Economic productivity in the Knowledge Society: A critical review of productivity theory and the impacts of ICT by Ilkka TuomiAccording to several widely publicized and influential studies, information and communication technologies (ICTs) were a major source of productivity growth during the 1990s in many developed countries. The diffusion of ICTs has been argued to permanently change the rate of sustainable economic growth, and they have frequently been described as core technologies of the emerging knowledge–based economy.
This paper examines critically the concepts and methods of ICT productivity studies. It concludes that current analytical techniques do not allow quantification of the productivity impacts of ICTs.
ICT productivity studies are problematic due to three main reasons. First, the current measures of economic output miss essential parts of output in knowledge–based economies. Second, productivity calculations measure inputs and outputs in ways that are conceptually and empirically problematic. Third, the theoretical models that have been used to analyze the impacts of ICTs often make assumptions that may be unrealistic; for example, they require that innovation can be neglected as a competitive factor and as a source of growth.
Although economic studies are only partially able to grasp the significance of ICTs, these technologies are transforming the foundations of economy and society. A number of important conceptual, methodological and empirical issues need to be studied. To fully analyze the socio–economic impacts of ICTs we may need a new "productivity paradigm."
Contents
Introduction
Why does productivity matter?
ICT as the driver of the New Economy
Disappearing consensus
Productivity and technical change in the neoclassical theory
Growth contributions from ICTs
The logic of growth contribution calculations
Empirical and conceptual limits of the Oliner–Sichel study
Price indices as a source of growth
Productivity paradox or paradigm anomaly?
ICTs as contextual and composite resources
Research challenges
Conclusion
Introduction
Common experience tells us that information and communication technologies (ICTs) are critically important for the modern economy. Firms are increasingly dependent on the effective use of ICT and it has a fundamental role in the ongoing socio–economic transformation. It is therefore important that we understand what we know about the economic impacts of ICTs and where are the limits of our current knowledge.
There exists a large and rapidly growing literature on the macroeconomic and industry–level impacts of ICTs. One might, therefore, expect that we have today a solid body of well–known and uncontested facts about ICT growth and productivity effects. This is not the case. A detailed review of existing studies reveals important conceptual and empirical challenges that have relevance also beyond ICT–related studies.
For example, empirical studies on the economic impacts of ICTs rely heavily on price indices that convert nominal prices and production into "real output" by adjusting for quality changes in products. A closer look on the sources of quality adjustments in computing prices reveals that such adjustments are conceptually ambiguous, as they make the value of money dependent on technical progress. This, in effect, translates the economic output into "real output" using technical characteristics that by default impute the assumed economic growth and points to a circular logic in the treatment of output value, growth, and productivity. Without quality adjustments, semiconductor and computer industries have expanded relatively slowly during the last decades, as rapidly dropping nominal prices have to a large extent cancelled the growth in volume. Much of the measured productivity growth depends on the very rapid technical improvements in semiconductors, which price indices translate into growth of real outputs and investments. Theoretically consistent quality adjustments also lead to prices that are not additive across product categories or time, indicating that the conventionally adopted approaches are conceptually inadequate. It is therefore not clear how much economic growth there has been during the last decade, and to what extent the growth can be associated with ICTs.
Economists have extensively discussed alternative explanations of the apparently limited impact of ICTs on growth and productivity during the last two decades. This discussion has made visible important measurement problems, and it has also highlighted potentially important conceptual challenges in the conventional ways of measuring growth and productivity in the knowledge society. The present paper argues that these challenges are particularly visible in ICT–intensive industries, but that they also require new approaches for understanding productivity and growth. The famous "Solow paradox" can be interpreted as an indication of a need for a new productivity paradigm.
A conceptualization of productivity that would allow substantial analysis of the impacts of ICT seems to require reconsideration of the links between growth and development. At the organizational level, ICTs are composite goods that consist of hardware, software, skills, systems integration, operational support, and infrastructure. Beneficial deployment of ICTs requires incremental innovation and reconfiguration of existing investments and resources. This makes it difficult to isolate the impact of specific investments in ways that typical growth accounting frameworks would require. ICTs themselves facilitate rapid recombination of existing investments, accelerating creative destruction within organizations and the overall economy, and making structural, contextual, and institutional factors increasingly important for growth. Conventional approaches in business accounting and national accounts treat some of the associated costs as consumption and some as investments, and they often remain blind to the historical and social investments that are needed for the productive use of ICTs. Research on knowledge management has highlighted the systemic nature and the interdependent elements that make ICT investments productive. As accurate measurement of investments in "ICT products" is critical for productivity studies, it is important that we conceptualize these products in an appropriate way. A multidimensional and holistic conceptualization of ICTs allows the researchers to address the different complementary elements that are needed to make ICTs productive.
In the knowledge society, the traditional concept of productivity faces both old and new challenges. The motivation for productivity measurement was to understand the sources and development of economic growth and welfare. This will remain a central issue in the knowledge society. Knowledge society, however, also makes visible productivity problems that are theoretically and empirically difficult and central, and which require careful consideration.
For example, although investments in research and development have been studied extensively since the 1960s by productivity researchers, knowledge creation is still only very inadequately accounted for. It has been estimated that investments in intangibles — not including human capital accumulated in the household sector — could be about $1 trillion of business fixed investment in recent years in the U.S. This is almost as much as the accounted business investment, about US$1.2 trillion in 2001 [1]. Productivity is measured as the efficiency of producing outputs from the inputs, but if we miss important outputs and inputs, our efficiency measures and policy recommendations easily become misleading.
The question what should be included in the inputs and outputs in analyzing production efficiency is an open question, with major socio–economic and political consequences. An important economic debate has centered on "externalities." They include unaccounted negative externalities such as pollution and depletion of non–renewable resources, and unsustainable use of renewable resources, such as ocean fisheries, but also positive externalities, such as knowledge, scientific advances, socio–economic institutions, infrastructure, and social capital. Measures of economic output that try to correct for traditionally unmeasured or mismeasured inputs lead to different productivity estimates than the traditional approaches. Sometimes the differences can be considerable [2].
Macroeconomic concepts such as growth and productivity are aggregate concepts and typically, by definition, understood to be independent of history and structural factors of production. As a consequence, the macroeconomic concept of productivity is conceptually blind to structural dimensions of productive activity. This is an important limitation when we try to understand networked and knowledge–based economies. Productive knowledge is often context dependent, situated, and embedded in configurations of material and symbolic artifacts. Contrary to the traditional idea of abstract context–independent knowledge, recent research on organizational knowledge has emphasized the importance of tacit and location specific knowledge. Similarly, researchers have increasingly emphasized the importance of "social capital" that is embedded in historically accumulated reputations, systems of trust, and social networks. Firm–level productivity often depends on historically developed informal social networks that extend beyond firm boundaries, and regional productivity often depends on international and inter–institutional networks of knowledge [3]. The contextual view on knowledge and productivity imply that productivity differences are often rooted in the position of social and economic agents in relation to material, social, and cognitive resources.
For example, task productivities of professional software programmers have typically been reported in various studies to vary at least an order of magnitude. To a large extent, programming productivity depends on specific knowledge about program libraries, collaborative work practices, and configuration of development tools into effective programming environments. Local improvements in such complex systems do not necessarily lead to productivity improvements. In fact, they easily lead to productivity losses, for example, through destruction of accumulated human capital.
Historically the most influential ICT–related productivity studies have been based on neoclassical growth models. The neoclassical theory understood the economy to be in or close to the equilibrium. The conceptual starting point in this framework is that resources are allocated optimally. In this framework, efficiency improvements come from outside the economy, as external "shocks" or "manna from heaven" that remain without explanation.
A truly networked and knowledge–based society does not easily accommodate the neoclassical abstractions. It is a mix of market transactions, business firms, and social networks that extend beyond organizational boundaries and markets of economic exchange. Decision–makers in firms, for example, may not always be able to reinvest the outputs generated in innovation networks that only partially remain within organizational boundaries. Markets may only see a small glimpse of all those social interactions that actually generate outputs and productive capital. In general, markets and firms, therefore, may be unable to optimally allocate resources.
This introduces a qualitative, structural, and social dimension to the problem of productivity, which becomes increasingly visible as networked information and communication technologies diffuse across societies and economies. Challenges for current theoretical approaches are increasingly visible, but no consensus exists about the right ways to formulate the core issues.
Although much research on productivity and its relationships with economic growth is available today, large uncharted terrains remain outside the currently known domains of productivity research. It is therefore important to try and outline the area within which current research sheds light, as well as to see the limits of our current knowledge. To see where we stand, we have to step outside the best explored areas and across disciplinary boundaries.
This study therefore attempts a challenging task. The present paper tries to clarify what we know today about ICT productivity impacts. This requires that we understand the basic theoretical concepts that underlie economic research on ICT and productivity. It, however, also requires that we discuss the particular ways in which these theoretical concepts are implemented in empirical research. For example, when economists talk about "technical change" or "constant prices" we have to understand what they actually mean by these concepts and where they get the data that is used to measure them.
The second part of the study, to be published separately due to space considerations, then, tries to provide some tentative ideas and elements of a broader productivity framework that could complement current ICT productivity frameworks.
This paper is organized as follows. The next three short sections will revisit the motivations for studying productivity and discussions on the "new economy" that argued that ICTs had an important role in the productivity revival in the second half of the 1990s. The basic concepts that underlie productivity research are then described in an attempt to clarify what concepts such as "productivity" and "technology" actually mean for economists working in this area. I will then move on to describe how the macroeconomic impacts of ICT are studied. For this I use an influential and representative study by Oliner and Sichel (2000). Walking through its procedures and architecture allows us to see the machinery of scientific knowledge production in operation, and to understand the theoretical and practical assumptions that produce the outcomes. This reconstruction of a scientific knowledge creation machine is followed by a discussion that evaluates the robustness of the reported empirical results both from conceptual and empirical points of view. In particular, I will show that the main results of neoclassical growth accounting studies critically depend on the way ICT price indices are handled in these studies. In a sense, we try to see whether the neoclassical productivity research machine works and whether it produces what we think it should produce.
The answer is to some extent negative, as productivity researchers know. This opens a question whether the problematic issues require just minor improvements in theory and practice, or whether they indicate a need for more profound revision of theory and research on ICT productivity impacts. I will explore this question by focusing on the famous Solow productivity paradox, reviewing known explanations for it. Based on the earlier sections, I will then draw up a list of open and potentially important research challenges.
