mediaentertainment_11_09bigshift_mediaentertainment_11_09
bigshift_mediaentertainment_11_09
 
The Human Side of Productivity Setting the Scene
The Human Side of Productivity Setting the SceneThe Human Side of Productivity Setting the Scene
The Human Side of Productivity Setting the Scene
 
EMPLOYEE AUTONOMY AND THE WITHIN-FIRM GENDER WAGE GAP: THE CASE OF TRUST-BASE...
EMPLOYEE AUTONOMY AND THE WITHIN-FIRM GENDER WAGE GAP: THE CASE OF TRUST-BASE...EMPLOYEE AUTONOMY AND THE WITHIN-FIRM GENDER WAGE GAP: THE CASE OF TRUST-BASE...
EMPLOYEE AUTONOMY AND THE WITHIN-FIRM GENDER WAGE GAP: THE CASE OF TRUST-BASE...
 
Debate Brasil e Alemanha - Formas Modernas de Contratação do Trabalho, 21/11/...
Debate Brasil e Alemanha - Formas Modernas de Contratação do Trabalho, 21/11/...Debate Brasil e Alemanha - Formas Modernas de Contratação do Trabalho, 21/11/...
Debate Brasil e Alemanha - Formas Modernas de Contratação do Trabalho, 21/11/...
 
BPIT-spez-II-2009-europe[1] - Kopieren
BPIT-spez-II-2009-europe[1] - KopierenBPIT-spez-II-2009-europe[1] - Kopieren
BPIT-spez-II-2009-europe[1] - Kopieren
 
Keynote Bartelsman
Keynote BartelsmanKeynote Bartelsman
Keynote Bartelsman
 
intelligent-workplace-white-paper
intelligent-workplace-white-paperintelligent-workplace-white-paper
intelligent-workplace-white-paper
 
Drivers of productivity: Information technology, firm size, and organizationa...
Drivers of productivity: Information technology, firm size, and organizationa...Drivers of productivity: Information technology, firm size, and organizationa...
Drivers of productivity: Information technology, firm size, and organizationa...
 

Viewers also liked

The effect of ict and e business on eu trade
The effect of ict and e business on eu tradeThe effect of ict and e business on eu trade
The effect of ict and e business on eu trade
Felipe Schmidt
 
Startzondag - ZIN
Startzondag - ZINStartzondag - ZIN
Startzondag - ZIN
silkink
 
La mediazione civile e commerciale evento 20 gennaio 2011
La mediazione civile e commerciale evento 20 gennaio 2011La mediazione civile e commerciale evento 20 gennaio 2011
La mediazione civile e commerciale evento 20 gennaio 2011
giuliashub
 
Ict opportunity profiler
Ict opportunity profilerIct opportunity profiler
Ict opportunity profiler
Felipe Schmidt
 
Presentazione Giulia's Hub sul coworking
Presentazione  Giulia's Hub sul coworking Presentazione  Giulia's Hub sul coworking
Presentazione Giulia's Hub sul coworking
giuliashub
 
Ook zo'n ZIN in de startzondag
Ook zo'n ZIN in de startzondagOok zo'n ZIN in de startzondag
Ook zo'n ZIN in de startzondag
silkink
 
JOHNSON LILLY PRESS RELEASE (Approved)
JOHNSON LILLY PRESS RELEASE (Approved)JOHNSON LILLY PRESS RELEASE (Approved)
JOHNSON LILLY PRESS RELEASE (Approved)
garydstratton
 
Th sounds-pronunciation-091006165657-phpapp01
Th sounds-pronunciation-091006165657-phpapp01Th sounds-pronunciation-091006165657-phpapp01
Th sounds-pronunciation-091006165657-phpapp01
letlady
 

Viewers also liked (15)

The effect of ict and e business on eu trade
The effect of ict and e business on eu tradeThe effect of ict and e business on eu trade
The effect of ict and e business on eu trade
 
Purdue's Neil Armstrong Building
Purdue's Neil Armstrong BuildingPurdue's Neil Armstrong Building
Purdue's Neil Armstrong Building
 
Startzondag - ZIN
Startzondag - ZINStartzondag - ZIN
Startzondag - ZIN
 
La mediazione civile e commerciale evento 20 gennaio 2011
La mediazione civile e commerciale evento 20 gennaio 2011La mediazione civile e commerciale evento 20 gennaio 2011
La mediazione civile e commerciale evento 20 gennaio 2011
 
Ict opportunity profiler
Ict opportunity profilerIct opportunity profiler
Ict opportunity profiler
 
Presentazione Giulia's Hub sul coworking
Presentazione  Giulia's Hub sul coworking Presentazione  Giulia's Hub sul coworking
Presentazione Giulia's Hub sul coworking
 
Presentazione Giulia's Hub 30 giugno
Presentazione Giulia's Hub 30 giugnoPresentazione Giulia's Hub 30 giugno
Presentazione Giulia's Hub 30 giugno
 
Salient
SalientSalient
Salient
 
Ook zo'n ZIN in de startzondag
Ook zo'n ZIN in de startzondagOok zo'n ZIN in de startzondag
Ook zo'n ZIN in de startzondag
 
Ppt 20 dicembre
Ppt 20 dicembrePpt 20 dicembre
Ppt 20 dicembre
 
Portugal ict sector
Portugal ict sectorPortugal ict sector
Portugal ict sector
 
Cintasa's introduction
Cintasa's introductionCintasa's introduction
Cintasa's introduction
 
JOHNSON LILLY PRESS RELEASE (Approved)
JOHNSON LILLY PRESS RELEASE (Approved)JOHNSON LILLY PRESS RELEASE (Approved)
JOHNSON LILLY PRESS RELEASE (Approved)
 
Th sounds-pronunciation-091006165657-phpapp01
Th sounds-pronunciation-091006165657-phpapp01Th sounds-pronunciation-091006165657-phpapp01
Th sounds-pronunciation-091006165657-phpapp01
 
Introduction and Brief History of Computers
Introduction and Brief History of ComputersIntroduction and Brief History of Computers
Introduction and Brief History of Computers
 

Similar to Productivity effects of ict in the german service sector

Impact of ICT and innovation on industrial productivity in Uruguay
Impact of ICT and innovation on industrial productivity in UruguayImpact of ICT and innovation on industrial productivity in Uruguay
Impact of ICT and innovation on industrial productivity in Uruguay
Dialogo regional sobre Sociedad de la Información
 
2009 - Interactive television A model of analysis of business.pdf
2009 - Interactive television A model of analysis of business.pdf2009 - Interactive television A model of analysis of business.pdf
2009 - Interactive television A model of analysis of business.pdf
ssuser4dafea
 
It Debate Final
It Debate FinalIt Debate Final
It Debate Final
UNU.MERIT
 
Information Technology Investment in Sustainability and Profitability
Information Technology Investment in Sustainability and ProfitabilityInformation Technology Investment in Sustainability and Profitability
Information Technology Investment in Sustainability and Profitability
Yasser Al Mimar
 
On the contribution of information and communicationtechnolo.docx
On the contribution of information and communicationtechnolo.docxOn the contribution of information and communicationtechnolo.docx
On the contribution of information and communicationtechnolo.docx
cherishwinsland
 
Llanza-A-2015-MScManagement
Llanza-A-2015-MScManagementLlanza-A-2015-MScManagement
Llanza-A-2015-MScManagement
Álvaro Llanza
 
Information Technology and Productivity
Information Technology and ProductivityInformation Technology and Productivity
Information Technology and Productivity
Chindavanh Vilivong
 
Dutch Construction Industry
Dutch Construction IndustryDutch Construction Industry
Dutch Construction Industry
Michael Hompes
 
Bus_transformation Session 1 ( LECTURE 1).pdf
Bus_transformation Session 1 ( LECTURE 1).pdfBus_transformation Session 1 ( LECTURE 1).pdf
Bus_transformation Session 1 ( LECTURE 1).pdf
NellyELHassani
 
How can 'IT' improve national competitiveness
How can 'IT' improve national competitivenessHow can 'IT' improve national competitiveness
How can 'IT' improve national competitiveness
Mike Backhouse
 

Similar to Productivity effects of ict in the german service sector (20)

ICT and labor productivity in Colombian manufacturing industry
ICT and labor productivity in Colombian manufacturing industryICT and labor productivity in Colombian manufacturing industry
ICT and labor productivity in Colombian manufacturing industry
 
it & Economic Performance a Critical Review of the Empirical Data
it & Economic Performance a Critical Review of the Empirical Datait & Economic Performance a Critical Review of the Empirical Data
it & Economic Performance a Critical Review of the Empirical Data
 
Impact of ICT and innovation on industrial productivity in Uruguay
Impact of ICT and innovation on industrial productivity in UruguayImpact of ICT and innovation on industrial productivity in Uruguay
Impact of ICT and innovation on industrial productivity in Uruguay
 
2009 - Interactive television A model of analysis of business.pdf
2009 - Interactive television A model of analysis of business.pdf2009 - Interactive television A model of analysis of business.pdf
2009 - Interactive television A model of analysis of business.pdf
 
Inapp cirillo fanti_mina_ricci_digital_technologies_and_firm_performance_indu...
Inapp cirillo fanti_mina_ricci_digital_technologies_and_firm_performance_indu...Inapp cirillo fanti_mina_ricci_digital_technologies_and_firm_performance_indu...
Inapp cirillo fanti_mina_ricci_digital_technologies_and_firm_performance_indu...
 
It Debate Final
It Debate FinalIt Debate Final
It Debate Final
 
Information Technology Investment in Sustainability and Profitability
Information Technology Investment in Sustainability and ProfitabilityInformation Technology Investment in Sustainability and Profitability
Information Technology Investment in Sustainability and Profitability
 
“Keeping pace with technological change: The role of capabilities and dynamis...
“Keeping pace with technological change: The role of capabilities and dynamis...“Keeping pace with technological change: The role of capabilities and dynamis...
“Keeping pace with technological change: The role of capabilities and dynamis...
 
On the contribution of information and communicationtechnolo.docx
On the contribution of information and communicationtechnolo.docxOn the contribution of information and communicationtechnolo.docx
On the contribution of information and communicationtechnolo.docx
 
Economic impact of e20 F. Biagi Tech4i2
Economic impact of e20 F. Biagi Tech4i2Economic impact of e20 F. Biagi Tech4i2
Economic impact of e20 F. Biagi Tech4i2
 
Llanza-A-2015-MScManagement
Llanza-A-2015-MScManagementLlanza-A-2015-MScManagement
Llanza-A-2015-MScManagement
 
Evaluation of Integrated Care: From methods to governance and applications - ...
Evaluation of Integrated Care: From methods to governance and applications - ...Evaluation of Integrated Care: From methods to governance and applications - ...
Evaluation of Integrated Care: From methods to governance and applications - ...
 
