Intangible assets: computers and organizational capital.

Brynjolfsson, Erik ; Hitt, Lorin M. ; Yang, Shinkyu 等

IN DEVELOPED ECONOMIES, production requires not only such traditional factors as capital and labor but also skills, organizational structures and processes, culture, and other factors collectively referred to as "intangible assets." Detailed investigation of some of these types of assets has found that they are often large in magnitude and have important productivity benefits. For example, Dale Jorgenson and Barbara Fraumeni found that the stock of human capital in the U.S. economy dwarfs that of physical capital and has grown over time. (1) Bronwyn Hall, Zvi Griliches, and Baruch Lev and Theodore Sougiannis found evidence that research and development (R&D) assets bring benefits in the form of positive marginal product and market valuation. (2) Timothy Bresnahan, Brynjolfsson, and Hitt have found that certain organizational practices, when combined with investments in information technology (IT), were associated with significant increases in productivity in the late 1980s and early 1990s. (3)

Investors also attempt to incorporate intangible assets into their valuation of firms, and this is one reason that the market value of a firm may differ markedly from the value of its tangible assets alone. In particular, stock market valuations of firms have increasingly diverged from their measured book value in the past decade or so. (4) Part of the explanation may be the growing use of IT and the associated investments in intangible assets. (5) Whereas early applications of computers were primarily directed at factor substitution (particularly of low-skill clerical workers), modern uses of computers have both enabled and necessitated substantial organizational redesign and changes in the skill mix of employees. (6) Collectively, this research argues for a complementarity between computer investment and organizational investment, and specifically a relationship between use of IT and increased demand for skilled workers, greater decentralization of certain decision rights, and team-oriented production. Moreover, case studies and a growing body of statistical analyses suggest that these complementary investments are large. (7)

This paper analytically explores the hypothesis that new, intangible organizational assets complement IT capital just as new production processes and factory redesign complemented the adoption of electric motors over 100 years ago. (8) To realize the potential benefits of computerization, investments in additional "assets" such as new organizational processes and structures, worker knowledge, and redesigned monitoring, reporting, and incentive systems may be needed. We study how the financial markets can be used to help identify such assets.

In some cases the costs of implementing the new processes, training, and incentive systems may be many times greater than the costs of the computer technology itself. However, the managers who decide to incur these costs presumably expect the present value of the resulting benefits to be no less than these costs, even if they accrue over a period of years and are uncertain. In this sense managers' behavior reflects their belief that they are investing in an economic asset.

Assets that are intangible need not be invisible. On the contrary, the presence of intangible organizational assets can be observed in at least three ways. First, some of the specific changes that firms make may be directly observable. In particular, previous work has used survey methods to document a relationship between technology and some aspects of organizational change, such as new business processes, greater demand for skills, and increased employee decisionmaking authority. (9) Firms sometimes try to highlight their investments in these areas, offering tours to customers, investors, and researchers who express an interest in them. A visit to the manufacturing operations of Dell Computer or of a steel mini-mill provides some insight into the effort these firms put into creating various kinds of organizational assets and the resulting productivity implications. Recently, researchers have begun more systematic efforts to help quantify the extent to which companies have adopted various organizational practices. (10)

Second, the effect of these changes on a firm' s market valuation should be measurable. If these new practices really represent the types of organizational assets we described earlier, one would expect the accumulation of these assets to be reflected in firms' market value, as revealed by voluntary transactions among buyers and sellers of the firms' financial securities.

Third, these assets should provide real returns in the form of higher output. Thus a production function framework should reveal that firms that have put in place more of these intangibles saw greater output in subsequent years, after accounting for standard inputs (such as capital, labor, and materials).