The overall result of this discussion is that there are conceptual and empirical problems in the current approaches to productivity research. Some of these appear to be quite fundamental, pointing to a need for alternative approaches for understanding ICT and productivity. The paper ends by summarizing the main results and by making concluding remarks.
Why does productivity matter?
Policy–makers use productivity studies to understand how productivity and economic growth could be increased. Although the links between welfare, economic output and productivity are complex in practice and in theory — requiring discussion on social distribution of wealth for example — conceptually the idea is simple. If productivity increases, other things equal, aggregate economic welfare increases. As Paul Krugman once put it, productivity is not everything, but it is almost everything.
Productivity measurement has also become important for monetary and fiscal policy. Productivity trends are used to forecast potential economic growth and, for example, tax revenues. If labor income grows faster than labor productivity, the expected result is inflation. Productivity measurement, therefore, is used in the difficult act of balancing unemployment and inflation. Long–term productivity growth is commonly viewed as the speed limit for sustainable economic growth.
By analyzing productivity developments in different firms, industries, and across regions and countries, we can try to find productivity bottlenecks and locate opportunities for improvement. Firm– and industry–level comparisons are important, for example, for effective firm–level decision–making and for understanding the real benefits of information technology. International comparisons of productivity trends, in turn, have often been used as sources of policy recommendations. Recently, for example, productivity researchers have argued that the institutional flexibility in the European countries should be increased if they want to catch up with the U.S. growth rates [4].
ICT as the driver of the New Economy
In the Knowledge Society an important question is how information and communication technologies (ICTs) impact growth and productivity. During the last decades, discussions about information society and knowledge–based economy have pointed out the increasing importance of information for economic growth. The more recent discussions on the "new economy" centered on the question whether information and communication technologies have irreversibly changed the productivity growth rate of the economy. If that were the case, accelerated investments in ICT could lead to increase in economic growth and productivity. This could imply a new balance between inflation and labor income growth. For example, Jorgenson and Stiroh (2000) argued in their influential article "Raising the speed limit" that ICT had, indeed, altered the speed of productivity growth.
On the other hand, Jorgenson and others also emphasized that the ICT productivity story has been closely connected to the advances in information technology, and in particular semiconductors [5]. If technical advances in these areas slow down, productivity growth will slow down, and perhaps turn negative. In particular, Jorgenson [6] argued that the rapid productivity increase in the U.S. during the second half of the 1990s resulted to an important extent from the accelerated product cycles in the semiconductor industry. Due to exceptional competitive conditions in these years, the traditional three–year semiconductor product introduction cycle was temporarily compressed into two years [7]. If we want to understand the future of productivity, we therefore need to understand the drivers of innovation in ICT production [8].
A closer study on the developments in semiconductor, computing and communication technologies shows, however, that developments have also to large extent been driven by innovative users of these technologies. The productivity potential associated with ICT is not created only by technical advances, per se. Instead, productivity opportunities become articulated and realized when technologies are taken into use. This means that economically important innovations are not simply technical advances. They are always social innovations. To understand the productivity potential of ICT, we therefore also have to understand social learning processes that underlie the adoption of new technologies.
In the Knowledge Society, economists are now starting to ask how learning, competencies, social and economic networks, and, for example, social capital, trust and reputation should fit the picture. Although the picture is not yet clear, it is becoming important to ask what we currently know about the productivity impacts of ICT.
Disappearing consensus
Macroeconomic studies on the effects of information and communication technologies have received considerable attention during the recent decade. These studies almost univocally refer to Robert Solow’s 1987 statement that computers can be seen everywhere except in the productivity statistics. This observation has become known as the productivity paradox. Despite a general belief on ongoing information technology revolution, since the early 1970s productivity growth appeared to be slowing down in the U.S. and in many other developed countries.
Firm–level studies on ICT productivity impacts in the 1990s revealed a different picture, where computers had major positive productivity effects [9]. These studies and historians of economy and technology also noted that productivity effects become visible only with a delay, after firms have made organizational changes and acquired skills and experiences that open the gates through which the productivity benefits flow. Investment in ICT can become productive only after organizations have adjusted their operations to take advantage of the productivity potential of new technologies and when complementary investments have been made. According to this view, the productivity impacts of ICT should only be seen with a delay and the impact would be contingent with complementary changes in organizational practices. Investment in ICTs, in itself, does not guarantee productivity growth.
After firms have had sufficient time to experiment with the possibilities of ICT and adjust their operations to take advantage of its productivity potential, firm level productivity impacts could be expected to become visible also at the aggregate level. Indeed, during the last years several macroeconomists have claimed that this has been the case. Towards the end of the 1990s, a widely accepted view emerged that the productivity paradox was a temporary event and that computers and other ICTs finally had became one of the main drivers of economic and productivity growth [10]. More specifically, the rapid price declines in computing equipment and semiconductors seemed to be an important driver in the process. As Jorgenson [11] noted, "despite differences in methodology and data sources, a consensus is building that the remarkable behavior of IT prices provides the key to the surge in economic growth." Gordon summarized the consensus:
"... by 1999–2000 a consensus emerged that the technological revolution represented by the New Economy was responsible directly or indirectly not just for the productivity growth acceleration, but also the other manifestations of the miracle, including the stock market and wealth boom and spreading of benefits to the lower half of the income distribution. In short, Solow’s paradox is now obsolete and its inventor has admitted as much." [12]More accurately, the emerging consensus seemed to be that ICT investments had increased productivity growth in durable goods manufacturing in the second half of the 1990s, and in particular in ICT equipment manufacturing. Some researchers (e.g., Van Ark, et al., 2003; Oliner and Sichel, 2002; Nordhaus, 2001; Colecchia and Schreyer, 2002) argued that the impact of ICT could also be seen in the ICT using sectors, at least in the U.S. Jorgenson and Stiroh, however, labelled popular accounts of widespread impacts of ICT the "phlogiston theory of new economy":
"... the evidence already available is informative on the most important issue. This is the ‘new economy’ view that the impact of information technology is like phlogiston, an invisible substance that spills over into every kind of economic activity and reveals its presence by increases in industry–level productivity growth cross the U.S. economy. This view is simply inconsistent with the empirical evidence." [13]ICT–producing and –using industries seemed to play very different roles in different countries, indicating that the level of ICT investments and the widespread use of ICTs were not directly related to growth or productivity [14]. Furthermore, the discussion on methodological and measurement problems started to erode the consensus. Vijselaar and Albers (2002), for example, pointed out that the observed differences in the U.S. and EU productivity studies were to an important extent created by the different methods of calculating computer prices, and by the simple fact that the U.S. happened to have a bigger ICT sector than most of the EU countries. They also noted that the data did not point in the direction of significant positive spillover effects of ICT investment to the rest of the economy in the Euro area, since the overall productivity growth apparently had slowed down in the second half of the 1990s. Although some researchers claimed that the Solow paradox had been solved, Gordon, for example, argued that:
"These results imply that computer investment has had a near–zero rate of return outside of durable manufacturing. This is surprising, because 76.6 percent of all computers are used in the industries of wholesale and retail trade, finance, insurance, real estate, and other services, while just 11.9 percent of computers are used in five computer–intensive industries within manufacturing, and only 11.5 percent in the rest of the economy ... Thus, three–quarters of all computer investment has been in industries with no perceptible trend increase in productivity. In this sense the Solow computer paradox survives intact for most of the economy ... ." [15]Gordon’s argument was based on his analysis of long–term productivity trends and cyclical productivity effects. Although the relevance of cyclical factors was widely accepted, no clear consensus existed about their importance. Gordon (2000:65) [16], for example, maintained that it was possible that computers had shown their biggest productivity impact already before the 1990s.
Researchers also noted that industry–level studies do not show any consistent influence of ICTs. O’Mahony and Vecchi (2002), for example, found that standard econometric approaches show negative impacts of ICT on output and productivity growth, arguing that this was caused by aggregating industries where ICTs have different effects. They also noted that most econometric studies on ICTs implicitly have assumed that the impacts of ICTs have remained the same during the last decades.
In general, researchers now agree that several important conceptual and empirical issues still require clarification. On a closer look, the apparent consensus about productivity trends seemed to be built of research footnotes that explained methodological and data limitations. The bust of the Internet boom also questioned the sustainability of the productivity patterns seen in the end of the 1990s. If there was a consensus, it started to fall apart. The U.S. Congressional Budget Office stated in its report on ICT productivity impacts that "In contrast to the unanimity about the effects of computer hardware manufacturing, no consensus exists yet on the degree to which computer use has boosted total factor productivity growth" [17]. As Mahadevan [18] put it: "Expert opinion is solidly divided on the IT–productivity debate. One view is that the IT–productivity paradox exists, and the other that there is no such paradox."
So what, exactly, was the paradox? It centered on total factor productivity growth, which economic literature associates with "technical advance" and the overall efficiency of economy. The long–term trend in total factor productivity can be interpreted as the change in the economy’s underlying productive capability when the inputs remain the same [19]. This somewhat complex sentence describes the reason why economists call the productivity paradox a paradox. Despite astonishing advances in technology, including ICT, total factor productivity grew slower after about 1974 than during the previous decades. This was in stark contrast with Solow’s classic calculations in the 1950s, which showed that in the last century most of the economic growth had been generated by increases in total factor productivity. After about 1974, technological advances, however, seemed to lose their economic importance. This paradox is visible in Figure 1, which shows this surprising disappearance of total factor residual in productivity growth.
Figure 1: Sources of labor productivity growth in the U.S., 1960–1994. [20]
No one has been able to explain why exactly total factor productivity growth slowed down after 1973. Two decades later the problem, however, seemed to go away. Between 1995 and 1999, labor productivity grew in the U.S. at nearly double the average pace of the preceding 25 years. According to neoclassical productivity studies, the main sources for productivity growth were increased investments in ICT and unexplained improvements associated with technical progress and ICT use. Productivity growth was also strongly focused on ICT producing industries and perhaps also on industries that used ICT extensively. Although it still remained unclear why the productivity paradox originally had appeared, it seemed increasingly clear that information technology made it disappear. The question then remained whether this was a temporary event, or something that we should take into account in economic forecasting and policy–making.
It is impossible to understand discussions about the productivity impacts of ICT without a clarification of the basic concepts that are used in this discussion. Economists use productivity–related concepts in very specific ways that often create confusion among non–specialists. In particular, when we try to assess the current state of knowledge about the impacts of ICT, it is necessary to understand what economists actually mean by terms such as production efficiency, total factor productivity, real output, and technical change. By clarifying these terms, we can try to see how far the traditional approaches can take us, and where we need to move beyond the traditional interpretations of these terms.