Information Technology and Productivity
Information Technology and ProductivityInformation Technology and Productivity
Information Technology and Productivity
 
VALUE-BASED-ASSESSMENT.pdf
VALUE-BASED-ASSESSMENT.pdfVALUE-BASED-ASSESSMENT.pdf
VALUE-BASED-ASSESSMENT.pdf
 
CO-EVOLUTIONARY COUPLING BETWEEN CAPTURED AND UNCAPTURED GDP CYCLES:CROSS LEA...
CO-EVOLUTIONARY COUPLING BETWEEN CAPTURED AND UNCAPTURED GDP CYCLES:CROSS LEA...CO-EVOLUTIONARY COUPLING BETWEEN CAPTURED AND UNCAPTURED GDP CYCLES:CROSS LEA...
CO-EVOLUTIONARY COUPLING BETWEEN CAPTURED AND UNCAPTURED GDP CYCLES:CROSS LEA...
 
Dutch Construction Industry
Dutch Construction IndustryDutch Construction Industry
Dutch Construction Industry
 
Bus_transformation Session 1 ( LECTURE 1).pdf
Bus_transformation Session 1 ( LECTURE 1).pdfBus_transformation Session 1 ( LECTURE 1).pdf
Bus_transformation Session 1 ( LECTURE 1).pdf
 
Wp inapp cirillo_fanti_mina_ricci_digitizing_firms_wp_53_2020
Wp inapp cirillo_fanti_mina_ricci_digitizing_firms_wp_53_2020Wp inapp cirillo_fanti_mina_ricci_digitizing_firms_wp_53_2020
Wp inapp cirillo_fanti_mina_ricci_digitizing_firms_wp_53_2020
 
TMT Valuations - Report
TMT Valuations - ReportTMT Valuations - Report
TMT Valuations - Report
 
How can 'IT' improve national competitiveness
How can 'IT' improve national competitivenessHow can 'IT' improve national competitiveness
How can 'IT' improve national competitiveness
 

More from Felipe Schmidt

More from Felipe Schmidt (9)

Influencias de las TIC's
Influencias de las TIC'sInfluencias de las TIC's
Influencias de las TIC's
 
Sql
SqlSql
Sql
 
Paper
PaperPaper
Paper
 
Modelo tres capas
Modelo tres capasModelo tres capas
Modelo tres capas
 
Laboratorios
LaboratoriosLaboratorios
Laboratorios
 
Ensayo de sistemas de información en ciencias empresariales
Ensayo de sistemas de información en ciencias empresarialesEnsayo de sistemas de información en ciencias empresariales
Ensayo de sistemas de información en ciencias empresariales
 
Data ware house
Data ware houseData ware house
Data ware house
 
Conceptos
ConceptosConceptos
Conceptos
 
¿Qué es el Modelo Tres Capas?
¿Qué es el Modelo Tres Capas?¿Qué es el Modelo Tres Capas?
¿Qué es el Modelo Tres Capas?
 