Although we will examine all three of these indicators, our focus will be on the relationship between intangibles and the financial markets. Just as investors can visit various factories and buildings owned by a firm and attempt to judge their profit-making potential, they can also form their own judgments about the existence, relevance, and value of various intangible assets owned or controlled by the firm. One difference, however, is that firms do not report a value for many of the intangible assets on their balance sheets, forcing investors to rely on other sources of information to value these assets. As a result, investors and analysts appear to devote relatively more time and effort to assessing the value of companies with larger stocks of intangible assets. (11)

Although the data can be noisy, the valuations provided by the public capital markets do have some advantages for researchers in this area. Whereas the effects on productivity or other measures of economic output may be spread over many years, the financial markets, which seek to assess the discounted value of companies' future revenues, provide an immediate indicator of whether these investments are expected to generate value for a firm's owners. In particular, the market value of a firm that has leveraged computer assets with organizational investments should be substantially greater than that of a similar firm that has not. A computer that is integrated with complementary organizational assets should be significantly more valuable to a business than a computer in a box on the loading dock.

An important characteristic of the organizational capital created by corporations is that its value may not be realized for years, if at all. Firms choose to invest in certain business models, organizational practices, and corporate culture. Later some of these investments turn out to be more productive and profitable than others. The financial markets recognize and reward those models that are well suited to the current technological and business environment. At that point, other firms may try to imitate the winners' best practices, but the complexity due to explicit and implicit complementarities among each collection of practices makes this difficult. Kmart may wish it could emulate Wal-Mart, and Compaq may try to learn from Dell, but their adjustment costs may prevent this from happening for years, even if they succeed in the end.

Thus it would be unwise to interpret high market values on an input such as IT as reflecting high adjustment costs for the successful investors. On the contrary, the market is mainly valuing the intangible assets correlated with IT; if anything, IT-intensive firms are likely to have lower adjustment costs than their rivals and hence higher levels of IT investment. At the same time, the higher investment costs of the rival firms are what prevents them from quickly dissipating the rents of the winners. When complex combinations of technology and organization are called for, the costs of imitation and investment are likely to be especially high. Furthermore, looking at the valuations of winning organizational strategies ex post can give a misleading impression of their returns. Many, perhaps most, efforts at organizational change fail, and projects involving extensive investments in IT often fall well short of expectations. (12) Ex ante, a rational manager must consider the substantial risk of failure before deciding whether a project is likely to deliver the required returns.

This has implications for how one should interpret the coefficients on asset variables in a market value equation. In the traditional interpretation, (13) when a high market value is found to be associated with IT, it reflects the adjustment costs of investing more in IT--the shadow value of installed IT capital. In our interpretation, however, relatively little of the market value is due to this effect. Most of the value springs from intangible assets, including factors such as business organization, that are disproportionately high in IT-intensive firms. At the end of our sample period, it was easy to see that Dell's business organization was a winning model; at the beginning of the period, as that organization was being created, its value was much less obvious.

This argument leads to five hypotheses, which are empirically testable:

--Each dollar of installed computer capital should be correlated with more than one dollar of market value, after controlling for other measured assets.

--Investments in computers should be correlated with increased investments in certain observable organizational practices.

--If these practices represent part of the productive assets of a firm, they should also be associated with increases in market value.

--If intangible assets are most common in firms that combine these specific organizational practices with investments in computer capital, those firms should have a higher market value than those that adopt these same practices in isolation.

--Any intangible assets correlated with computerization and these specific organizational practices should also result in higher measured output in future years, reflecting the returns to these intangible assets.

Using data on 1,216 large firms over eleven years (1987-97), we find evidence supporting all five hypotheses:

--Each dollar invested in computers is associated with an increase in firm market valuation of over $10 (depending on the assumptions of the estimation models), compared with an increase of just over $1 per dollar of investment in other tangible assets.

--Firms that are intensive IT users are also more likely to adopt work practices that involve a specific cluster of organizational characteristics, including greater use of teams, broader distribution of certain decision rights, and increased worker training.

--This cluster of organizational characteristics increases a firm's market valuation beyond what can be accounted for by tangible assets.

--Firms that have adopted both these organizational characteristics and have a large computer capital stock have disproportionately higher market valuations.

--Firms with higher levels of computerization, especially when they also have higher levels of these organizational characteristics, have significantly higher output in subsequent years.