Productivity and technical change in the neoclassical theory
There are many alternative ways to measure economic efficiency [21]. Labor productivity is a "single–factor" productivity measure, which can be measured either from total output or value–added. When labor productivity is measured as value–added per hour worked, or per person employed, productivity becomes closely related to the average per capita income [22]. This is the main reason why labor productivity is understood to be a central concept in economics.
Labor productivity can also be measured from gross output. When measured from gross output, labor productivity simply gives the ratio of output to labor inputs.
In general, changes in labor productivity reflect many different types of changes in the production process. For example, when more capital is used in the production process, the relative amount of labor in the total production inputs decreases and the measured labor productivity increases. In addition to such "capital deepening," labor productivity can also increase when existing production equipment is used more efficiently, for example, during economic growth periods, when the utilization of production capacity typically increases. It is therefore important to note that labor productivity, despite its name, does not measure workers’ efficiency or labor–related changes in production. As a single–factor productivity measure, it measures all changes in output using labor inputs as a reference point [23].
It is also useful to note that although macroeconomic labor productivity is often understood to reflect task productivity, today we do not know how they are related [24]. The concept of labor productivity is perfectly agnostic about the causes of output growth. In recent years, an important source of labor productivity improvements in the ICT using sectors, for example, has been the increased importance of self–service. As more work is allocated to customers, whose labor is not compensated, the measured labor productivity increases [25].
As noted, productivity can also be measured using value added instead of gross output. Value–added productivity indicators are common in studies that try to isolate sources of productivity within and across specific industries or firms. Instead of measuring total gross output, they subtract intermediate inputs and measure as output only the incremental production that occurs within the measured process. For example, total sales of computer manufacturers do not necessarily tell much about industry output or labor productivity. A better measure of produced output is the value added in the process [26]. Subtraction of intermediate inputs is necessary, for example, when productivity differences between specific industries are studied. This is typically the case in research on sectoral impacts of ICT. Productivity changes in the computer assembly and retail industries, for example, need to be distinguished from productivity changes in the semiconductor and software industry, if we want to understand the real sources of productivity growth.
Most macroeconomic analyses of the productivity impact of ICTs are based on the neoclassical growth accounting framework originally proposed by Solow in 1957. Solow’s calculations indicated that over 80 percent of labor productivity growth remained unexplained by increases in labor and capital inputs [27]. The difference between the observed growth rate and the theoretical growth rate generated by increases in labor and capital inputs became subsequently known as the "Solow residual." The residual reflects economic growth that is left unexplained after increases in labor and capital inputs are taken into account. As was noted above, it has commonly been interpreted as technical progress.
The Solow residual is closely related to the concept of total factor productivity (TFP). In Solow’s original formulation, output growth is completely determined by changes in labor and capital inputs, and the change in total factor productivity. In Solow’s framework, economic output is represented as:
Total output = TFP * F(L,K);where the "production function" F describes how labor, L, and capital, K, are converted into total economic output, and a "total factor productivity" multiplier describes changes in the overall efficiency of the conversion process.
The concept of multifactor productivity (MFP) emerges when labor and capital are separated into many qualitatively different types and these different types of labor and capital are explicitly added to the production function. In practice, TFP and MFP are used synonymously [28].
Solow’s residual appears when one studies the growth rates of inputs and outputs. Without changes in total factor productivity, increased labor and capital inputs directly determine the growth of total output. Under the assumptions of full competition, perfect and efficient allocation of resources, constant returns to the scale of production, and independence of total factor productivity of the relative composition of capital and labor, the rate of change in total factor productivity multiplier becomes the Solow residual. As the list of conditions indicates, the link between TFP growth and the Solow residual is not trivial. It is, however, almost trivial in the neoclassical framework that typically takes these conditions as starting points [29].
In contrast to labor productivity, total factor productivity is a combined measure of production efficiency. The importance of total factor productivity in ICT productivity studies is to an important extent related to the fact that it is associated with technical progress [30]. In fact, economic literature often states that total factor productivity represents the current level of technology. This is a common source of misinterpretations. Total factor productivity does not measure technology or technical progress; instead it measures all those factors that are not explicitly taken into account when we describe the way in which the economy turns inputs into outputs.
In other words, the growth rate of total factor productivity, or the Solow residual, incorporates all those elements of output change that do not result directly from increased capital or labor inputs. The residual collects productivity effects that are not modeled, as well as those that are mismeasured or modeled incorrectly. Strictly speaking, therefore, the Solow residual is a measure of our ignorance [31]. Much of the research on productivity has in fact tried to improve economic growth models to get rid of the residual. Economists know this well, but they still frequently make the claim that the residual reflects the "current level of technology." In fact, this usage is so common in the research literature that it has become the definition of the economic concept of "technical advance." As this confusing terminology has often led to misunderstandings in the interpretation of ICT productivity studies, it is important to note that there is no empirical or conceptual link between technological change and the residual. The residual will show productivity effects that are unrelated to technology, such as earthquakes, global warming, international trade agreements, new diseases, the aging of population, sunspot activity, forest fires, and wars [32]. If we remember that "technical advance" does not have anything to do with technical advances as they are understood by scientists or engineers, the confusion goes away.
The importance of the Solow residual is in the fact that it filters out the growth impacts of pure labor and capital increases. Savings that are profitably reinvested in production by hiring workers or through investing in capital equipment and materials lead to growth in production. The Solow residual accounts for growth that is not explained by such increases in inputs. New tools and production capital obviously embed technology and when workers are given more capital in the form of better tools, production technology, and plants, such capital deepening increases labor productivity. If it is fully paid for, it, however, does not change total factor productivity. The "current level of technology" that is captured in the Solow residual represents "costless technological progress" or, more accurately, "costless improvements in productive efficiency," and excludes technological progress that is embedded in capital investments. The Solow residual therefore is, for neoclassical economists, a "free lunch" [33].
In fact, productivity researchers often point out that the "technological progress" represented by the Solow residual does not imply technological progress. It can simply mean more efficient ways to organize work, better institutional structures, or good weather. Economic historians have also frequently pointed out that the reverse is true as well: technological advances do not imply changes in productivity [34]. New technologies that become important in many domains of economic activity often require lengthy adoption processes and institutional change before they become visible in the macroeconomic measurements [35].
Solow originally distinguished only two different types of inputs, labor and capital. In Solow’s model, capital included only fixed tangible capital. Denison (1962) and Griliches (cf. Griliches, 2000) extended this model to include several different types of labor, and Jorgenson (1963) and Jorgenson and Griliches (1967) further to different types of capital [36]. Such additions try to accommodate the fact that workers have different levels of education and skills, and that different types of capital may have very different productivity characteristics over time. Researchers therefore now commonly talk about multifactor productivity instead of total factor productivity.
More recent productivity studies have also used models that view technical progress as an inherent part of economic growth, and not just an externally given "current level of technology." Such endogenous growth models have, for example, emphasized research and development spillovers, network effects, and changes in the technical quality of products. As more sophisticated models are developed and more parameters are used to describe the economic translation from inputs to outputs, in theory the unexplained component of economic growth shrinks. A perfect description of the ways in which knowledge, organizational and institutional factors, and improved work tools and methods influence work outputs would therefore imply that the Solow residual disappears [37]. On the other hand, in the 1970s it already did disappear in many developed countries, and this, exactly, was the source of the productivity paradox. As Griliches [38] noted: "First we wanted to get rid of the residual, now we want it back!"
It is important to realize that most influential studies on ICT productivity impacts are based on neoclassical models [39]. They typically start from the assumption that economic actors allocate their resources optimally, price their products in fully competitive markets, and maximize their profits in an economic equilibrium. This theoretical set-up has a somewhat paradoxical consequence: by definition, productivity cannot be increased by improved use of current technologies. If the economic actors really are perfectly rational economic actors, and if the economy really is in the state of allocative equilibrium, as the conventional mathematical treatment in this framework requires, the actors cannot become more productive. The drivers of efficiency improvement cannot be economical in this framework; instead, they require magic acts that kick the economic system from outside, keeping it in the growth path. Neither can the neoclassical framework describe the productivity impact of qualitatively new products and technologies (Hulten, 2000; Pakes, 2002). As Schreyer [40] notes: "equilibrium concepts may be the wrong tools to approach the measurement of productivity change, because if there truly was equilibrium, there would be no incentive to search, research and to innovate, and there would be no productivity growth."
This is an interesting challenge if indeed innovation is important for growth in the Knowledge Society. Although productivity studies often start from the neoclassical assumptions, innovation researchers commonly adopt the Schumpeterian approach to economics of innovation, which, in contrast, starts from the idea that innovators and entrepreneurs systematically create disequilibrium and extraordinary profits through innovation. So, although researchers working in the neoclassical framework have made important contributions in the economics of innovation, the applied theoretical framework is conceptually limited in its capacity to describe innovative processes and technical change.
Some productivity studies, therefore, adopt a radically different perspective. Whereas the traditional neoclassical studies assumed that all firms operate with optimal efficiency, defined by a "current level of technology," some recent studies conceptually divide production efficiency into two components. First, the unobservable theoretical "production frontier" gives the maximum output that could be achieved using the available inputs. Actual firms rarely operate at this theoretically optimal production frontier. This allows for, for example, bad management decisions and work practices, as well as institutional and historical constraints. In actual organizations, production efficiency is usually only a fraction of the theoretically possible maximum [41]. This fraction characterizes the "technical efficiency" of the firm in question [42]. Productivity change, therefore, has two components: one related to a "current maximal level of productivity" and another related to the "productive efficiency" of the firms (Färe, et al., 1994; Mahadevan, 2002).
In this framework, the theoretically best current productivity can be approximately found by studying the most efficient producers. For example, if two firms have exactly the same inputs, and one produces twice as much as the other, the productivity of the worse producer is half of the more efficient one. By studying the existing inputs and outputs of real producers, one can find those input combinations that would lead to maximal outputs. The maximal outputs can then be used to define the best current theoretically possible efficiency of production. The actual technical efficiency of real firms can then be defined as their distance from that maximal level of production. As the maximal level of production is also known as the production frontier, this productivity analysis framework is therefore often called "frontier analysis" [43].
In the traditional neoclassical "production function" analysis, changes in TFP result from "costless progress" that was associated with technical progress or advances in knowledge. In the production "frontier analysis" framework, productivity changes also result from more efficient use of existing technologies and knowledge. As advances in technology can also lead to short or long–term losses of technical efficiency, technical progress can also decrease TFP. Furthermore, in this framework, technical efficiency can depend on organizational processes and the quality of management. Managers can, for example, over–invest in new technologies [44].