Productivity effects of ict in the german service sector

  • 1. Does Experience Matter? Productivity Effects of ICT in the German Service Sector Thomas Hempell Centre for European Economic Research (ZEW), Mannheim February 2002 Abstract In this paper, it is argued that ICT investment is closely linked with complementary innovations and most productive in firms with innovative experience. In an analysis based on firm–level panel data covering the period 1994–99, system GMM estimates for an extended production function framework reveal significant productivity effects of ICT in the German service sector. Moreover, there is strong support for the hypothesis that the experience gained from past process innovations is a specific complement that makes ICT investment more productive. The results suggest that ICT may have been contributing to productivity differentials both between firms and countries. Keywords: Information and Communication Technologies, Services, Production Func- tion Estimation, Panel Data JEL–Classification: C23, D24, O32 Address: Centre for European Economic Research (ZEW) Research Group of Information and Communication Technologies P.O. Box 10 34 43 D–68034 Mannheim Germany Phone: +49/621/1235–233 Fax: +49/621/1235–225 E-Mail: hempell@zew.de
  • 2. 1 Introduction In spite of the current economic woes and emerging disillusions about a ‘New Economy’, productivity effects of information and communication technologies (ICT) continue to play a key role in assessing the prospects and growth potentials of both firms and whole economies. In fact, the economic downturn currently experienced by some countries shows that ICT are far from being a panacea that yields permanent growth and the end of business cycles, as some analysts suggested at the peak of the hype. Rather, there is growing support for the view that it is the specific economic circumstances that determine the success of exploiting the productivity potentials induced by new technologies. ICT have been compared to other important historical inventions like the steam engine or the electric motor (David and Wright, 1999). These inventions are interpreted as ‘general purpose technologies’ (GPT) with three main characteristics: they are pervasive, they carry an inherent potential for technical improvements and they lead to innovational complementarities and scale economies (Bresnahan and Trajtenberg, 1995). According to this view, GPT should be understood primarily as ‘enabling technologies’, “opening up new opportunities rather than offering complete, final solutions” (Bresnahan and Trajtenberg, 1995:84). Various empirical studies seem to support the ‘general purpose’ hypothesis. Microe- conomic investigations by Brynjolfsson and Hitt (1995; 1996) and Lichtenberg (1995) found significantly positive contributions of ICT investment to output and productivity at the firm level. A macroeconomic study by Schreyer (2000) concludes that ICT investment has fostered output and productivity growth in the OECD countries during the 90s. Similarly, Colecchia and Schreyer (2001) find evidence that a broad application of ICT within an economy plays a key role in this context. They conclude that a rapid diffusion of ICT depends less on the existence of an ICT producing sector but rather on the flexibility of product and labour markets as well as the business environment. The ICT boom at the end of the 90s was accompanied by many exaggerations. The stock markets soared, the market capitalization of various dot.com firms spectacularly exceeded the value of several giants of the ‘Old Economy’. Bank credits and venture capital were given to new firms on the basis of fashionable ideas rather than sound business plans. In the meanwhile, about two years after the end of the ‘hype’ in 2000, both managers and economists have become less enthusiastic about the emergence of a ‘New Economy’. On the one hand, numerous of the once highly praised dot.com firms have declared bankruptcy. Moreover, many firms of the ‘Old Economy’ have not survived the wave of innovation and increased globalization and competition induced by the new technologies. On the other hand, some brick–and–mortar firms have been able to use 1
  • 3. ICT to boost productivity and market power. Obviously, not all firms have been equally able to benefit from the productivity potentials of ICT. More recent studies have tried to explain these differences in the productivity effects between firms. Brynjolfsson and Hitt (2000) emphasize the role of organizational changes that are necessary to reap the potential benefits of ICT. Brynjolfsson and Yang (1999) and Yang and Brynjolfsson (2001) point to the role of intangible assets that are complementary to the firm’s use of ICT. Bresnahan, Brynjolfsson and Hitt (2001) find evidence for significant synergies between ICT, workplace reorganization and new services and products. These studies have an important message in common: in order to assess the impact of ICT on firm performance it is crucial to investigate the firm–specific circumstances in which ICT are used. Hence, complementary factors are essential for the ICT to unfold its productivity impacts as a ‘general purpose technology’. The purpose of this paper is to shed more light on the factors that determine the success of improving a firm’s performance by ICT use. Unlike previous studies of the topic, the theoretical model proposed in this study stresses the importance of ICT being part of the innovation process within a firm. With reference to evolutionary approaches in innovation theory it is argued that firms that have introduced innovations in the past are better prepared for ICT–induced innovations, like process improvements and organizational restructuring, than firms without any innovation experience. As a consequence, the model predicts productivity effects of ICT to be higher in experienced firms. The empirical analysis of this hypothesis is based on a sample of firms from a representative survey in the German business–related and distribution service sector covering the period 1994 to 1999. In this regard, the study distinguishes itself from most previous studies that have concentrated on the manufacturing sector or on samples of large firms only.1 However, there are several reasons to draw more attention to the service sector. Firstly, in the service sector, ICT investment is most dynamic and most extensive in volume (OECD, 2000a; EITO, 2001). Moreover, as pointed out in OECD (2000b), the service sector — and in particular business–related services — have been the most important driver of economic growth over the last decades in most industrialized countries.2 Finally, recent macroeconomic studies — for example by Baily and Lawrence (2001) and McKinsey Global Institute (2001) — conclude that services have accounted 1 See for example Pilat (2001), Brynjolfsson and Yang (1996) and Brynjolfsson and Hitt (2000) for surveys of the empirical literature. 2 In Germany, the fraction of value added contributed by financial services and business services rose from 24,1% in 1991 to 30,3% in 1999 (Sachverst¨ndigenrat zur Begutachtung der gesamtwirtschaftlichen a Entwicklung, 2001). 2
  • 4. for a substantial part of productivity acceleration in the U.S. during the 90s. Therefore, in order to better assess the impact of ICT on the economy as a whole, a more detailed understanding of the underlying processes in the service sector seems to be particularly important. A further advantageous feature of the sample employed in this study is that it covers firms of all size classes instead of only large firms. This broader focus is likely to increase the variation of characteristics and implementation strategies between firms and thereby helps to identify the role of the firm characteristics. Moreover, the broader focus en- hances the likelihood of receiving a more representative picture for the economy as a whole. Even though the German economy continues to play a key role within the European Union as well as within the OECD, few attempts have been made to quantitatively assess the productivity effects of ICT use in German firms. To the knowledge of the author, the only related study is the paper by Bertschek and Kaiser (2001) who find significant contributions of ICT investment in a cross–section analysis of business–related services firms.3 Hence, a further objective of this paper is to shed more light on the impact of ICT use in the German economy. The paper is organized as follows. In section 2, the theoretical background is discussed and the hypothesis of innovative experience playing a key role in efficient ICT use is implemented into a production function framework. Section 3 gives an overview of the data employed, that is the Mannheimer Innovation Panel in the Service Sector (MIP-S), and describes the methods used to construct different stock values for ICT and conventional capital separately. Section 4 discusses the econometric issues and presents the empirical results for both a simple ICT–extended production function framework and the more specific model about the role of innovative experience. Section 5 concludes with some comments on the implications of the findings concerning theoretical, methodological and policy issues. 2 Methodological and Theoretical Framework Like many other important inventions in history, the rapid technological progress of the semi-conductor industry and the fast expansion of the internet evoked high expectations about their impact on productivity and growth. The adoption and diffusion of these new technologies were expected to boost productivity substantially. However, for a long time anecdotal evidence of computerized workplaces did not show up in aggregate 3 Licht and Moch (1999) use qualitative data to assess the impact of ICT on the performance of German firms. They find that the main impact of ICT use are quality improvements. 3
  • 5. productivity statistics, which led to the proclamation of the “productivity paradox” of ICT. While investment in ICT started to grow rapidly in the U.S. since the 1970s, labour productivity growth slowed down substantially and remained at a low level until the beginning of the 90s (Jorgenson and Stiroh, 1995). The service sector was at the heart of this paradox. Around 1990, the U.S. service sector accounted for nearly 80 percent of total IT investment in the U.S. but experienced productivity growth only slightly above zero (Brynjolfsson and Yang, 1996). In this regard, theoretical and empirical approaches that focus on the firm level rather than aggregate level appear very promising. Theoretically, the productivity impact of ICT is likely to vary substantially between firms. Like in the case of many other innovation waves in economic history, some firms are better prepared than others to take productive advantage of new technologies. For the particular case of ICT, it has been argued that complementary factors like organizational restructuring and intangible assets play a key role for ICT to unfold its largest benefits (Brynjolfsson and Hitt, 2000; Yang and Brynjolfsson, 2001). By treating all firms equally (assuming equal input coefficients in a production function framework, for example), the true impact of ICT on “well–prepared” firms may be understated. Therefore, in order to assess the determinants of successfull ICT use, it may be useful to explicitly model the influence of various firm characteristics on the efficient use of new technologies. A further — rather empirical — source of the productivity paradox is that official statistics do not adequately reflect quality changes. As pointed out by Brynjolfsson (1994) and Licht and Moch (1999), quality improvements — in particular improved customer service — are an especially important goal for ICT investment decisions. Griliches (1994) shows that the problem of unmeasured quality improvements is especially important in the case of ‘unmeasurables’ services like trade and F.I.R.E. (finance, insurance, real estate) where ICT investment has grown most rapidly. As a consequence, the contribution of ICT to real output growth inferred from aggregate data are likely to be biased towards zero. For applications to firm–level data, this problem will be less severe. If a firm invests in ICT in order to improve the quality of a product while its competitors continue to offer their old products, the innovating firm will try to charge a higher price for its new product. If the quality improvement is approved by customers, they will in fact be willing to pay a higher price for the good and the value added of the innovating firm will increase accordingly.4 Consequently, Brynjolfsson and Hitt (2000) argue that microeconomic studies will capture this effect and variations in output quality will contribute to measuring a higher output elasticity of ICT investment. This relationship 4 A firm level study by Brynjolfsson and Hitt (1995) did not find any significant differences in IT produc- tivity between “measurable” and “unmeasurable” sectors, indicating that appropriate quality measurement is mainly a problem at the aggregate level. 4
  • 6. is independent of the price deflator employed for the corresponding industry. Moreover, firm–level data sets are particularly suitable for empirical assessment of productivity effects if — apart from input and output data — they contain information on additional firm characteristics. These data may allow to identify the determinants of success in ICT adoption. On the contrary, this identification will be much more difficult with aggregate data since differences between firms will even out by aggregation. For these reasons — both theoretical and empirical —, firm level analysis seems to be the appropriate approach for a more detailed investigation of productivity effects of ICT investments. In the following first subsection, it is argued on the basis of evolutionary approaches of innovation that a distinction between firms according to their innovative experience may play an important role for the benefits of ICT use in services. In the second subsection, this distinction is formalized within an extended microeconomic production function framework that is implemented empirically in the following. 2.1 Innovation in Services: The case of evolutionary approaches In earlier studies on innovation, the service sector has been characterized as a mere applier of technological innovations developed in the manufacturing sector (see for example the influential taxonomy by Pavitt, 1984). Recent studies, however, tend to confirm a more active role of the service sector in the process of technological change (see Blind et al., 2000; Sirilli and Evangelista, 1998). Though new technologies of the manufacturing sector play an important role in initiating innovation processes in the service sector, the restructuring of processes and improvement of existing services often leads to the development of completely new services (Barras, 1986). Therefore, like in the manufacturing sector, differences in technological opportunities, appropriability conditions, and cumulativeness of innovative capabilities may lead to differences in the innovation paths between firms. As pointed out by Cohen and Levinthal (1990, p. 128), “the ability of a firm to recognize the value of new, external information, assimilate it, and apply it to commercial ends” is critical to its innovative capabilities. They argue that this ‘absorptive capacity’ is largely a function of the level of prior related knowledge.5 Firstly, the absorptive capacity 5 Building on studies in cognitive and behavioral sciences, Cohen and Levinthal (1990) state that at the level of individuals, “learning is cumulative, and learning performance is greatest when the object of learning is related to what is already known” (131). At the level of organizations, for example firms, the absorptive capacity is not just the sum of absorptive capacities of its individual employees but also depends on the transfer of knowledge across and within subunits. As a consequence, there is a trade off between the aim of highly diversified capacities of the individuals in order to recognize the value of various new external information on the one hand, and the need of specialization or a “shared language” to ensure the transfer of knowledge within the organization. Therefore, the effect of specialization on the learning effect 5