Growth contributions from ICTs
In the neoclassical framework, ICTs influence labor productivity in three different ways. Other things being equal, more ICT production means more total output. When ICT manufacturers learn to produce more powerful computers without increasing their economic inputs, this learning becomes recorded as total factor productivity growth in the ICT producing sector. Second, for industries that use ICTs as end users, they are capital investments. The increasing use of ICT capital in the user industries implies capital deepening and labor productivity growth. Third, if the user industries become more efficient because of their new ICT investments, this can also lead to increase in productivity, for example, by accelerating knowledge creation and by decreasing coordination and transaction costs.
The importance of these effects depends on the importance of ICTs in the economy. The overall effect may be small, if ICT production and investments represent only a small fraction of the total economy. In fact, this has been the case until recently. In the EU, ICT manufacturing and services are less than six percent of GDP, Ireland and Finland being the most ICT manufacturing intensive countries [45]. The value added in ICT manufacturing and services reached about 8.6 percent of the Finnish GDP in the year 2000, whereas it was about 8.1 percent in the U.S. [46] The role of ICT investments, however, has been growing during the last decades, and as a result ICTs now start to have a considerable impact on economic growth. In OECD countries, the contribution of ICT capital to GDP growth increased from about 16 percent to about 20 percent from the first to the second half of the 1990s [47]. The impact has been pronounced in the U.S., where the level of ICT investments has been relatively high already for several decades.
Although the visible effects of ICT use have been small until recently at the level of aggregate national economies, ICT production, however, has been very important for total factor productivity growth already since the early 1980s. This is because technical developments in computer technologies have been extremely rapid. Technical improvements have doubled capacities of computer memory and disk storage and halved microprocessor feature sizes roughly every two or three years during the last three decades [48]. When such improvements have not led to price increases, they are recorded as increases in total factor productivity. Although the total volume of ICT production has been a relatively small part of total economy in most countries, its extremely rapid growth rates have made it important for TFP growth. According to the U.S. Department of Commerce, ICT manufacturing and services grew by an average 22 percent per year, and was responsible for an average 29 percent of the country’s overall real economic growth during the 1996–99 period [49]. In fact, Oliner and Sichel [50] estimate that the in the 1974–95 period, ICT production contributed about half of the TFP growth in the U.S.
In practice, productivity calculations make a simplified assumption that ICT production consists of computer manufacturing, software production, and telecommunication equipment manufacturing. Often the studies have focused on computer manufacturing, arguing that telecommunication equipment prices and software are not well measured in the current statistics. Most of what we know about economy–wide ICT productivity impacts comes from such studies. It is therefore useful to see how such a simplified analysis can be conducted. In the next section, I will use the widely quoted study by Oliner and Sichel as an example to discuss the basic logic of dissecting economic growth into its various elements, including ICT. This discussion allows us to evaluate the robustness of those theoretical and empirical assumptions that underlie our current knowledge on ICT productivity impacts, and point out important areas where we need further analysis and research.
The logic of growth contribution calculations
Oliner and Sichel (2000) calculate the contributions of ICT to output growth using five inputs: computer hardware, computer software, communication equipment, other capital, and labor hours. In addition, they adjust labor inputs for labor quality changes. In the neoclassical growth accounting framework, this leads to a simple equation that describes the total output growth rate as a weighted sum of input growth rates.
The question about appropriate weights is a central question in the neoclassical framework. If we only had one type of input, the output growth rate would equal the growth rate of the input, plus the total factor productivity Solow residual. In the case of several inputs, the weights are called output elasticities. They are multipliers that describe how much the output grows for an incremental increase of the specific type of input.
Output elasticities cannot be directly measured, as we cannot freely experiment with the inputs and outputs of the economy. Solow, however, pointed out in his landmark article in 1957 that the neoclassical assumptions can be used to estimate the elasticities. In fact, he showed that output elasticities equal the income shares earned by each type of input. This is because in the neoclassical market equilibrium, each input is paid its marginal productivity. Economic actors invest in different types of capital and labor until the last marginal investment cannot increase profit. In this equilibrium, adding, for example, one hour of work should increase the total output by the cost of one work hour. Output elasticity of labor, therefore, equals the cost that has to be paid to cover the cost of the added hour of labor. Similarly, the output elasticities of different types of capital equal the "rents" that should be paid for the marginal investment. These rents are also called "the user cost of capital" [51].
To calculate the total income of a specific type of input, we need to estimate the rate of return and amount of capital that generates the input. Firms do not always rent their capital from the market, but instead invest in capital and receive income from their investments over a period of time. Capital rents therefore cannot be directly observed, and we have to estimate them.
Capital stocks that are used in productivity calculations have to reflect economic services generated by the investments. Such "productive capital stocks" do not equal the market value of accumulated capital. For example, although an old computer may have very little market value, it may still generate economic services that reflect a large productive value. On the other hand, old computers may be unable to run new programs and therefore their value may depreciate even when they do not deteriorate much physically. Productive stocks are therefore calculated by correcting investment costs for depreciation and price changes. Such calculations require quite sophisticated modelling of time–dependent characteristics of productive value for each type of capital and detailed price indices. The results, however, are readily available, at least for the U.S. economy. In other countries, researchers typically use the U.S. price and volume indices as the starting point.
Contributions of the different types of investments to total economic growth can be calculated by multiplying the growth rate of capital by its weight, which equals to its income share, as was noted above. When the neoclassical assumptions are valid, income shares can be estimated using the cost of capital. According to the neoclassical assumptions, in the equilibrium each type of capital has to earn the same net rate of return. If this were not the case, the investments could be reallocated for increased profit, and by definition this cannot happen in the equilibrium. To generate the same net rate of return, the different types of capital, however, have to generate very different amounts of gross return. The gross return has to cover basically three different factors. One is the loss of productive value due to wear, tear, and obsolescence. The second factor accounts for the loss or gain that results from the change of the price of the asset across time. The third factor is the net rate of return that the capital would earn if it would generate returns on the market, perhaps with adjustments for possible taxes. Computers, in particular, have a relatively short investment lifetime and their prices drop rapidly, so that the gross rate of return has to be high. Oliner and Sichel estimate that computer investments depreciate roughly about 30 percent and that computer prices drop about 30 percent per year. This means that computer investments must earn about 60 percent above the net rate of return for investments in the economy, which Oliner and Sichel assume to be four percent.
After we have estimates for productive capital stocks for the various years and corresponding estimates for the income, the share of income for each type of capital stock can be calculated by dividing the income earned by the total income of the economy. Oliner and Sichel further assume that the income share of labor is what remains after the incomes earned by different capital stocks are subtracted from the total income. This, then, allows them to say how much each type of input has contributed to the growth rate of the economy. The rest is the total factor productivity residual. The resulting numbers are shown in Table 1.
Table 1: Contributions to Growth of U.S. Non–farm Business Output, 1974–99 (Oliner and Sichel, 2000, Table 1).
1974–90 1991–95 1996–99 Growth rate of output 3.06 2.75 4.82 Contributions from: 2. Information technology capital 0.49 0.57 1.10 3. Hardware 0.27 0.25 0.63 4. Software 0.11 0.25 0.32 5. Communication
equipment0.11 0.07 0.15 6. Other capital 0.86 0.44 0.75 7. Labor hours 1.16 0.82 1.50 8. Labor quality 0.22 0.44 0.31 9. Multifactor productivity 0.33 0.48 1.16
From Table 1, one can easily see that ICT was a major factor in the extremely rapid growth of the U.S. economy in the second half of the 1990s. An even more important factor was growth of labor inputs. The decline in the importance of labor quality indicates that the labor markets grew by employing also people who were not categorized as highly skilled labor. To avoid misinterpretations, one should however note that, strictly speaking, labor quality does not measure the level of skills in productivity studies. Instead, the labor force is categorized using a combination of characteristics that cluster workers in groups that have similar levels of labor costs. Labor quality therefore varies with, for example, the average level of formal education, industry of employment, age, and gender. "Labor quality" in productivity studies can often be interpreted as "wage category." The visible decline in labor quality contribution in the 1996–99, therefore, can be also interpreted as a "trickle–down" in the labor market. More generally, decreasing labor productivity can sometimes indicate that the members of the society broadly benefit from economic growth. In the neoclassical framework, wage categories, of course, were assumed to perfectly reflect the marginal productivities of different worker categories, so that labor costs and labor quality are more or less synonymous [52].
As was noted above, the growth impact of ICTs may also become visible in the total factor productivity residual, which Oliner and Sichel call multifactor productivity. In the above table, the growth contribution of ICT capital reflects the increasing investments in ICT capital by user industries. To analyze the impact of ICT production, we have to dissect the total factor productivity residual into components that originate from ICT manufacturing industries and from other sources. To achieve this, Oliner and Sichel decompose the economy into three segments: semiconductor manufacturing, computer manufacturing, and the rest. They further assume that the total factor productivity growth rate of the total economy, shown in Table 1, is a weighted sum of TFP growth rates of these three sectors. For the weights, they use the gross output share of each sector.
To implement the decomposition, Oliner and Sichel need estimates of the sectoral total factor productivity growth rates. Here they use a common and theoretically interesting assumption. This is that price decreases in computer and semiconductor manufacturing result from improvements in total factor productivity. The basic assumption is that if the semiconductor industry, for example, is able to continuously cut prices without losing its profitability, and if the input prices for the industry do not decrease, the source of the output price decline has to be better productivity. Output price decline, therefore, becomes a simple measure for total factor productivity increases in the industry. For researchers, this has the major benefit that output prices can be observed directly.
Using this procedure, Oliner and Sichel find out that computer and semiconductor manufacturing, indeed, have been major sources of total factor productivity growth. In the U.S., in the 1974–90 and 1991–95 periods they accounted for about half of the growth and in 1996–99 about two–fifths.
Empirical and conceptual limits of the Oliner–Sichel study
The Oliner and Sichel study is an exemplary academic piece of work, and it has been extremely influential in recent productivity discussions. It is, however, also clear that it leaves several interesting questions open. In the above described procedure, the neoclassical assumptions and accurate measurement of inputs and outputs are of fundamental importance. In this section I discuss points that have particular relevance in the context of the Oliner–Sichel study, and for studies that have adopted similar approaches for ICT productivity analysis.
The first point is that the study assumes constant returns to the scale of production. This is a common assumption in productivity studies, but it is also a quite unintuitive assumption for semiconductor and software production. The rapid declines in semiconductor prices result, to a large extent, from the fact that there are large scale benefits in semiconductor manufacturing. In fact, the most important semiconductor industry products, such as memory chips and microprocessors, are quite similar to packaged software, where additional copies can be printed with low cost. The assumption of constant returns to scale might be particularly misleading in studies on ICT productivity impacts.
Oliner and Sichel use price and capital stock estimates provided by the U.S. Bureau of Economic Advisers (BEA) and Bureau of Labor Statistics (BLS). Although these estimates are based on extensive studies, they are also known to have conceptual and empirical problems. One important conceptual problem is that the underlying price estimation models are not able to account for qualitative changes in the use of computers. Although the price indices that provide the basis for calculating ICT investments and capital stocks are commonly thought to model quality change, in reality they are only able to measure improvement in existing qualities [53]. Oliner and Sichel therefore implicitly assume, for example, that personal computers from different decades can be categorized as members of the same class. Empirically, however, it looks probable that the use of broadband–connected PCs with advanced audio and video capabilities is essentially different from the use of 1980s PCs, for example. The price indices that underlie the Oliner and Sichel study have theoretical difficulties in crossing such qualitative discontinuities. It is therefore possible that the existing capital stocks and the user costs of ICT are over– or underestimated. As the neoclassical framework does not say anything about discontinuous qualitative change, we actually cannot say much about the size of the resulting estimation errors in general. One could, however, speculate that when the importance of innovative activities grows in the economy, these discontinuities become increasingly important. Although they could have been relatively invisible when production was largely based on mass–produced goods, they perhaps need to be taken into account in knowledge– and innovation–based economies.
An important and well–known empirical problem in the ICT price indices is that there is no accurate information about software output. Average labor costs are therefore used to derive the price index for in–house developed software [54]. The price index for custom–developed software is then derived by averaging these in–house development cost estimates with estimates for prepackaged software [55]. Pre–packaged software indices, in turn, are based on spreadsheet and word processing software prices. A conceptual problem with this approach is that it measures output using labor input as a proxy for much of the produced software output [56]. With this assumption there can be no changes in computer programmer and systems analyst productivity [57]. Empirically, this is a troublesome assumption, for example, as it is often noted that professional software programmers can have very big task productivity differences, on the order of 1 to 20.
In their analysis of the growth contributions of ICT production, Oliner and Sichel use the assumption that price declines reflect productivity improvements. This assumption is called the "dual approach" for measuring total factor productivity growth (cf. Jorgenson and Stiroh, 2000; Barro, 1998) [58]. Instead of trying to measure TFP changes directly, this approach simply uses prices, and assigns all price change to TFP improvements. Implicitly, this procedure relies on a model of economic activity that does not necessarily closely reflect the realities in the semiconductor, software, and computer industries.
The underlying idea of the dual approach is that declines in relative ICT prices reflect productivity growth in the ICT sector. The semiconductor industry, for example, has continuously been able to produce more outputs with the same inputs. Without productivity increases that compensate the price declines, the industry would have gone bankrupt. This intuition is supported by the common sense observation that the semiconductor industry has been amazingly successful in inventing new product generations that have almost exponentially increased computing capability without increasing prices.
The public awareness of dizzying technical progress in ICTs does not, however, immediately translate to economic progress. Strictly speaking, the price developments in computers and semiconductors do not show much decrease. If we simply look the product prices and use the dual approach of measuring TFP growth using price declines, the result is that there has not been much TFP growth in the computer or semiconductor industries.
In nominal terms, the prices have remained relatively stable for new ICT products. Hard disk drives show a constant decline per sold unit [59], but until recent years, microprocessor prices have been relatively constant at introduction [60]. The median price of desktop computers sold in the U.S. has been about US$2000 since the 1970s, although recently the nominal prices have declined [61]. As there obviously have been amazing technical developments in the semiconductor and computer industries, price indices are therefore used to correct the nominal prices. This has led to rapid decrease in the estimated "real prices" of ICT products. A large fraction of the growth that is measured in the national accounts and related to ICT originates from these corrections. This is particularly visible in the U.S., where prices and investment stocks are extensively adjusted for quality. It is therefore important to understand whether ICT prices actually have been measured right. Indeed, this is a critical element in the ICT productivity story and therefore deserves a closer look.
Price indices as a source of growth
Simple price indices can be used for homogenous products, such as wheat and sugar. Many products, however, change across years. Price researchers therefore try to develop price indices for carefully controlled homogenous product groups, which can then be aggregated to generate, for example, consumer price indices. A common approach is to use "matched model" price indices, where the price for the exactly same or closely similar product is compared across time.
Matched model indices have been used for computer products, but they have important deficiencies. In practice, new computer models and versions emerge frequently, and completely new functionality is regularly introduced in ICT products. Only few product models exist on the market long enough that matched model price indices could reliably be developed for them. As the matched model approach cannot count new products for which historical data does not exist, it also weights old products whose prices may develop differently from the current ones.
The hedonic method tries to overcome these problems by creating statistical estimates of the value of product characteristics. For example, price researchers can observe the prices of ten different PCs that are otherwise similar except that they have hard disks that are of different capacity. Using the prices and the data on hard disk capacity, the researchers can fit a statistical model that estimates the current market value of hard disk capacity. By extending this approach to multiple technical characteristics — for example, processor clock speed, amount of random access memory, and other similar characteristics — the researchers can develop a mathematical model that describes how the different technical parameters influence the price.
If we plug in a specific product specification to the hedonic equation, it tells us how much the product would cost. Hedonic equations, therefore, can also give prices for products that actually do not exist on the market. Specifically, they can tell us what would have been the price of a product model if exactly the same model would have been on the market already a year ago. The price difference can then simply be used to derive a price index for that specific product. More generally, the same method can be used to derive price indices for products that consist of "bundles" of constant technical characteristics. The estimated price change of the constant bundle of product characteristics, therefore, gives a "quality–adjusted" or "constant–quality" estimate of price change.
Computers have been important for measured growth because computer prices have been aggressively adjusted for quality improvements. In other ICT products and services the adjustments have been much less prominent. This can be seen in Figure 2, which shows the U.S. price indices for computers, communications, software, and other products using the year 1996 as the base year. These indices are commonly used as the starting point in international ICT productivity studies. To put it very simply, the reason for the rapid productivity growth in the second half of the 1990s is the rapid decline in computer price indices, shown in the picture. In neoclassical productivity studies, this decline becomes extremely influential. This is because it affects both the growth rate of quality–adjusted productive assets and the user costs that multiply the growth of these assets. Further, as the user cost is calculated by multiplying the total quality–adjusted volume of productive assets with its gross rate of return, which itself depends on the rate of quality–adjusted depreciation and revaluation of the assets, the quality adjustments effectively have an impact that is in the third power.
Figure 2: Price indices used to adjust the productive value of different products.
The reason why productivity studies find ICTs as the driver of productivity improvements, therefore, is to be found in the fact that neoclassical growth accounting studies allocate productivity growth to those sectors where productive assets grow fast and where price declines are rapid. Productive computer assets, in turn, have been growing fast because the quality–adjusted prices that are used to measure the discrepancy between the actual market value of computers and their assumed "productive value" have been declining rapidly. The contribution of ICTs to labor productivity has been large as the rapid increase of ICT assets — or more exactly computer assets — becomes in this theoretical framework translated into capital deepening. Similarly, as the framework associates total factor productivity growth rate with the speed of price declines, it should not be a surprise that those industries where price indices decrease rapidly become important for productivity growth. As such, these results follow purely from the mechanics of growth accounting and there is nothing that would differentiate ICTs in this framework from any other products that have similar price dynamics.
Without quality adjustments, the story would be quite different. First, the amount of computer capital would be only a fraction of what productivity studies now assume it to be. Second, the growth rate of computer capital would have been much slower. This can be seen in Figure 3. The value of U.S. computing assets has roughly doubled over the two decades since the 1980s, while growth in the 1990s was relatively modest. The estimated value of productive assets that generate computing services, however, grew extremely rapidly in the second half of the 1990s. This rapid growth, in fact, has been the main source of research results that show that ICTs became important for economic growth and productivity improvements in the 1990s.
Figure 3: Computer assets in the U.S. Market value vs. value used in productivity studies.
The fundamental question, then, is whether the hedonic price indices measure correctly the productive value of computers. If productivity increase is conceptually independent of improvements in technical parameters, there should be no obvious reason why hedonic indices would correctly estimate productive computer assets, or that the "dual approach" would, in fact, measure total factor productivity change.
The hedonic price indices do not directly measure productivity. Instead, they measure a set of technical characteristics that empirically seem to be correlated with price, and which conceptually are associated with the production of valuable outputs. For microprocessors, for example, price indices are constructed by characterizing different microprocessor chips using their clock speed, internal bus bandwidth, existence of multimedia capabilities, and other similar characteristics. An important factor for semiconductors is also the time that has passed from the introduction of the chip [62]. Using the hedonic approach, the semiconductor industry output, therefore, becomes represented as a bundle of technically defined characteristics, such as the total volume of megahertzes and, for example, megabits that the industry has produced in a year.
To understand this question, it is necessary to understand what actually drives technical improvement in, for example, semiconductors. If the prices drop because of productivity improvement, price indices could be used as proxies for productivity. If, on the other hand, prices drop for reasons that are essentially unrelated to productivity improvements, the dual approach does not work.
It is also critical that the technical characteristics are productive and valuable for the buyers. For example, if semiconductor prices drop because old chips are being substituted by new ones, this does not necessarily lead to real output growth. From the productivity point of view, new chips reflect growth only to the extent that they lead to more valuable products.
The neoclassical theory assumes that market prices are determined by the productive value of products. In this framework, if producers pay for the production of some characteristics, they have to be productive, and as the buyers are fully rational, they pay exactly the productive value of the characteristic. In other words, the theory requires that price changes, in fact, are directly associated with productivity changes.
The problem, however, is that the nominal prices of these products actually do not drop as much as we believe they should. The conventional way to deal with this problem is to assume that the improvements in technical parameters can be interpreted as growth of output. Although the market prices do not drop, the argument goes, the buyers get more for their money. Furthermore, this adjustment is made in a way that exactly explains the observed price changes by imputing a component of quality change that makes neoclassical economists happy.
In a simplified way, the hedonic approach could, for example, count the number of transistors on a chip, and when more transistors are shipped, this would by definition mean more output. The fact that this output growth would not be captured by nominal sales would then be corrected using price indices that lead to "real" growth numbers that multiply nominal dollars until the output seems to measure the number of transistors shipped, instead of the chips that contain these transistors. The difference between sold chips and sold transistors would then be measured as an increase in total factor productivity.