  • 7. accumulated in a particular area in one period will permit a more efficient accumulation in the next. Secondly, ‘experienced’ firms will be able to better predict the nature and commercial potential of technological advances in an uncertain environment. Both aspects of absorptive capacity — cumulativeness and the impact on expectation formation — “imply that its development is domain–specific and is path– or history–dependent” (136). There are further studies that stress evolutionary aspects of innovation in a similar way. Mansfield (1968) and Stoneman (1983) argue that a firm’s innovative success enhances its technological opportunities and thereby makes further success more likely. This ‘success breeds success’ hypothesis is supported by an empirical study by Flaig and Stadler (1994). They derive a stochastic dynamic optimization model with intertemporal spill–overs of innovations within firms. In their empirical application, they find that firms that have introduced innovations in the past are indeed more likely to innovate in subsequent years. Similar to these arguments and findings, innovative experience may be an explanation for substantial differences in firms’ ability to boost productivity by the application of ICT. As emphasized by Bresnahan et al. (2001), the use of ICT is closely linked to innovations within the firm. Case studies show that organizational changes and process reengineering play a key role as complementary innovational efforts (Brynjolfsson and Hitt, 2000). Both the ‘absorptive capacity’ and the ‘success breeds success’ hypotheses suggest that firms that have introduced such innovations in the past are probably better prepared than non-innovative firms to reap the potential benefits of ICT. There are several more specific reasons supporting this view. Firstly, managers are likely to have learned from past mistakes.6 They are better prepared to assess the potentials and limits of introducing major changes within their company, being aware of possible reactions of their employees and their traditional customers. In short, they have varied experience concerning the optimal speed of innovation of their firm. Secondly, innovative firms will have more experience in how to implement innovations. In most cases, the efficient introduction of ICT requires complementary changes of organizational structures and processes as pointed out by Brynjolfsson and Hitt (2000). Innovative firms might be more successful in training and motivating their employees to take part actively in the innovation process. This argument suggests, in fact, that it will mainly be experience in process and organizational innovations that increase the ability of at the level of firms is ambigous. 6 The returns to innovations are far from being safe. However, the likelihood of a success may increase with the experience gathered from past innovations. In this regard, the learning process in introducing innovations may be compared to the search model proposed by Nelson (1982) for the case of R&D. In- novation is modelled as a search process in which knowledge helps to lower the search costs by focussing search on more relevant alternatives. 6
  • 8. exploiting the productivity potentials of ICT successfully. Thirdly, firms that are more innovative may have a higher share of employees that accept an intensive use of ICT at their workplace. This may be due to both special recruitment strategies and to specific experience and greater flexiblity of the workforce from former innovation activities.7 Finally, innovative firms may have gained some degree of ‘innovative reputation’ or branding in new business areas. In particular, it might be easier for them to attract new customers by exploiting ICT–based sales channels (B2B, B2C) or by introducing new ICT-based services or products.8 Since innovative reputation can most likely be acquired by offering innovative products, this argument would imply that past product innovations raise the ability of benefiting from ICT. Taken together, these arguments imply that the returns to ICT investment are higher in innovative firms than in firms with little or no innovative experience. In the empirical testing of this hypothesis it will be important to distinguish between different kinds of innovative experience, in particular between product and process innovations. In the following formal derivation of the model, however, this distinction is abstracted from and left for the empirical implementation in section 4. 2.2 The Model In this part, a theoretical framework of production is proposed focussing on two specific questions raised in the previous part. Firstly, is innovative experience a crucial prereq- uisite to implement ICT efficiently? And if so, secondly, is ICT capital ‘special’ in the sense that other capital inputs are much less dependent on innovative experience? In order to assess these questions, the traditional Cobb-Douglas production setup is used as a starting point: Yi = F (Ai , Ki , Li ) = Ai Lγ1 Kiγ2 i (1) where Yi is value added of firm i, Ki is conventional (non-ICT) capital, Li is effective labour and Ai is the multifactor productivity of firm i.9 ICT capital does not enter 7 Nelson and Winter (1982) suggest that much of the knowledge of a firm’s organizational routines and objectives is tacit. They summarize this observation by claiming that organizations ‘remember by doing’. As a consequence, innovational capabilities are difficult to raise by hiring new personnel. Also in this regard, the firm’s innovative history plays a key role (see Cohen and Levinthal, 1990). 8 Smith and Brynjolfsson (2001) find that brand is an important determinant of consumer choice in internet transactions. They find that “consumers use brand as a proxy for retailer credibility in non– contractible aspects of the product and service bundle, such as shipping reliability” (541). Interestingly, in the study the strongest brand effect was found for amazon, which has gained reputation as one of the leading internet retailers. 9 F (•) may be such that Y exhibits constant returns to scale in K and L (γ1 +γ2 = 1), but not necessarily. 7
  • 9. the production function directly, but rather affects multifactor productivity jointly with innovative experience as a complement: Ai = A(ICTi , Ei , mi ) (2) where Ai represents firm’s i (time–invariant) multifactor productivity, ICTi rep- resents its amount of ICT capital, Ei innovative experience and mi collects other unobserved efficiency parameters of firm i like management skills, location, branding, ∂Ai etc. Furthermore, the following properties of function A(•) are imposed: (1) ∂ICTi ≥ 0, ∂Ai ∂ 2 Ai (2) ∂Ei ≥ 0, (3) ∂ICTi2 ≤ 0. That is, a firm’s productivity depends positively on both ICT input (1) and innovative experience (2). Marginal contributions to productivity are assumed to decrease with the amount of ICT employed (3). That is, the productivity potentials of ICT are limited. Finally — and most importantly — it is predicted that the potential benefits from ICT are higher for firms with innovative experience. More specifically, the productivity contributions of ICT are increasing in the degree of ∂2 i innovative experience: (4) ∂ICTA∂Ei ≥ 0. A very convenient functional form satisfying i these properties is the following specification: x(Ei ) Ai = C · mi · ICTi (3) with C as a common scale factor and x(•) being a strictly monotone function of Ei such that ∂x(•) > 0 and x(•) ∈ [0, 1]. Past innovation activities are considered as a proxy ∂Ei of a firm’s innovative experience such that Ei = 1 if firm i has been an innovator and Ei = 0 otherwise. Denoting x = x(1) and x(0) = x, it follows that x = x + ∆x with ∆x > 0. Then equation (3) becomes: Ai = C · mi · ICT x+∆x·Ei (4) Inserting equation (4) into (1) and taking logs then yields the following extended production function equation: ln Yi = ln C + ln mi + x ln ICTi + ∆x(Ei · ln ICTi ) + γ1 ln Li + γ2 ln Ki (5) Thus, the model corresponds to an ICT–extended Cobb–Douglas function except that the coefficient of ICT is predicted to be different for experienced firms (x = x + ∆x) and unexperienced firms (x). One implication of this framework is particularly noteworthy. The theoretical property that innovative firms exhibit a higher elasticity of output with respect to ICT (∆x > 0) does not necessarily imply the marginal product of ICT to be higher in innovating firms since the marginal product depends also on the ICT 8
  • 10. intensity (ICTi /Yi ) of the individual firm.10 What a higher elasticity does imply is that in innovating firms the marginal output contributions of ICT decrease less rapidly when ICT–input is raised. In the context here, this is equivalent to saying that there are more potential benefits to be exploited by the use of ICT within innovative firms. In the extreme case, innovative experience is an essential condition for ICT to yield benefits at all. If so, the output elasticity of ICT would be zero for non–experienced firms (x = 0) or even negative11 , whereas experienced firms exhibit positive returns to ICT capital (∆x > max(0, −x)). Under the assumption that marginal products of ICT (MPI) are equal across all firms (that is ICT is earning equal returns), the implication is that x(E) experienced firms will produce their services with a higher ICT intensity ICT = M P I and Y a lower intensity in other factors. Finally, equation (5) is transformed into the following econometrical model: ln Yit = c + ηi + β1 ln Lit + β2 ln ICTit + β3 ln Kit +β4 (ln ICTit · Ei ) + [β5 (ln Kit · Ei )] + it (6) with c = ln C, ηi = ln mi , β1 = γ1 , β2 = x, β3 = γ2 , β4 = ∆x and it as a normally distributed disturbance term. In this framework, the answer to the question whether innovative experience influences the output contributions of ICT depends crucially on whether the coefficient β4 turns out to be significantly positive or not. As far as the second initial hypothesis is concerned, the effect of experience on ICT is contrasted by the comparable effect on non–ICT capital (see the β5 -term in brackets). If ICT is a ‘special’ capital input by its dependence on past innovations, β5 is expected to be zero and β4 to be positive in a simultaneous estimation of both coefficients. 3 The Data In order to implement the production framework empirically, data from the Mannheimer Innovation Panel in Services (MIP-S) are employed. This survey is conducted by the Cen- tre of European Economic Research (ZEW) on behalf of the German Federal Ministry for Education and Research (bmb+f). The data has been being collected annually since 1994 in a representative survey of innovation activities in the German business related-service and distribution sector and includes information of more than 2000 firms (Janz et al. 10 Formally, the marginal returns to ICT (MPI) are the product of the output elasticity of ICT and the ∂Yi Yi inverse ratio of ICT capital in output: M P Ii = ∂ICTi = x(Ei ) · ICTi . Therefore, the MPI increases with the output elasticity x(Ei ) but decreases with the share of ICT capital in output ICTi . Yi 11 A negative coefficient would imply that the introduction of ICT leads to efficiency losses, for example due to high costs of internal restructuring or long lags in ICT causing measurable productivity effects. 9
  • 11. 2001). It has an (unbalanced) panel structure in important key variables for the years since 1994. The survey methodology is closely related to the guidelines proposed in the Oslo- Manual on innovation statistics (OECD/Eurostat, 1997). Furthermore, the 1997 wave of the survey in the service sector formed part of the Community Innovation Survey (CIS). For the particular purpose of the empirical analysis, the MIP-S data set contains annual data on sales, number of employees (full-time equivalents), skill structure, expenditures for gross investment and for ICT-capital (hardware, software and telecom- munication technology). Furthermore, firms were asked whether they had introduced a technologically new or substantially improved products (product innovation) or processes (process innovation) within the last three years. In order to estimate the model equation, some data transformations were necessary. Most importantly, capital stocks for ICT capital and conventional (non-ICT) capital were constructed separately. For this purpose, investment on conventional capital was defined as total investment expenditures minus ICT expenditures.12 In order to derive the corresponding real investment, conventional investment is deflated by the deflator of the German Statistical office.13 As far as deflators for ICT goods are concerned, German official price statistics on ICT goods tend to understate the real price decline of this product class (Hoffmann, 1998). Therefore, the price deflators for total ICT in the U.S., which are calculated on a hedonic basis (see Jorgenson, 2001), are used. Additionally, these prices were corrected for exchange rate changes, multiplying the U.S. deflators by the annual averages of the $/DM exchange rates (taken from German Bundesbank, time series wj5636). Given the deflated investments for both types of capital, the perpetual inventory method with constant, linear depreciation,14 was applied to construct the capital stocks 12 Some firms reported investment expenditures in ICT that exceeded total investment (leading to a negative gross investment in conventional capital). These inconsistencies were most frequent for the years 1994, 1997 and 1998 (6.3%) but almost absent in 1995 and 1996 (0.2%). The most likely explanation seems to be questionnaire design: for the years 1995 and 1996, the question on total investment was immediately followed by the question on ICT investment, thereby drawing the respondent’s attention to the consistency issue. The most likely underlying source for the inconsistent answers seems to be that respondents had not included ICT investment in total investment expenditures. In particular, expenditures for software may have been left out. In order to avoid the loss of many observations, the inconsistent cases were interpreted as having equated total investment with conventional investment only. 13 The index of the producer prices, investment goods, from the Statistical Yearbooks with 1996 as the base year, was employed for this purpose. 14 It may be argued that especially for the case of ICT capital it might be more adequate to apply a vintage model in which computers maintain their productive efficiency over the lifetime of computers (see for example Jorgenson and Stiroh, 1995). However, the definition of ICT includes a very broad range of technologies. Furthermore, the length of the time series available is very short. The assumed life cycle of ICT capital (1/δ2 = 4 years) would exceed the time series available for the vast majority of the firms. 10
  • 12. for ICT and non-ICT. Accordingly, capital Kt in period t with investment It−1 is defined as: Kkt = (1 − δk )Kk,t−1 + Ik,t−1 (7) with k = 1 for conventional and k = 2 for ICT capital and investment. There are two potential problems in this approach. Firstly, reasonable values for the depreciation rates of both types of capital have to be found. Secondly, since no information is available on the level of capital stocks, initial capital stocks have to be constructed for all individual firms. Therefore, the method proposed by Hall and Mairesse (1995) for the construction of an R&D stock was followed since the problems in both contexts are very similar. Under the assumption that investment expenditures in capital good k have grown at a similar, constant average rate gk in the past for all firms, by backward substitution equation (7) can be rewritten for period t = 1 (1994) in the following way:15 Kk1 = Ik0 + (1 − δk )Ik,−1 + (1 − δk )2 Ik,−2 + . . . (8) ∞ ∞ s 1 − δk = Ik,−s (1 − δk )s = Ik0 s=0 s=0 1 + gk Ik1 = gk + δk Constant linear depreciation rates are assumed for conventional capital (δ1 = 0.06) and ICT capital (δ2 = 0.25) correspondingly. In particular, with δ1 < δ2 it is taken into account that the fast technological progress in ICT implies more frequent replacement of ICT inventory than of conventional capital (including buildings and office furniture among others). In order to derive the initial capital stocks, assumptions about pre-period growth rates of both type of investments must be made. For non-ICT investment expenditures, an annual growth rate of approximately 5% (g1 = 0.05) is assumed.16 For ICT investment, no time series are available for Germany. In order to get a rough idea of the evolution of ICT investments during the last decades, U.S. data are referred to as Therefore, there is little gain in trying to exploit the potential advantages of the vintage approach in this context. 15 In fact, the initial value of investment in conventional capital I1,1 was replaced by the average of the observed values of conventional investment for each firm. With this “smoothing” it was aimed to correct for cyclical effects which might have affected the estimated capital stock due to different initial years in the unbalanced panel. The underlying assumption is that long term growth of investment in conventional capital (g1 = 0.05) is relatively low compared to cyclical variations in this variable. On the contrary, the first observation on ICT capital was not replaced by the corresponding averages since long-term growth (g2 = 0.4) rates of ICT investment are more likely to dominate changes that are due to cyclical fluctuations. 16 Calculations on capital data provided by M¨ller (1998) show that gross capital stock in German services u has grown on average by 4.8% annually between 1980 and 1991. 11
  • 13. a rough guideline. Jorgenson and Stiroh (1995) calculate an average annual growth rate of 44.3% in real computer investment and of 20.2% for OCAM (office, computing, and accounting machinery) between 1958 and 1992 for the U.S. Since the share of computers in OCAM has been steadily rising and is more closely linked to ICT investment as defined in the context here, an annual pre-period growth rate of — somewhat arbitrarily — g2 = 0.4 for ICT investment is assumed.17 Since there are time lags between the installation and productive contribution of capital goods, the capital stock at each period’s beginning (or at the end of the corresponding fore-period) are taken as measures for ICT and conventional capital input. For labour input, the annual average of the number of the firm’s employees (full-time equivalences) is used. Moreover, for some firms, data on human capital — based on formal education — were available. Two particular variables are used to proxy human capital: the fraction of employees with vocational training (Berufs- or Fachschulabschluss) on the one hand and the fraction of employees with a university degree including universities of applied sciences (Hochschul- or Fachhochschulabschluss) on the other. However, there is a vast amount of item non-responses in these variables, leading to a substantial decrease (nearly 50%) in the number of firms with sufficient observation for panel analysis. In the remainder, this data set will be referred to as the “small sample”. Estimates based on this sample are just used to explore the effects of omitted human capital variables. As far as output is concerned, a firm’s value added, deflated by prices at the industry level, would serve as an appropriate measure for output. However, the empirical analysis had to rely on information on the firm’s undeflated total sales only since the survey does not contain further information on intermediate inputs, which would allow to derive a firm’s value added. Furthermore, there are no official deflators available for output in the German service sector. Finally, there is no information on factor utilization available. In the next chapter, it is shown that under some quite weak assumptions it is still possible to control for these influences econometrically. In order to make results comparable between different types of econometric techniques, only firms for which consistent information on at least three consequent periods were available were included in the sample. The resulting unbalanced sample consists of 1246 firms with a total of 5355 observations, that is with an average of 4.3 observed periods per firm and is referred to as the “full sample” in the remainder. Some firms reported a share of ICT investment in total investment expenditures equal to zero for all the periods surveyed. Since the econometric specification is in logs, 17 In fact, later results in the production function estimates turned out to be robust to variations in both g and d. 12
  • 14. these firms must be excluded from the full sample. However, there are reasons to assume that ICT investment in these firms is not zero, in fact, but rather very low and rounded to zero by the respondents. In order to prevent potential biases in the results (in particular in those regressions discriminating between experienced and not experienced firms) the ICT stock per worker in these firms was assumed to be equal to the corresponding industry minimum and the corresponding values were imputed. This sample is referred to as the “extended sample” in the remainder. Several variables of the MIP-S data set make it possible to distinguish between firms with more and less innovative experience. In each wave of the survey, firms were asked whether they had introduced new or significantly improved services on the one hand and new processes on the other. These questions refer to the past 3-year-period of each wave. These variables were used to construct two different classifications of firms in order to proxy for its innovative experience. According to the first (broader) distinction, a firm is classified as a “panel product innovator” (PPD) or “panel process innovator” (PPC) if it had declared itself an innovator in one of the periods surveyed. On the contrary, according to the second, narrower distinction, a firm is classified as an “experienced product innovator” or an “experienced process innovator” (EPD or EPC respectively) only if it has declared itself an innovator in the first period surveyed.18 By this more restrictive definition, it is intended to focus more strongly on innovation experience as a history–oriented concept since the first definition also includes firms that have been introducing innovations in the course of implementing ICT technologies. Therefore, the broader definition can be viewed rather as the firm’s characteristic of how ICT capital goods are implemented. By comparing the results for both of these classifications, it can therefore be concluded if it is really experience that matters for ICT output contribution or rather the way the new technologies are implemented. The statistics of the sample are summarized in the tables in the appendix. Table 4 shows the (pooled) summary statistics of the logs of the variables on input, output and innovation that are employed in the regressions, and Table 5 gives an overview over the shares of innovating firms according to the different classifications. Tables 6 and 7 show that the sample reflects industry and size structure of the German distribution and business–related services fairly well.19 Finally, in table 8, the medians20 of the firms averages of capital and output intensity are displayed for the full sample.21 The figures 18 Note that an “experienced innovator” must also be a “panel innovator”. 19 Retail trade is slightly oversampled whereas traffic and postal services as well as software and telecom- munication are slightly oversampled. As far as firm size is concerned, large firms are oversampled in their mere number and undersampled in their respective share in sales (see last two columns of table 7) 20 The corresponding mean values are substantially higher, since some firms — in particular real estate — display very high values for both inputs and output per employee. 21 The corresponding intensities for the other samples (not reported) are very similar. 13
  • 15. indicate that at the median, workplace in services is equipped with ICT capital worth slightly more than DM 3000, and with non–ICT capital worth about DM 55,000. The median turnover per employee is DM 200,000. These figures show that — as suggested in Griliches (1994) — the small share of ICT input (less than 6% of the value of other capital goods) together with measurement errors may make it difficult to determine the productivity effects of ICT use. However, the columns 2 and 3 of table 4 indicate that, in fact, the distinction according to firms’ innovative experience may help to identify the productivity effects. In fact, the median of the per capita value of ICT stock in experienced firms (defined according to its narrowest definition) is more than twice as high than among non–experienced firms whereas sales and conventional capital hardly differ. These simple summary statistics coincide with the outlined hypotheses: if ICT use is more productive in experienced firms, these firms will tend to spend a higher amount on ICT per worker than other firms. 4 Empirical Results Before the full model from equation 6 is implemented empirically, the data problems described in the previous section have to be addressed methodologically. Besides, potential biases from various sources are to be addressed by using different econometric approaches. These investigations may also shed some light on the potential sources of the ‘productivity paradox’. To keep things simple, this empirical section is organized in the following way. Firstly, econometric issues arising from data characteristics are discussed. In the second subsection, estimation results abstracting from firms’ innovative experience are discussed in order to investigate the effects of various potential sources of biases on the estimates. Finally, the initial hypotheses about the role of ICT for a firm’s exploiting its “experience benefits” are analyzed in more detail. 4.1 Econometric Issues One of the main drawbacks in the MIP-S data set is that it contains only information on undeflated nominal sales as a measure of firms’ output instead of deflated value added (see section 3). If the deflators of output and the share of value added in sales did not vary substantially, this lack of data would not cause any substantial bias: both the deflator and the common share of value added in sales would enter the constant term in the logarithmic specification.22 However, the share of value added in 22 If, furthermore, these factors were constant over time, they could be “differenced away”. 14
  • 16. sales varies substantially across industries. Industries that are typically at the end of the value chain (like wholesale and trade) will exhibit much lower shares than other service industries that are much less dependent on intermediate goods in quantitative terms. Moreover, prices have evolved quite differently between industries in Germany in the period concerned. Prices of telecommunication, for example, are likely to have declined to a larger extent than prices in other services have. Finally, different cyclical effects between industries are very likely to have induced different evolvements of in- tensity in factor utilization between industries over the observed period from 1994 to 1999. To control for these potential distortions econometrically, it is assumed that the most substantial variation in the share of value added, price deflators, and factor utilization is due to differences (both in levels and temporal change) between industries.23 More specifically, firms are classified with respect to 7 industries (j = 1 . . . 7)24 , such that nominal sales Yijt of firm i from industry j at time t can be expressed as: Pjt Ujt Yijt = Vit (9) Sj where Pjt is the price deflator for industry j in year t, Ujt is the corresponding index of factor utilization, Sj the industry’s average share of value added in sales (equal to one minus the share of intermediate goods and materials in sales), and Vijt is the value added of firm i belonging to industry j at time t. By taking logs, the industry- and time-dependent factors can be separated from V : Pjt Ujt ln Yijt = Djt + ln Vit with Djt = ln (10) Sj Therefore, by introducing interacted industry and time dummies Djt , the potential biases induced by prices, business cycles and differences in the share of value added in sales can be controlled for. In fact, the allowance for interacted dummies might be particularly important for analyzing the output contribution of ICT. ICT use tends to have increased especially strongly in industries for which substantial price declines during the past decades can be observed, such as telecommunications and software. If the price 23 Similar, but more restrictive assumptions were made by Lichtenberg (1995). He corrected for differences in the share of value added in sales between industries by introducing industry dummies on the one hand and controlled for price movements by introducing common time dummies on the other. However, differences in price movements and in business cycles between industries cannot be captured with this specification. In the discussion of the empirical results, the empirical relevance of including interacted dummies will be considered separately. 24 These are (with the corresponding nace-codes in brackets): wholesale trade (51), retail trade (50, 52), transport and postal services (60-63, 64.1), electronic processing and telecommunications (72, 64.2), consultancies (74.1, 74.4), technical services (73, 74.2, 74.3), and other business-related services (70, 71, 745-748, 90). Since there are no output data available for banking and insurance (only the balance sheet total and insurance premiums respectively), these industries must be excluded from the analysis. 15