Of course, transistors are rarely sold separately these days. Most transistors, indeed, are on semiconductor chips that use transistors in multiple different ways, for example, for building memory cells and microprocessor logic. Improvements in technical functionality depend on technical characteristics, such as the size and speed of transistors, internal data bus bandwidths, but also on skillful organization of the transistors on the chips and innovative designs. In fact, the processing power of a microprocessor depends to a large extent on the software that uses the underlying hardware capabilities. The compiler — that converts high–level programs into code that actually runs on the microprocessor and uses its transistors — is often the main source of processing power. Even a relatively technically unsophisticated microprocessor can easily beat state–of–the–art chips if the latter use compilers that are not skillfully optimized for the underlying hardware capabilities.
"Task level productivity" or "processing power" of a microprocessor, therefore, cannot be captured simply by looking at the technical characteristics of the chip. The technical characteristics make sense only in a context that combines hardware capabilities, software architectures, and specific types of applications [63]. As long as this configuration remains stable, we may be able to forget the operating environment of the microprocessor and focus simply on the microprocessor itself. Empirically, the configurations, however, change often. As was noted, in such environments, price indices become theoretically difficult to define.
In general, one should therefore ask in what sense total factor productivity increase could be associated with technology improvements. In the neoclassical productivity framework, total factor productivity was associated with the overall efficiency of the economic process. As was noted above, although total factor productivity often is interpreted as "the level of technology," economists are well aware of the fact that it is not in any direct way associated with technology. The dual approach, however, makes the link between "level of technology" and total factor productivity in a different way. It becomes a definitional link. The efficiency increases when technological change in products becomes interpreted as growth in real output.
Hedonic price indices often lead to confusing interpretations because they mix two essentially different concepts of technology. In the economic framework where these indices are used, "technical advance" is purely an economic parameter that is independent of any technical considerations. Hedonic indices, on the other hand, rely on an independent concept of technical advance that is deeply rooted in knowledge of engineering processes and the uses of technical products. The information that is needed to derive hedonic indices is not modeled in economic theory because the basic historical assumption of economics was that it can be an autonomous domain of study, where technical, social, or, for example, ethical sources of values do not have to be considered. In this sense, hedonic indices import into economic theory information about values that the theory itself assumed to be irrelevant. Strictly speaking, consistent use of hedonic indices, therefore, would also require that we revise the economic theory of value.
Of course, if we want to say something about the relative impact of ICT industries, we also need to do similar output corrections in other industries as well. If we only adjust one industry and get high growth rates for it, our productivity studies pick up this specific industry and show that it has been an important source of growth. This, in fact, is what typically happens in ICT productivity studies. A broad and consistent adjustment for quality in economic outputs would make ICTs less prominent sources of growth.
A fundamental question is to what extent the technical improvements really reflect conventional ideas about output growth. As was noted, implicitly the price indices, and therefore also the accounted output and investment growth, assume that, for example, increased microprocessor clock speed implies economic growth. End users, however, are not necessarily interested in technical characteristics, per se. If they were, technical improvement could be expected to lead to increasing demand. Although the demand in semiconductors is very cyclical, it however, seems to have been surprisingly independent of technical improvements over the years (Gordon, 2000). The apparent demand growth exists to a large extent because the "real output" is corrected by the price indices. In other words, we get growth because the conceptual system that we use to measure growth puts it where we want to see it.
In a simplified way, the use of quality–adjusted price indices essentially implies that smaller transistors mean growth. Strictly speaking, quality–adjusted indices measure the decline of prices in the ICT products, but this becomes interpreted as increase in the real value of the new products that replace earlier products, or as efficiency improvement when old product generations are dumped on the market at rapidly declining prices. The semiconductor industry has continuously been able to ship memory chips that have smaller transistors and microprocessors that have higher clock speeds, but the jump from purely technical characteristics to economics requires a crossing of an interesting conceptual boundary. Why, indeed, we think that smaller transistors imply more economic output, but smaller cars, for example, do not? Why do we count the increasing millions of transistors on a chip, instead of counting the chips?
The reasons, of course, are complex, and price indices are derived using many technical characteristics, not only the size or number of transistors on a chip. It is important to note that the arguments for making specific quality adjustments, however, cannot in any conventional sense be economic arguments. They are arguments about the usefulness and value of technical characteristics in specific uses of technology. To be able to generate a list of potentially valuable technical and functional characteristics, we need to specify a particular way and context of using the product and ask engineers and business managers about the alterative ways to produce these functionalities within currently known constraints.
If the uses and the constraints for production are stable, these contextual factors may sometimes be taken for granted and they do not have to be explicitly described. Conceptually, the dual approach would seem to be best justified in industries that compete on product price in relatively perfect markets, where innovation does not matter. Assuming that firms would try to retain or grow their profits in such a setting, they would attempt to squeeze more outputs from a given input. This, indeed, could reasonably be called production efficiency improvement.
In the ICT industry, the setting is, however, quite different. Firms compete by introducing new innovative products and product variations. "Products," therefore, are only ambiguously defined [64]. The economic meaning of products is continuously reinvented by the users and as the uses evolve continuously, existing value systems become reconfigured. A large fraction of the total profits are usually generated in the first months after the introduction of a new product, when it has limited competition and when the product price can be high. As soon as competitors enter the new product category, prices start to decline extremely rapidly. For example, to cover the development costs, semiconductor manufacturers sell their products with very high margins at the beginning of the product lifecycle and try to keep the product price above the manufacturing cost by effective use of scale effects [65]. Intel’s microprocessor chips, for example, were typically introduced at prices between US$600 and US$1000 during the 1990s. When the chips were discontinued, their prices had usually fallen to under US$100. The manufacturing cost for the chips has typically been much lower. Aizcorbe [66] notes that for the Pentium I chips, which were introduced in the first quarter of 1994 for US$1000, the manufacturing cost was at the fourth quarter of the same year about US $53 per chip [67]. Productivity increase would probably be reflected on the decline of the manufacturing costs, but the link between product price and productivity is not clear when product price is fifteen times the manufacturing cost. All this indicates that innovation and qualitative change are central in this industry.
Neoclassical assumptions normally require that producers operate in markets where producers have no influence on prices, and where the production of one additional product costs as much or more than the first exemplar of the product. Oliner and Sichel, and many other influential studies, implicitly assume that all firms are perfect users of ICT, allocate their resources with perfect economic rationality, and that they realize without delay all possible productivity opportunities. Furthermore, due to difficulties in data collection, they start from the assumption that price changes perfectly reflect total factor productivity changes. Theoretically, this might be so, if the markets were perfect and if the competitive environment would be in equilibrium. It is, however, not clear whether such a competitive equilibrium would conceptually make sense, or be a contradiction in terms in industries that compete through innovation [68].
The link between economic growth and technical change is not a trivial one also because quality adjustments are necessarily based on a retrospective selection of quality parameters. For example, power consumption became important with the introduction of portable PCs, but it is not included in the microprocessor price index calculations in the U.S. In fact, Berndt, et al. (2000) show that the emergence of portable PCs created a discontinuity in price indices for PCs around 1987. The historical contingency of the relevance of technical characteristics means that potentially we should re–evaluate historical price indices whenever new product functionalities or uses emerge. Intel, for example, has recently moved to describe its microprocessors in terms of MIPS (millions of instructions per second) per Watt, reflecting increasing problems with power consumption [69]. Strictly speaking, the accumulated ICT capital stocks should now be computed anew, as we understand that important parameters, such as investments in chip cooling and computer room air conditioning, were not taken into account.
In general, it is not conceptually clear what we mean by constant prices when technological change becomes the main source of price changes. This makes the value of money dependent on qualitative advances in technical parameters, and the conventional theoretical economic frameworks move to a new uncertain terrain [70]. This terrain is particularly shaky in the domain of ICT, where measured price changes have been driven by extremely rapid improvements in semiconductor quality [71].
Indeed, if we do not correct for quality improvements in semiconductors, both overall growth and productivity numbers look quite different [72]. Much of the productivity increase goes away if we roll back the effect of quality improvements in investment stocks and consumption. For example, the global sales of semiconductors do not show any obvious trend after the beginning of the 1990s when measured in current dollars, as can be seen from Figure 4. In year 2003, the world semiconductor market was 166 billion U.S. dollars, or some 11 percent more than in year 1995, without adjusting for inflation. If we account for the fact that there are now more computers in the world than a decade ago, and that annual sales in the semiconductor industry therefore increasingly replace rapidly decaying computer investments, it seem possible that new semiconductor demand has slowed down [73]. Although there was a clear upsurge in semiconductor sales during the last years of the 1990s, this was perhaps a temporary peak. Of course, if we correct for quality improvements and treat replacement sales as new sales, the picture looks quite different. This, indeed, is what happens when national accounts and productivity studies convert the actual sales numbers into "real" investments. The point, however, is that the picture crucially depends on the accuracy of the quality adjustments. It is possible, for example, that the rapid increases in semiconductor quality result from the fact that demand for computing has saturated, and that competition has forced semiconductor firms to aggressively cut their prices to keep the total volume from shrinking [74]. Neoclassical productivity analysis cannot easily shed light on such issues, as it does not tell about the causal sources or drivers of productivity change.
Figure 4: World semiconductor sales, 1982–2002.
Data: World Semiconductor Trade Statistics (WSTS).Accurate estimates for the impact of quality corrections in price indices are difficult to make, as multiple potential sources of error should be considered simultaneously. Landefeld and Grimm (2000), for example, have argued that quality adjustments cannot to an important extent explain the rapid GDP growth in the US in the 1995–99 period. This result is based on their estimate of the size of the errors created by the quality adjustment methods [75]. As the selling prices of typical computers has declined about five–nine percent annually, and as the hedonic price indices used for computers have declined about 33 percent annually during this period, one could estimate that the difference created by quality adjustment could be perhaps 25 percent per year. If such a potential error would in fact be realized when computer assets are calculated in national accounts, the impact on GDP growth would be relatively small. This is because the share of computers of total assets is small. Landefeld and Grimm estimate that the use of hedonic quality adjustments could have introduced no more than a one quarter of a percentage point to the average annual 4.14 percent GDP growth over the 1995–99 period.
On the other hand, the main result of growth accounting studies was that the rapid growth in the second half of the 1990s was strongly associated with ICT production. Although quality adjustments may have little importance for overall growth, they have central importance for studies that focus on ICTs.
One way to see this is to compare ICT asset growth rates, measured in their historical–cost values, as they would be recorded in the company reports, with the growth rates of quality–adjusted productive assets that are used in productivity calculations [76]. Computers and peripheral equipment grew in the 1990–2000 period 5.3 times faster when measured in quality–adjusted quantities than when measured in their historical–cost values. For software and communications equipment, where quality adjustments play a smaller role, the growth in historical–cost values was almost exactly equal to the growth of quality adjusted productive assets. This can be seen by comparing Figure 5 and Figure 6.
As the figures show, in historical–cost values, computers and peripheral equipment assets grew three–fold in the 1980–90 period and they doubled in the 1990–2000 period. There was growth, as we could expect. Quality–adjusted quantity, however, grew 10.5 and 10.7 times during these periods. Software stocks increased about five– and three–fold in historical values, and communication equipment about three– and two–fold, respectively. In quality–adjusted values, software stocks increased about three–fold and communication equipment stocks about two–fold in both periods.
Figure 5: ICT assets in the U.S. historical–cost value, 1980–2001.
Data: Bureau of Economic Affairs (BEA).Simply looking the historical–cost values, one can see that software became more important than hardware in year 1991, but that communication equipment still represents about 1.6 times bigger assets in the U.S. than software. Software assets and investments have also grown faster than hardware since the early 1980s. This picture, however, turns upside down after quality adjustments. Computers and peripheral equipment assets grew three times faster than software in the 1990–2000 period, when measured using productive stocks. In historical–cost money, software assets, however, increased twice as fast as hardware.
Figure 6: ICT productive assets in the U.S., 1996 dollars, 1980–2001.
Data: Bureau of Labor Statistics (BLS).In current–cost value, which attempts to measure the replacement value of these stocks, computer and peripheral assets have been relatively stable. The replacement or market value of computers and peripheral equipment assets was close to US$100 billion for the 1987–98 period in the U.S., peaking at about US$150 billion in year 2000. The replacement value of software assets was US$345.5 billion, communication equipment US$499 billion, and computers and peripheral equipment US$138.6 billion in 2001.
Quantity indices try to measure the "physical–volume" or the "real" stocks of assets by taking into account quality change. The differences between quantity indices of assets and historical values of the same assets therefore reflect the impact of quality adjustments. Depreciation rates, however, also play a role. In ICT investments, the expected lifetime is short and depreciation rates are high. The average age of computer and software stocks was less than two years in the U.S. in year 2001 [77]. To a large extent this was because computers and software are assumed to depreciate extremely quickly. In this sense, the U.S. economy has perhaps wasted money by investing heavily in rapidly depreciating stocks that lose their value in just a couple of years. If all computers in the U.S. had been thrown to a garbage can in year 2000, the ongoing investment rate would have been sufficient to rebuild these assets in about two years. On the other hand, this movement towards rapidly decaying investments has generated growth opportunities, as more production is needed to compensate the increasing consumption of capital. This rapid decay of many ICT products makes them somewhat of a borderline case between capital and consumption goods. Indeed, some hard disk industry managers have complained that they are in the fish business, as product prices constantly drop about one percent per week and products on the shelf begin to stink [78].
For communication equipment, depreciation rates are estimated to be much smaller than computer and software depreciation rates. This is reflected in the fact that the average age of communication equipment investments was about five years in the U.S. in year 2001. The reasons for the slow decay of communications equipment are to some extent historical. Depreciation includes declines in value that results from wear, tear, accidental damage, obsolescence, but also from aging. If incumbent telecom operators would depreciate their assets at computer and software rates in their accounts, they probably in many cases would go bankrupt, as the accelerated depreciations would destroy their profits and assets.
Accurate price indices are critically important for productivity studies, and there has been extensive theoretical and empirical research on price indices in ICTs during the last two decades. As was noted above, it is, however, not clear what we measure with price indices. Theoretically correct price and volume indices have to be calculated by "chaining" changes from one time period to the next one. This implies that prices start to measure the value of the particular type of good, and the value becomes incompatible with the values of other goods. Chained indices can measure theoretically correctly year–to–year changes in the prices of goods that they measure, but they also become independent of price changes of other goods that they do not measure. "Microprocessor money," "hard disk memory money," and, for example, "car money" become different in this treatment. As a result, quality adjusted Euros or dollars cannot be added anymore in the traditional sense. Quality–adjusted values become particularly incompatible when the quality adjustments are large, and when we move away from the time period that acts as the base year used for defining the original values. In technologies such as computers, where the value of investments decays in just a few years because of the introduction of new innovative technologies, quality adjustment of price indices is therefore a theoretically interesting challenge.
Quality adjustments look natural in products such as computers or semiconductors, but conceptually we also have to jump from a world of physical characteristics to the world of economics when we start to use statistical regressions to derive hedonic prices for products. The fact that quality money is not additive, indicates that the basic concepts of economical theory are not valid anymore.
In sum, the reality of semiconductor industry does not easily accommodate the assumptions of the dual approach of identifying total factor productivity increases with price decreases. Nor does it fit well with the assumptions of the neoclassical growth accounting framework. Hedonic adjustments lead to quite profound questions about the ways in which value and capital investments are conceptualized in ICT productivity studies. As a result, the findings of the Oliner and Sichel study, which crucially depend on the dual approach, the neoclassical equilibrium assumptions, and quality–adjusted price indices, do not necessarily well reflect the productivity impacts of ICT. Faster processors or bigger MIPS ratings do not necessarily mean more productivity, at least in a world where processors may be idle over 90 percent of time.
Productivity paradox or paradigm anomaly?
When we analyze existing ICT productivity studies it becomes clear that the results of these studies depend both on broad conceptual assumptions and the details of data collection methods. As Kuhn (1970) and other historians of science have shown, existing research paradigms can always be extended by adding new explanatory variables. This process of continuous improvement expands the boundaries of current theories without any fixed limit. With a sufficient number of crystal spheres and good computers, the Ptolemaic model of the solar system is as accurate as the best Keplerian models. However, as Kuhn also noted, the approaching breakdown of a theoretical paradigm is often indicated by persistent anomalies that remain unexplained in the present theoretical framework. The Solow paradox, the mysterious disappearance of the total factor residual after 1973 and the invisible impact of ICTs, is a potentially interesting example of such an anomaly. If it truly is a paradox, the theoretical approaches that are used to discuss productivity may require rethinking. If, on the other hand, the paradox can sufficiently be explained within the current framework, no fundamental revision is needed. It is therefore interesting to see how productivity researchers associated with the neoclassical approach have explained the paradox.
Triplett (1999) has reviewed several explanations for the paradox. The first explanation is that ICT is only a small fraction of GDP and national accounts. Brynjolfsson (1993), for example, calculated that IT investments in the U.S. contributed perhaps no more that 0.06 percent to aggregate GDP in the 1980s [79]. In fact, it was only around 1982 when the annual investments in computers and peripheral equipment started to exceed investments in farm tractors in the U.S. [80] As was discussed above, if ICT production and investments are relatively invisible parts of the total economy, also their productivity impact should be quite invisible [81].
The second explanation is that the methods used to measure computer investments and costs of ICT perhaps mismeasure the value of these investments. As was pointed out above, if we measure ICT investments in nominal current dollars, they have increased only relatively slowly [82]. The picture changes radically when we adjust computer prices for quality improvements.
ICT production may also be difficult to measure, for example, because productivity studies that use national data may not be able to see all productivity impacts of international division of labor. Another example, which is more difficult to handle within the neoclassical framework, is open source software. Open source software now forms over half of the core software on the Internet, but it often has no transaction price [83]. It therefore remains invisible and unaccounted in national accounts and investment stocks.
The third possible source of the productivity paradox is that computers are often used in economic sectors were productivity is not easy to measure. Output in services, in general, is difficult to measure. As the service sector of the economy increases, the measured productivity growth may be slowing down simply because qualitative improvements in services are underestimated.
The fourth explanation is that ICT impact is poorly captured in economic statistics. Software firms have, for example, invested heavily in the usability of their products, and these qualitative improvements could be interpreted as increased consumption. Increased product variety may also imply more valuable products for their users, and variety of choice itself could be valuable. Such qualitative improvements are not necessarily captured in current output measures. Although systematic mismeasurement of quality improvements may cancel out in calculations of productivity growth rates, Brynjolfsson and Hitt [84], for example, have argued that computers are associated with an increasing degree of mismeasurement. According to Brynjolfsson and Hitt, this is likely to lead to increasing underestimates of productivity and economic growth.
The fifth explanation is that the effective use of ICT requires learning and adjustment costs, and the impact of ICT should become visible only with delay. As the widespread use of ICT is only a relatively recent phenomenon, according to this explanation the productivity paradox is a temporary phenomenon, and will go away soon.
The sixth explanation could be called the "Dilbert theory of productivity paradox." This is a more serious conceptual challenge for productivity research. According to Dilbert (cartoon of 5 May 1997), as quoted by Triplett (1999) "the total time that humans have waited for Web pages to load ... cancels out all the productivity gains of the information age." In fact, this theory could be extended to a Schumpeterian model, where the extremely rapid obsolescence of ICT leads to negative productivity impacts through destructive creation. In this model, advances in technology might push Dilbert into a cyberspace singularity where he has to go for shopping for new computers faster than his computer seller’s e–commerce Web pages become visible [85].
The final explanation is that there is no paradox. Although there are all kinds of new products around us, and the world seems to be transforming into a "new economy" driven by ICT, in economic terms this revolution may be an illusion. The Solow residual measures the growth rate of economic efficiency, and it grows only if the rate of growth of economy grows faster than inputs. In other words, to have an impact on productivity growth rate and the Solow residual, the rate of "technical advance" should be increasing. Although it is common to talk about the increasing speed of technical change, historical data does not necessarily support this view. It is clear that there are now more new products introduced every year than before. The rate of change may, however, have remained quite stable for centuries. Triplett quotes Diewert and Fox (1999), who pointed out that the growth in the number of products in the average grocery store had actually fallen from its 1948–72 level in the 1972–94 period [86].
Brynjolfsson (1993) added an important point to this discussion by highlighting the point that it is quite possible that decision–makers in business firms can also make bad decisions. They can, for example, over–invest in ICTs. If economists are unable to accurately measure the benefits of ICT investments, why would an average manager be a more perfect decision–maker? [87] In fact, a 1995 study by Standish Group estimated that 32 percent of corporate IT projects were abandoned before completion in the U.S., leading to a cost of $81 billion. A U.K. Study by OASIG estimated in 1996 that 40 percent of software projects were totally abandoned before completion and that a further 25 percent were substantially truncated and simplified during the implementation (cf. Ewusi–Mensah, 2003).