  • 17. declines in these industries are not taken into account, the increase in real output for firms of these industries will be understated. As a consequence, the output contribution of ICT in general will be underestimated. Furthermore, variations in factor utilization due to cyclical effects may add substantial ‘noise’ to the residual variation in output. Measuring the output contributions of ICT, however, has been compared to looking for ”the needle in the haystack” (Griliches, 1994). Even though real quantities of ICT used by firms have grown dramatically over the last decade, ICT still represents a relatively small portion of overall inputs. In the U.S., where ICT has diffused much more rapidly than in Europe, the share of ICT services in GDP has risen from 4.4% in 1994 to a modest 5.5% in 1999 (Jorgenson, 2001). Griliches (1994) points out that the small share of ICT combined with poor measurement in output and deflators makes it difficult to distinguish the output contributions of ICT from stochastic events. Thus, by omitting industry specific cyclical effects and price deflators, the stochastic noise may be too strong to identify the contributions of ICT econometrically.25 4.2 Evidence from the ICT–augmented production function In order to explore the effects of different potential biases, the ICT-augmented production function is estimated firstly in a simple pooled OLS regression.26 In the regression equation, interacted industry and time dummies as well as a dummy variable for firms located in East Germany are included. This last variable is expected to be significantly negative since the transformation process in the Eastern part of Germany is still lagging behind in both productivity and wages when compared to West Germany. The results for the pooled regression are summarized in the first column of table 1. The coefficients of all three inputs are significantly different from zero at the one percent level. The output elasticity of labour takes the reasonable value of 61%.27 What is most striking in the results is that the point estimate of the coefficient of ICT capital (24.2%) exceeds the coefficient of conventional capital (12.6%).28 Given that the share of the average invest- 25 However, the costs of including interacted time and industry dummies are substantial as well. There are n = J ·T −1 = 7·6−1 = 41 additional variables that must be included additionally, with J representing the number of industries and T the number of years. 26 All estimations were computed with the DPD98 programme developed by Arellano and Bond (1998) running in GAUSS. For all the results, heteroskedasticity–consistent standard errors are reported. 27 Under the assumption of constant returns to scale and perfect competition, the income share of labour in an economy must equal its labour coefficient in the production function. For the German economy as a whole, the average share of labour payments in national income between 1994 and 1999 amounted to 72.4% (Statistisches Bundesamt, 2001). 28 Similar results have been found in cross section regressions by Bertschek and Kaiser (2001) for a sample of firms in the business-related service sector taken from a different survey. On the contrary, Brynjolfsson and Hitt (1995) report point estimates of the coefficients of ICT capital (10.9%) that were only about half of the value for non–ICT capital (20.9%) in the pooled regression for a sample drawn from the Fortune 500 Manufacturing and Service listings. There are two main reasons for these differing results. First, 16
  • 18. ment expenditure in conventional capital exceeds the average IT-investment, and given that the average of the estimated non-ICT stock per worker exceeds the corresponding value for ICT capital by a factor close to twenty, these results would imply that firms heavily underinvest in ICT capital goods. However, a more reasonable explanation is that the results from both pooled and simple cross–section regressions are biased. There are five potential sources that will be considered step by step in the empirical exploration: unobserved heterogeneity between firms, simultaneity of the decisions about inputs and output, measurement errors in the input variables, autocorrelation of exogenous shocks, and biases from omitted variables. Unobserved heterogeneity may bias the results if the investment strategies of highly productive firms are systematically different from their less productive competitors within the same industry.29 In particular, if highly productive firms tend to invest more in new technologies than firms with low productivity do, the ICT coefficient will be biased upwards in a pooled or cross section OLS regression.30 Brynjolfsson and Hitt (1995) find that unobserved heterogeneity may explain as much as half of the productivity effects attributed to ICT in their pooled regressions. In order to control for the firm–specific effects, the within–estimator was used.31 In the second column of table 1, the corresponding results are summarized. The figures indicate that once unobserved heterogeneity is controlled for, the output contributions of both types of capital are no longer significantly different from zero whereas the labour coefficient rises slightly (67.6%).32 The figures show that all the output contributions assigned to both types of capital in the pooled regression were in fact due to unobserved heterogeneity. These results coincide with very similar findings by Black and Lynch (2001) and Wolf and Zwick (2002) for production function estimates with one type of capital only. Brynjolfsson and Hitt (1995) referred to a very different sample of large firms only. They drew their sample from the Fortune 500 listings of both the manufacturing and the service sector which consists of large firms only. Second, they constructed the estimate of the firms’ ICT capital stock in a very different manner. While the ICT capital in the study of Bertschek and Kaiser (2001) and this paper ultimately builds upon ICT investment data, Brynjolfsson and Hitt (1995) used data on the stock of mainframes and PCs and converted these data into corresponding estimates of their market values. In particular, their measure of ICT capital is much narrower than the measure employed in this study in that it excludes factors such as telecommunication hardware, peripherals and software. With this background, the much higher ICT coefficient in the results presented here comes at no surprise. 29 Productivity differences between different industries are captured by the industry dummies. 30 In fact, the highly significant first– and second–order autocorrelation in the errors of the pooled regression indicate that fixed effects exert a strong effect on the results. 31 More specifically, the orthogonal deviations from the corresponding firm’s mean of the variables were employed. 32 Since there is no variation in the East dummy over time, this variable is excluded from the within estimation. 17
  • 19. Table 1: Results for the ICT-augmented production function Dep. Variable: Sales production inputs pooled within GMM SYS–GMM SYS–GMM (2) SYS–GMM (3) constant 1.230*** -0.005 -0.025 0.615 0.456 0.551* (0.136) (0.022) (0.028) (0.299) (0.293) (0.294) labour 0.610*** 0.676*** 0.549*** 0.665*** 0.675*** 0.664*** (0.019) (0.063) (0.099) (0.066) (0.066) (0.066) ICT capital 0.242*** -0.016 0.040 0.075** 0.050 0.041* (0.018) (0.021) (0.053) (0.036) (0.037) (0.023) non-ICT capital 0.146*** -0.002 0.102 0.233*** 0.230*** 0.232*** (0.014) (0.072) (0.234) (0.043) (0.044) (0.042) East -0.101*** — — -0.396*** -0.397*** -0.404*** (0.041) (0.046) (0.047) (0.045) observations: 5355 4109 4109 5355 5355 5529 number of firms: 1246 1246 1246 1246 1246 1292 R-square 0.848 0.315 0.211 0.851 0.848 0.840 joint significance, Wald-test [df]: 6906[4]*** 125.5[3]*** 35.1[3]*** 567.8[4]*** 573.1[4]*** 601.0[4]*** Sargan (p–values): — — 0.043 0.091 0.073 0.046 AR(1) of errors 0.000 0.000 0.002 0.001 0.001 0.001 (p–values) AR(2) of errors 0.000 0.362 0.400 0.291 0.279 0.271 (p–values) ***,**,*=significant on the 1,5 and 10 per cent level All regressions except SYS-GMM (2) contain industry dummy variables interacted with year dummy variables. Heteroscedasticity consistent standard errors reported. The unplausibly low estimates of the capital coefficients in the within estimates may be due to a downward bias of the point estimates due to measurement errors as argued by Griliches and Hausman (1986). Measurement errors, however, are very likely to be substantial in both types of capital stocks. First, since there is no information available about the share of expansion investment in total investment expenditures, common depreciation rates were assumed for all firms. This may have induced a significant (though presumably not systematic) measurement error into the construction of the capital stocks. Second, there was no initial value of the capital stocks available in the data employed here. For the calculation of approximate values for the initial stocks, however, both the depreciation and the pre–sample growth rates of the capital stocks had to be assumed equal across firms. Again, deviations from this assumption are very likely and will add much noise to the calculated values of ICT– and non–ICT stocks. On the contrary, the measurement errors for labour input will be less severe, even though the transformation of part–time workers into full–time equivalents may — apart from the well–known problem of overtime accounting — add some measurement error here as well. 18
  • 20. On the other hand, the downward bias due to measurement error may be offset by a second type of error which is simultaneity. If input and output are determined simultaneously, exogenous demand shocks result in an increase of both input and output. In this case, the output elasticities of the corresponding factors would be biased upwards.33 In order to correct for these two potential biases, the GMM estimator is applied for the production function in first differences. Similar to the proceeding proposed in Mairesse and Hall (1996), all three inputs are assumed to be predetermined and the (log) levels of their lagged values xt−2 , xt−3 , ..., x0 are used to instrument the input in differences ∆xt = xt − xt−1 .34 The corresponding results in column 3 of table 1 show that in this specification the point estimates for the capital coefficients increase while the labour elasticity decreases as compared to the within estimator. These findings indicate that the measurement error bias in the capital coefficients exceeds the counteracting simultaneity bias.35 However, the capital coefficients remain insignificant from zero in this specification. Furthermore, the low p–value of the Sargan test rejects the validity of the instruments employed in this specification. A further estimation issue related to the simultaneity issue is the potential au- tocorrelation of the errors. If the exogenous shocks (demand shocks, cyclical effects etc.) are autocorrelated and this effect is not taken into account, the estimates will not be consistent. Therefore, the p–values are reported for the corresponding AR(1)– and AR(2)–tests of the errors in the corresponding specification. In the specification in first differences of the variables, however, the first–order correlation of the errors will be induced by the data transformation.36 Therefore, the relevant test for equations in first differences is whether the corresponding errors are AR(2) or not. As shown in table 1, autocorrelation of the errors can be rejected for all specifications except simple pooled regression. A possible reason for the insignificant capital coefficients found in the GMM regressions is the small power of the instruments used.37 Blundell and Bond (1998b) 33 The simultaneity bias might apply in particular to those factors that can be adjusted rapidly which is not so much the case for capital stocks. For a simple formal derivation of the origin of the simultaneity bias, see Griliches and Mairesse (1995). 34 This means that the firms’ corresponding fixed effect are eliminated by explaining output growth by the growth rates of the inputs. 35 These findings coincide with similar results in Black and Lynch (2001) for estimates of the production function with one type of capital only. 36 It is easy to see that if the errors it are i.i.d. with variance σ 2 their corresponding first differences will be AR(1): E(∆ it · ∆ i,t−1 ) = E(( it − i,t−1 )( i,t−1 − i,t−2 )) = −σ 2 . 37 Since capital stocks within firms are highly persistent over time, the correlation of the first differences with the second lag in level is close to zero. Formally, this can be illustrated by assuming Kt being 19