Brynjolfsson and his colleagues also noted that if the main impact of ICT investment is to redistribute profits among competing firms, investments do not necessarily lead to increase in total industry output. The competitive impact of ICTs is conceptually independent of the aggregate productivity impact (Hitt and Brynjolfsson, 1994). In fact, much of the ICT investment in the 1990s was justified and motivated from the perspective of competitive advantage. Large firms invested heavily, for example, in business intelligence, data mining, market information, intellectual property management, and executive information systems. Investments in ICT can also be defensive investments that are necessary to avoid deterioration of competitive positions. The competitive impact of such systems may have been considerable but they have not necessarily increased total output, except by perhaps generating demand for ICT products, software, and consulting.
Such competitive strategic use of ICTs, indeed, could partly explain why ICT manufacturing saw major productivity improvements in the 1990s, when total factor productivity stagnated in most industries. Non–ICT industries may simply have given some of their profits to ICT industries [88].
In the second half of the 1990s, productivity grew impressively in the U.S. and also in other countries (Colecchia and Schreyer, 2002; Jorgenson, et al., 2003). This recent productivity revival has switched the focus of productivity paradox explanations somewhat. Steindel and Stiroh (2001) point out four important questions that remain open within the neoclassical growth framework.
First, there is the question whether the productivity revival in the second half of the 1990s was purely cyclical [89]. During the upswings in the economy, firms tend to invest more and this leads to capital deepening and increase in labor productivity. Total factor productivity is also pro–cyclical. As the economy grows, firms use their existing production capacity more efficiently. This typically becomes registered as an increase in TFP growth rate. It is quite clear today that the growth rate in the second half of the 1990s was unsustainable, and therefore it is possible that much of the productivity revival, in fact, is unrelated to the long–term productivity impacts of ICT. As Steindel and Stiroh note, this is particularly a challenge to the neoclassical productivity framework, which assumes that the economy is in equilibrium. For example, correct estimates of capital services are critically important in this framework, but they are calculated using the theoretical assumption that prices perfectly reflect the marginal products of different types of capital goods, which is clearly not true, at least over short time periods. In particular, it is not easy to argue that in the second half of the 1990s ICT investment decisions reflected perfect economic rationality [90].
The second issue highlighted by Steindel and Stiroh is the question why almost all productivity effects were seen in high–tech manufacturing [91]. As was pointed out above, the observed increase in TFP during the second half of the 1990s was strongly focused on ICT manufacturing industries in the U.S. [92] As Steindel and Stiroh note, the neoclassical productivity framework is by definition unable to answer this question. In this framework, total factor productivity change is defined as the residual growth that cannot be attributed to any known causes.
The third issue is why the measured productivity growth outside ICT manufacturing has been slow or negative. A common explanation for this is that service productivity growth is underestimated, as qualitative improvements in services are rarely visible in national accounts [93].
The fourth issue highlighted by Steindel and Stiroh is the estimation of sustainable productivity growth rates. Recent projections of economic growth have sometimes been based on the rapid productivity growth rates of the second half of the 1990s. If, for example, the slow rates of total factor productivity growth of the 1980s would be used instead, the estimates of future growth would look quite different.
Basically, the neoclassical explanations for the productivity paradox, then, argue that the theoretical apparatus used for measuring and analyzing productivity works, but the data, or our interpretations of it, might be inaccurate. It is, however, important to keep in mind the actual logic of growth accounting calculations. The growth accounting equations only output what we put in. If the starting point is that the equations exactly describe the contributions of different sources of growth, changes in the contributions of the various growth factors exactly reflect the changes we make in measuring these different factors.
This is important for two reasons. First, growth accounting studies typically very heavily rely on the validity of the used theory. As data that would be needed for empirical studies usually is not available, researchers in practice use the assumption that the theory is empirically accurate to switch to alternative data that are available and to fill in missing data sets. For example, because the weights that are needed to add growth factors together are not measurable, researchers assume that all income is distributed exactly according to the marginal productivities of the different productive factors. Furthermore, because marginal productivities of many capital investments are not available, researchers assume that firms adjust their capital stocks in real time, based on perfect knowledge about the future revenues generated by different investments. Missing empirical data are therefore filled in by defining how they should look if the theory would be correct. These results are then used to describe what has happened in the actual economy. Obviously, there can be no empirical test that could falsify the theory or its results, if we start from the assumption that the theory, in fact, is correct. As long as the theory defines what productivity is, and what kinds of jumps from theory to empirical world are allowed, the measured productivity growth is exactly what the theory tells it to be.
Therefore, it becomes also very important that we do not leap from these theoretically derived results back to empirical world without proper caution. For example, if we multiply the value of computer services by ten, we simply reallocate some of the total output toward computers, and the share of productivity that becomes associated with computers grows. In this framework, there is no causal connection between the different growth components and common sense "productivity" concepts [94]. As long as, for example, Barbie dolls would represent a similar share of investments in the economy as computers, the studies would show that Barbie dolls have been a major source of capital deepening and thus labor productivity growth. If the quality adjusted prices of Barbie dolls would be mismeasured, Barbies would also show up in total factor productivity growth. Although we might quickly protest that computers are obviously more productive than Barbie dolls, there is nothing in the theoretical structure of the growth accounting framework that could be used to argue for this [95]. Similarly, if we would have corrected the price indices for cars and transport equipment as aggressively as they have been adjusted for computers, productivity studies would reveal that cars have been a major driver of productivity in the 1990s in most developed countries. In this sense, ICT, in fact, is the "story" that underlies the productivity miracle of the 1990s. This is because in the neoclassical growth accounting framework any investment goods or services whose prices are deflated as fast as semiconductors and which become obsolescent as fast, become a disproportionately important factor of productivity growth, in particular if the adjustments are done only in this specific industry.
Computers, however, are special partly because of their rapid decay. Steindel and Stiroh, for example, argue, following the U.S. Economic Advisers, that the capital deepening component of labor productivity growth could be sustained also in the future. This is because computer prices drop rapidly and the current productive ICT assets can be maintained and increased with small nominal growth in investment spending. Here the logic is that because we do not need much money to keep on investing in ICT, capital deepening probably could go on. On the other hand, if the average age and lifetime of these assets is about three years, the total depreciation of these assets grows rapidly when their accumulated stock increases. Capital deepening requires that we invest more than all historically accumulated ICT stocks decay annually. The sustainability of capital deepening as the source of labor productivity growth, therefore, greatly depends on the time ICT products become obsolescent.
Steindel and Stiroh also highlight the open issue why the productivity has been concentrated in high–tech manufacturing. As was show above, an important source of growth in the ICT sector has been the way price indices are used. There is actually not much productivity increase in "high–tech manufacturing." The productivity increase strongly focuses on the semiconductor industry, for the reasons discussed above. Another potential reason, however, is that ICT manufacturing industries are exceptionally global. The biggest productivity increases are often seen in industries that use extensively international outsourcing of production. Some recent studies have argued that total factor productivity differences depend to large extent on international terms of trade and productivity flows across countries [96].
ICT producing industries and several ICT using industries are tightly connected international networks, where intermediate products and services flow constantly across industry and country borders. This creates major challenges for correctly allocating productivity changes among industries and national accounts. Productivity researchers have conventionally focused on national data, describing productivity developments in national "industries" that do not always exist in practice. Comparative assessments of productivity developments, and the impact of ICTs, therefore remain difficult to interpret. Productivity researchers may have been measuring the impact of globalization, instead of ICT [97].
ICT–related industry sectors also typically rely on labor compensation schemes that produce large errors in studies that use traditional sources for labor compensation data. When ICT developers are hired with stock options, for example, the recorded average income may be a minor part of the actual compensation [98]. Similarly, the average hours that are used to measure labor productivity may only inadequately reflect actual working hours. In the second half of the 1990s, these effects were clearly visible in the ICT industry. Their impact on productivity statistics is, however, unclear [99].
One important conceptual challenge in productivity studies is the apparent empirical mismatch between firm–level and macroeconomic productivity impacts. From the perspective of industry practitioners, international trade, financial services, semiconductor design, scientific research, and the production of movies, for example, are obviously much more efficient when we have telephones, computers, and data networks. On the other hand, we do not know how to link such task–level efficiency with macroeconomic measures of efficiency. As David (1999) has noted, there is a conceptual disconnect between our views on productivity at these different levels. Although it appears intuitively clear that computers allow us to be more productive at work, it is not clear what the macroeconomic effects of improved task productivity are.
In interpreting empirical research on ICT productivity impacts, it is also important to note that the neoclassical productivity framework is by definition unable to account for "non–market" factors that may influence productivity [100]. For policy–makers a particularly relevant invisible factor is policy. Although it is, in principle, possible to incorporate policy–related variables in the neoclassical framework, conceptually this often contradicts the basic neoclassical assumptions [101]. In fact, sometimes productivity changes result more from policy choices than from technological progress. Studies on productivity developments in the ICT manufacturing sector usually do not, for example, take into account the semiconductor trade agreements between Japan and the U.S., which led to price increases at the end of the 1980s [102]. Instead, policy changes become registered as changes in TFP, and sometimes interpreted as productivity impacts of technical advances in ICT.
ICTs as contextual and composite resources
If we look beyond traditional growth accounting for possible explanations of the invisibility of ICT productivity impacts, one important factor can be found from the way we have conceptualized ICT products. The accounting for ICT investment has mainly focused on ICT equipment purchasing costs, neglecting complementary costs in organizational change, skills, and system operating costs. If direct software and hardware costs account for less than 20 percent of the cost of new information system deployment, as for example Brynjolfsson and Hitt [103] suggest, the true cost and investment in ICTs is perhaps considerably higher than commonly estimated. Most of the ICT–related investments in firms and the society have perhaps therefore been left unaccounted for. More fundamentally, we do not know today how costs and benefits should be allocated in systems of production that consist of interoperating technical components and software applications, but also non–technical elements, such as accumulated human and social capital and organizational routines and work practices. For example, studies on electronic commerce and collaboration show that investment in learning and trust may be crucially important in realizing the productivity potential of ICT. Such investments are often accounted as consumption and expense, if at all.
From the user and investor point of view, ICT investments could be described as "composite resources." ICT hardware, which is commonly the focus of economic studies, has no productive value as such. It has to be combined with other resources before its productivity potential can be released. Furthermore, the optimal ways to combine ICT with other resources depend on historically accumulated st