  • 21. show that this may result in substantial finite–sample biases when the GMM estimator in first differences is used. In a specific application to production function estimation, Blundell and Bond (1998a) argue that the weak instruments will bias the differenced GMM estimates in the direction of the within group estimation, that is towards zero in the case of the capital coefficients. They use an (extended) SYS–GMM estimator in which both the equation in differences is instrumeted by suitably lagged differences (like in the simple GMM–estimation) and the equation in levels is instrumented by suitably lagged differences additionally. These two specifications are then estimated simultaneously. This SYS–GMM estimator was originally proposed by Arellano and Bover (1995). The additional moment conditions required for the equation in levels are not very restrictive. As shown by Blundell and Bond (1998b), only weak assumptions about the initial distribution of the variables used are necessary. In particular, the joint stationarity of the dependent and the independent variables is a sufficient, yet not necessary condition for the validity of the moment conditions for the equation in levels. The corresponding results for the SYS–GMM estimation are given in column 4 of table 1. In this specification, all three factor inputs are significantly positive. The output elasticity of labour amounts to two thirds which is consistent with the share of income from labour in the aggregate statistics (see footnote 27). The coefficients of ICT and non–ICT capital amount to 7.5% and 23.3% respectively, which coincides with the income share from total capital goods of roughly one third. The null–hypothesis of constant returns to scale (CRS) cannot be rejected at the 1%–level (not reported). A further very robust result is that East–German firms in services are significantly less productive than their West–German counterparts. The coefficient of the East–Dummy (roughly -0.4) implies that the multifactor productivity in East–German firms is still only about two–thirds of the West–German level. This finding coincides with aggregate statistics on productivity differentials in Germany. The corresponding Sargan–statistic (p = 0.093) does not reject the validity of the instruments at the 5%–level. These robust results in- dicate that there are substantial output contributions of ICT in the German service sector. In order to further investigate the sources of potential biases in assessing the productivity effects of ICT, the effect of allowing for different business cycles and inflation rates between industries by including interacted time and industry dummies has also been analyzed. In order to illustrate the importance of this procedure, the SYS–GMM estimation with simple (not–interacted) time and industry dummies were AR(1): Kt = ρKt−1 + t with ∼ i.i.d. If Kt is weakly autocorrelated (|ρ| 1 and ρ = 0), the past levels are correlated with the contemporaneous levels. For the first available instrument Kt−2 , this is: E(∆Kt · Kt−2 ) = E((Kt − Kt−1 ) · Kt−2 ) = E(Kt · Kt−2 ) − E(Kt−1 · Kt−2 ) = ρ2 − ρ. However, if the evolution of Kt resembles a random walk (ρ ≈ 1), the correlation between the variable in differences and its past values in levels will disappear (ρ2 − ρ ≈ 0) and the instruments will therefore turn out to be weak. 20
  • 22. repeated. The corresponding results reported in column 5 (“SYS–GMM [2]”) of table 1 show that the coefficient of ICT capital is substantially affected by this change in the estimation specification. The corresponding point estimate reduces to roughly 5% and, more importantly, is not even significant at the 10% level. In contrast, the coefficients of the other explanatory variables do not exhibit any remarkable changes.38 These results suggest that the inclusion of interacted industry and time dummies is an important prerequisite for assessing the contributions of ICT correctly. In fact, the impossibility of controlling for these effects might be an important methodological reason for the insignificant output contributions of ICT and the ‘productivity paradox’ found in earlier studies. Another source of distortion is considered in the last column of table 1 (SYS–GMM [3]). The corresponding results are based on exactly the same estimation method as column 4 (SYS–GMM). However, the sample was extended by those 46 firms that have reported zero ICT investment for all years observed (‘extended sample’, see section 3). The inclusion of these firms substantially lowers the point estimate for ICT (4.1%) compared to the values reported for the original sample (SYS–GMM). Moreover, the ICT coefficient is significantly positive at the 10% level only. These results appear very reasonable if one considers that firms may differ in their output elasticities. Those firms with a low output elasticity of ICT are maximizing profits with a lower share of ICT capital in output; excluding these firms might therefore overstate the ICT coefficient due to sample selection bias.39 Finally, an important issue of estimating the productivity effects of ICT in the production function framework is the potential bias in the estimates from omitted variables that are potentially complementary to the firm’s use of ICT. In particular, recent studies find that human capital plays an important role in this regard (Bresnahan et al., 2001). Furthermore, Brynjolfsson and Yang (1999) argue that the use of ICT is strongly complementary to intangible assets. On the one hand, ignoring these complementary factors might lead to an overestimation of the true impacts of ICT on production if the output contributions of these factors are wrongly assigned to the mere use of ICT. On the other hand, a firm’s human capital and intangible assets are likely to be quite persistent. If these complementary assets hardly vary over time, their effect will not be distinguishable from other factors like management skills etc. which are controlled for as unobserved heterogeneity between firms. In this case, no dis- tortions are to be expected from the omission of these variables in the estimation equation. 38 Again, the Sargan statistic does not reject the validity of the instruments (p = 0.073). 39 The Sargan statistic of the extended sample, however, indicates that the validity of the instruments is rejected at the 5%–level (p = 0.046). 21
  • 23. Table 2: The effects of including human capital Dep. Variable: log sales production inputs full small small w. skill constant 0.615 0.426 0.637 (0.299) (0.418) (0.303) labour (log) 0.665*** 0.653*** 0.584*** (0.066) (0.101) (0.061) ICT capital (log) 0.075** 0.019 0.087** (0.036) (0.045) (0.039) non-ICT capital (log) 0.233*** 0.244*** 0.172*** (0.043) (0.070) (0.051) East-Germany -0.396*** -0.364*** -0.363*** (0.046) (0.063) (0.058) % university – – 0.472*** (0.170) % vocational – – 0.318*** (0.106) observations: 5355 2060 2060 number of firms: 1246 639 639 R-square 0.851 0.826 0.835 joint significance, Wald-test [df]: 567.8[4]*** 196.6[4]*** 364.9[6]*** Sargan (p-values): 0.091 0.679 0.206 ***,**,* = significant at the 1, 5 and 10 per cent level All regressions are based on SYS-GMM and contain industry dummy variables interacted with year dummy variables. Heteroscedasticity consistent standard errors reported. In order to assess the potential biases from omitting the probably most important source, human capital, the list of independent variables was extended by the share of employees with vocational training and with university degree or equivalent correspond- ingly. As discussed in section 3, the resulting ‘small sample’ consists of only 639 firms. The results of the corresponding regression are surveyed in table 2. The first column replicates the “SYS–GMM” results obtained from the large sample (with interacted time and industry dummies, fourth column of table 1) to facilitate comparisons while the second column displays the corresponding results for the small sample. The most striking result is that while the coefficients of labour and non–ICT remain more or less unaffected by the sample reduction, the point estimate of the ICT coefficient decreases substantially and becomes insignificant. This is likely to be the result of the information loss due to the much smaller number of observations. The effect of including the proxies for human capital in the regression becomes obvious from the results displayed in the third column 22
  • 24. of table 2. Including the human capital variables reduces the coefficients of both labour and non–ICT capital but leads to a substantial increase in the estimated ICT elasticity. These quite astonishing results seem to imply a complementary relationship between human capital and non–ICT capital rather than ICT input. Most importantly for the explorative analysis here, however, the omission of human capital obviously induces an underestimation rather than an overestimation of the productivity impacts of ICT. Taken together, the findings of this subsection indicate that unobserved hetero- geneity, measurement errors and industry–specific time effects may lead to substantial distortions in assessing the productivity impacts of ICT. Ignoring heterogeneity may lead to a substantial overestimation of these impacts while measurement error and the omission of industry–specific cyclical effects and price deflators work in the opposite direction. The potential upward bias from simultaneity issues is – at least compared to the measurement error bias – relatively low. When all these effects are controlled for in an adequate SYS–GMM estimation (column 4 of table 1), reasonable results are obtained that are consistent with the rough income shares of the different inputs in aggregate statistics. A further explorative analysis shows that the omission of human capital does not lead to an over– but rather an underestimation of the productivity impacts of ICT. 4.3 The role of innovative experience After assessing the methodological robustness of the SYS–GMM estimator for the production function framework, this estimator is applied to the regression equation 6. In order to investigate the role of innovative experience for the success in implementing new technologies, ICT capital interacted with innovator dummies is included in the regression as an additional variable (see section 2.2). A potential bias in favour of the hypotheses to be tested may arise by the mere fact that innovators will be more productive than other firms due to the returns to their (intangible) R&D or innovation capital that is likely to be higher for firms with innovative experience. In order to control for this effect, the corresponding innovator dummy was included as a further variable that captures the direct productivity contribution of innovations.40 The results are based on the extended sample in order not to exclude firms with potentially low output elasticities of ICT (see comments in the previous subsection). In the first column, of table 3, the results are replicated for the specification in which a ‘surplus’ of ICT contributions (∆x) is allowed for in firms with experience in introducing process innovations according to the narrow definition (epz). The most 40 Both product and process innovations may lead to an increase in productivity by either raising the sales due to improved service quality or by lowering input costs due to more efficient processes. 23
  • 25. Table 3: The role of innovative experience Dep. Variable: sales production inputs (epc) (epd) (ppc) (epc) (epd) epc w/o imp constant 0.717*** 0.495* 0.497* 0.749*** 0.616** 0.794*** (0.238) (0.266) (0.257) (0.233) (0.257) (0.254) labour 0.607*** 0.658*** 0.670*** 0.582*** 0.641*** 0.611*** (0.238) (0.063) (0.060) (0.051) (0.057) (0.057) ICT capital 0.032 0.012 0.027 0.051** 0.038 0.066 (0.022) (0.025) (0.029) (0.022) (0.026) (0.037) non-ICT capital 0.237*** 0.246*** 0.248*** 0.226*** 0.240*** 0.230*** (0.041) (0.042) (0.039) (0.055) (0.059) (0.043) innovator 0.188* 0.087 0.006 0.296 -0.007 0.127 (0.098) (0.101) (0.124) (0.223) (0.135) (0.110) ICT capital 0.087** 0.061* 0.015 0.060* 0.012 0.074 (innovator) (0.039) (0.033) (0.035) (0.034) (0.042) (0.047) non-ICT capital — — — -0.008 0.018 — (innovator) (0.074) (0.078) East-Germany -0.390*** -0.396*** -0.403*** -0.368*** -0.397*** -0.383*** (0.043) (0.042) (0.043) (0.045) (0.042) (0.044) observations: 5529 5529 5529 5529 5529 5355 number of firms: 1292 1292 1292 1292 1292 1246 R-square 0.842 0.842 0.841 0.834 0.842 0.852 joint significance, Wald [df]: 1135[6]*** 919.4[6]*** 901.0[6]*** 1098[7]*** 1102[7]*** 1076[6]*** Sargan (p-values): 0.085 0.069 0.124 0.046 0.123 0.194 ***,**,* = significant at the 1, 5 and 10 per cent level. All regressions are based on SYS-GMM and contain industry dummy variables interacted with year dummy variables. All variables in logs, except innovator and East-Germany dummies. Heteroscedasticity consistent standard errors reported. The definition of innovating firms (‘innovator’) varies between columns according to the abbreviation given in the top row of each column (see text). striking result is that the ICT coefficient, in fact, differs significantly between experienced and not experienced firms. For epz–firms an implicit output elasticity of x = 11.9% is found.41 that is significantly higher than the coefficient of the non–experienced (p–value of the difference: 0.024). On the contrary, the corresponding ICT coefficient for not experienced firms is quite low (3.2%) and insignificant. Interestingly, the dummy for process innovation experience is highly significant as well. Obviously, beyond the importance for ICT use there are direct benefits from new processes introduced in the past as well. The significant coefficient of 0.188 of the epz–dummy implies that, on 41 Note that the estimated ICT coefficient for innovators represents the difference between the elasticities of experienced and non–experienced firms ∆x (see equations 5 and 6). The implicit ICT elasticity of the experienced firms x = x + ∆x is just the sum of the two ICT coefficients in each regression. 24
  • 26. average, firms with process innovation experience are about 20% more productive than other firms.42 In the second column of table 3, the results are reported for the corresponding distinction of firms according to their experience of product innovations (epd). The results are very similar to those in column 1. However, the coefficient marking the difference in ICT contributions is significant only at the 10% level and the implicit output elasticity for epd–firms (7.3%) is considerably smaller than in the regression with the classification according to process innovation experience (epc).43 Given that more than 70% of the product innovators are also experienced in process innovation (see table 5), this difference may be a consequence of the simultaneous process innovation experience rather than the experience collected in the turn of product innovations. As a first preliminary result, it may be concluded that past process innovations have the biggest impact on a firm’s productive use of ICT. In column 3, the results for the less restrictive differentiation between firms with some process innovation (“panel process innovators”) are replicated. As discussed in section 3, this wider definition captures both firms with innovative experience and firms that have in- troduced innovations at a later point in time during the period observed. Therefore, apart from ‘experience’, this classification variable also includes implementation strategies. In the application of this definition, there is no significant difference between innovators and other firms regarding the productivity of ICT use. This finding suggests that it is indeed the experience generated by a firm’s innovation history that facilitates an efficient ICT use. Obviously, ICT productivity is higher in firms with experience in process innovation. But is this a special feature of ICT as opposed to conventional capital? In order to address this second question, non–ICT capital interacted with the dummy for experienced process innovators has been added to the regressors in a further specification (col. 4 in table 3). The results show that, in fact, there is a remarkable difference between both types of capital. While ICT capital — like in column 1 — continues to be more productive in experienced firms (at the 10%–level however), no such difference can be observed in the case of conventional capital. In the corresponding specification for product innovators (column 6), however, there is no significant difference between both types of firms for either kind of capital. Taken together, these findings indicate that the effects of innovation experience are obviously very specific with respect to two aspects: it is mainly experience from process innovations that matters, and it is the productivity of 42 Compared to the SYS–GMM results in table 1, the other coefficients are hardly affected by the addi- tional variables included in the regression, which points to the robustness of the results. 43 A further difference to the preceding results is that there are no significant direct productivity effects resulting from past product innovations. 25
  • 27. ICT capital that is essentially affected. Finally, the robustness of the findings of the first specification (process innovation experience) was investigated with respect to changes in the sample. If firms that have reported zero ICT investments are excluded from the regression (that is if the ‘full’ instead of the ‘extended’ sample is used; see column 6 of table 3), the difference in ICT productivity between experienced and non–experienced firms becomes only marginally significant (p=0.110, not reported). However, the point estimates of the coefficients do not change greatly, indicating that the exclusion of firms with a potentially low ICT productivity mainly affects the precision of the distinction between the two groups of firms analyzed.44 To sum up, the results for the empirical model specified in equation 6 deliver strong support in favour of the initial hypotheses. Firstly, innovative experience is found to significantly influence the productivity potentials of ICT use in services. Secondly, this dependence on a firm’s innovative history apparently is a feature that distinguishes ICT from other capital goods. Beyond this, the findings suggest that it is experience collected from past process innovations which is particularly worthwhile for the efficient use of ICT. This finding is in line with other studies that emphasize the close link between ICT use on the one hand and organizational change and restructuring of business processes on the other. Because of this close link, experience from past innovations may reduce mistakes and will improve expectation formation with regard to the costs and benefits of ICT–induced changes. 44 In some further excercises to validate the robustness of the results, only insignifcant differences between the ICT coefficients of experienced and not experienced firms were found for the small sample of 639 firms that includes information on human capital. This result was independent of the actual inclusion of human capital. Furthermore, in the small sample specification without human capital, the result yielded no significant output contribution of ICT at all. These findings can be interpreted as an indicator that the precision loss due to a smaller sample size is an important issue in this empirical framework. Not surprisingly, the need for a large sample for more precise estimates is most important for the capital inputs for which measurement error was found to be an important issue. 26
  • 28. 5 Conclusions In this paper, the productivity effects of ICT use in the German business–related and distribution services are analyzed with firm–level data. Using a production function framework, two types of models are analyzed. Firstly, for the simple Cobb–Douglas specification with labour and two types of capital as inputs, a highly significant output elasticity of ICT–capital of 7.5% is found, indicating substantial productivity effects of ICT in the service sector. These estimates are based on a SYS–GMM estimator that controls for a variety of potential estimation biases, like unobserved heterogeneity, simultaneity issues and measurement errors. Secondly, based on a theoretical model, the production function framework has been extended to allow productivity contributions of ICT capital to vary between firms. This more detailed analysis reveals that firms that have introduced process innovations in the past — labelled ‘experienced’ firms — are especially successful in ICT–use. The output elasticity of ICT in these firms amounts to about 12% and is significantly higher than for non–experienced firms (3%). On the contrary, no such difference can be observed for conventional capital. These findings support the hypotheses developed in this paper which assign ICT the role of a ‘special’ capital input: unlike other capital goods, the productive use of ICT is closely linked to innovations in general and the re-engineering of processes in particular. Firms with experience in process innovations are therefore predicted to exploit the potential benefits of new technologies more successfully than other firms. There are several implications of these findings concerning theoretical, methodological and policy issues. At the theoretical level, the results contribute to a clarification of the role of ICT as a general purpose technology (GPT). In spite of the diverse uses and the rapid diffusion of ICT throughout all industries, the productivity effects of ICT are far from self–enforcing but rather demand an active implementation strategy within firms. The role of innovative experience found in this paper indicates that the determinants for the efficient use of ICT are in the range of a firm’s long–term strategies rather than characteristics that can be changed easiliy in the short term. Innovative experience is likely to be acquired within years rather than months. Furthermore, the role of innovative history found at the micro level may also be useful to shed more light on the differences of ICT–induced productivity effects found between countries. In fact, the competitive and innovative business environment in the U.S. may be one reason that helps explain why the productivity impact of ICT has been much higher there than in continental Europe. The higher innovation pressure in the U.S. over the last decades may have led firms to collect much more diverse innovative experience than the more protected firms in Europe. This may have enabled firms in the U.S. to recognize the productive value of the wave of ICT–induced innnovations faster 27
  • 29. and to react more flexibly with an appropriate restructuring and re-engineering of the internal processes in order to reap high benefits from the use of ICT. In this respect, ICT may have led to a further widening of the productivity gap between the U.S. and Europe. As far as the empirical methodology is concerned, the findings of this paper illustrate that unobserved heterogeneity, measurement errors and omitted variables (including differences in price and quality evolutions over time) are important sources of estimation biases in assessing the impacts of ICT within a production function framework. While unobserved heterogeneity may result in a substantial overestimation of ICT impacts, measurement errors forcefully work into the opposite direction and may lead to an apparent affirmation of the ‘productivity paradox’. Furthermore, as shown in this paper, particular firm characteristics may play a key role for the potential impact of ICT on firm performance. The results suggest that firm–level studies form a promising basis of analysis that may add more detailed insights and further complementarities between ICT and firm characteristics in future research. As far as economic policy is concerned, the findings of this paper point to the importance of an innovative business environment that is needed to lay the fundamentals for an efficient use of ICT. New technologies like ICT may be compared to the invention of a new fertilizer in farming: though its potential uses may be very general and its costs quite low, a sound climate, a cultivated soil and a gifted farmer will still be needed to actually increase crop yield. Unlike the case of farming, however, the climate in economics may be favoured to a large extent by sound policies. The results of this study suggest that the enhancing of competition and innovation incentives may serve as an important driver of both the rapid diffusion and a productive use of ICT. 28
  • 30. References Arellano, M. and Bond, S. (1998). Dynamic Panel Data Estimation Using DPD98 for GAUSS: A Guide for Users. ftp://ftp.cemfi.es/pdf/papers/ma/dpd98.pdf. Arellano, M. and Bover, O. (1995). Another look at the instrumental variable estimation of error-components models, Journal of Econometrics 68: 29–51. Baily, M. N. and Lawrence, R. Z. (2001). Do We Have a New E-conomy, AEA Papers and Proceedings 91(2): 308–312. Barras, R. (1986). Towards a Theory of Innovation in Services, Research Policy 15: 161– 173. Bertschek, I. and Kaiser, U. (2001). Productivity Effects of Organizational Change: Mi- croeconometric Evidence, Discussion paper no. 01-32, Center for European Economic Research (ZEW). Black, S. E. and Lynch, L. M. (2001). How to Compete: The Impact of Workplace Practices and Information Technology on Productivity, Review of Economics and Statistics 83: 435–445. Blind, K., Czarnitzki, D., Ebling, G., Gottschalk, S., Janz, N. and Niggemann, H. (2000). Quellen f¨er Innovationen im Dienstleistungssektor, in N. Janz (ed.), Quellen f¨er u u Innovationen: Analyse der ZEW–Innovationserhebungen 1999 im Verarbeitenden Gewerbe und im Dienstleistungssektor, ZEW Dokumentationen, Centre for European Economic Research (ZEW), Mannheim, pp. 27–55. Blundell, R. and Bond, S. (1998a). GMM Estimation with Persistent Panel Data: an Application to Production Functions, Working paper series no. w99/4, Institute for Fiscal Studies. Blundell, R. and Bond, S. (1998b). Initial conditions and moment restrictions in dynamic panel data models, Journal of Econometrics 87: 115–143. Bresnahan, T. F. and Trajtenberg, M. (1995). General purpose technologies ’Engines of growth’ ?, Journal of Econometrics 65: 83–108. Bresnahan, T. F., Brynjolfsson, E. and Hitt, L. M. (2001). Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-level Evidence, Forthcoming in the Quarterly Journal of Economics. Brynjolfsson, E. (1994). Technology’s true payoff, Informationweek, October 10 pp. 34–36. 29
  • 31. Brynjolfsson, E. and Hitt, L. (1995). Information Technology as a Factor of Production: Thr Role of Differences among Firms, Economics of Innovation and New Technology 3: 183–199. Brynjolfsson, E. and Hitt, L. (1996). Paradox Lost? Firm-level Evidence on the Returns to Information Systems Spending, Management Science 42/4: 541–558. Brynjolfsson, E. and Hitt, L. M. (2000). Beyond Computation: Information Technol- ogy, Organizational Transformation and Business Performance, Journal of Economic Perspectives 14: 23–48. Brynjolfsson, E. and Yang, S. (1996). Information Technology and Productivity: A Review of the Literature, Advances in Computers 43: 179–214. Brynjolfsson, E. and Yang, S. (1999). The Intangible Costs and Benefits of Computer Investments: Evidence from the Financial Markets, Proceedings of the international conference on information systems, atlanta, georgia, MIT Sloan School. Cohen, W. M. and Levinthal, D. A. (1990). Absorptive Capacity: A new Perspective on Learning and Innovation, Administrative Science Quartely 35: 128–152. Colecchia, A. and Schreyer, P. (2001). ICT Investment and Economic Growth in the 1990s: Is the United States a Unique Case?, STI Working Papers 2001/7, OECD. David, P. A. and Wright, G. (1999). General purpose technologies and surges in produc- tivity: Historical reflections on the future of the ICT revolution , Discussion papers in economics and social history (number31), University of Oxford. EITO (2001). European Information Technology Observatory 2001, EITO, Frank- furt/Main. Flaig, G. and Stadler, M. (1994). Success Breeds Success.The Dynamics of the Innovation Process, Empirical Economics 19: 55–68. Griliches, Z. (1994). Productivity, R&D, and the Data Constraint, American Economic Review 84: 1–23. Griliches, Z. and Hausman, J. A. (1986). Errors in Variables in Panel Data, Journal of Econometrics 31: 93–118. Griliches, Z. and Mairesse, J. (1995). Production Functions: The Search for Identification, NBER Working Paper No. 5067, National Bureau of Economic Research. Hall, B. H. and Mairesse, J. (1995). Exploring the relationship between R&D and pro- ductivity in French Manufacturing firms, Journal of Econometrics 65: 263–293. 30
  • 32. Hoffmann, J. (1998). Problems of Inflation Measurement in Germany, Discussion paper no. 01-98, Economic Research Centre of the Deutsche Bundesbank. Janz, N., Ebling, G., Gottschalk, S. and Niggemann, H. (2001). The Mannheim Innovation Panels (MIP and MIP-S) of the Centre for European Economic Research (ZEW), Schmollers Jahrbuch 121: 123–129. Jorgenson, D. W. (2001). Information Technology and the U.S. Economy, American Economic Review 91: 1–32. Jorgenson, D. W. and Stiroh, K. (1995). Computers and Growth, Economic Innovation and New Technology 3: 295–316. Licht, G. and Moch, D. (1999). Innovation and Information Technology in Services, Cana- dian Journal of Economics 32: 363–383. Lichtenberg, F. R. (1995). The Output Contributions of Computer Equipment and Per- sonnel: A Firm-Level Analysis, Economics of Innovation and New Technology 3: 201– 217. Mairesse, J. and Hall, B. H. (1996). Estimating the Productivity of Research and De- velopment in French and United States Manufacturing Firms: An Exploration of Simultaneity Issues with GMM Methods, in K. Wagner and B. van Ark (eds), In- ternational productivity differences: Measurement and explanations, Elsevier Science Publishers B.V., Amsterdam, pp. 285–315. Mansfield, E. (1968). Industrial Research and Technological Innovation: An Econometric analysis, Norton, New York. McKinsey Global Institute (2001). US Productivity Growth 1995–2000 — Understandig the contribution of information technology relative to other factors, report. M¨ller, A. A. (1998). Kapitalstock und Produktionspotential im privaten und ¨ffentlichen u o Sektor Deutschlands, Nomos, Baden–Baden. Nelson, R. R. (1982). The Role of Knowledge in R&D Efficiency, The Quarterly Journal of Economics 97/3: 453–470. Nelson, R. R. and Winter, S. (1982). An Evolutionary Theory of Economic Change, Harvard University Press, Cambridge MA. OECD (2000a). A New Economy? The Changing Role of Innovation and Information Technology in Growth, OECD, Paris. OECD (2000b). The Service Economy, Business and industry policy forum series, STI/OECD. 31