The primary alternative hypothesis for the high market-to-book ratio of many firms is some sort of investor mispricing, perhaps due to a market bubble, fads, or irrationality. Our argument is not that investors never make mistakes in pricing assets; undoubtedly this does happen. However, the five findings above are collectively difficult to explain as being due entirely to mispricing.

Furthermore, our examination of fixed-effect specifications and of specifications using differences over long periods suggests that our results are not driven by other types of unobserved firm heterogeneity or short-run correlated shocks between market value and computer investment. Similarly, the evidence suggests that they are not driven by a general complementarity between capital and skill: these results appear to be unique to IT capital and are not important for ordinary capital. Because our sample consists predominantly of large, established firms rather than new high-technology entrants, and because the time period of our data predates the large increase in the value of technology stocks in the late 1990s, our results are not driven by the possible presence of a bubble in high-technology stocks in the late 1990s. Moreover, our results are qualitatively similar for each individual year when estimated separately over our eleven-year sample period, which includes the peak and trough of a business cycle. This argues against the possibility that our results are simply driven by short-term stock market fluctuations. The results are consistent with earlier case-based research as well as with recent econometric work using production functions, which suggests an important role for IT-enabled organizational changes in increasing productivity and the value of firms. Taken together, these results lend quantitative support to the idea that IT is most valuable when coupled with complementary changes in organizational design.

Related Literature and Cases

We begin by summarizing some of the related literature, including some case examples that help provide some perspective and texture to the statistical results reported in that literature and later in this paper.

IT and Organization

This paper draws primarily on two strands of research and seeks to link them. Here we review some studies of the interaction of IT and firm organization; later, when we develop the model, we will draw on studies that use financial markets to provide insight into the size and nature of intangible assets.

For U.S. businesses the most important technological change in the last twenty years has been the increased power and ubiquity of computers and related technologies. The quality-adjusted price of logic and memory chips has declined by about an order of magnitude every five years, and that of many other components such as magnetic storage and data communications has declined at a comparable or faster rate; these trends show no signs of abating in the near future. Indeed, there is some evidence of an acceleration since 1995.

IT has effects on the organization that adopts it that are disproportionate to its share of the organization's costs. A firm's business processes, internal organization, and relationships with outside parties are significantly determined by the economics of information and communications. (14) For instance, lower-cost access to data and communications can exacerbate the information processing bottleneck at the tops of hierarchies and therefore increase the value of delegation and of decentralized, nonmachine decisionmaking. (15) It can also have direct and indirect effects on the value of skilled labor, job design, and incentive systems, in particular, Bresnahan, Brynjolfsson, and Hitt analyze data on IT, organizational practices, and productivity from over 300 large firms and conclude that

 IT use is also correlated with a pattern of work organization involving
 more decentralized decisionmaking and greater use of teams. Increases in
 firms' IT capital stock are associated with the greatest increases in
 output in firms that also have high levels of human capital or
 decentralized work organization, or both. However, firms that implement
 only one complement without the others are often less productive than firms
 which implement none at all. (16)

In other words, there is evidence of a complementarity between the use of IT and certain changes in work organization. Of course, the falling quality-adjusted price of IT raises the return to investments not only in IT but also in its complements. A significant literature, mostly outside of economics, has explored various aspects of the interaction between IT and organization, business processes, and even corporate culture. (17)

Although the organizational complements are valuable and, in some cases, even essential to the success of IT innovations, implementing organizational changes is costly and risky, yielding both successes and failures. (18) Both the case evidence and the econometric results suggest that the costs of these organizational complements to IT investments typically exceed the direct financial costs of the IT investments themselves. (19) Importantly, although many of these organizational practices may be readily visible to competitors and are copiously documented in articles by business school professors and consultants, they are notoriously difficult to imitate successfully. (20) This reflects complementarities and large effects of seemingly minor characteristics. Intel, for example, has adopted a "copy exactly" philosophy for any chip fabrication plant built after the first plant in each generation. Wholesale replication of even seemingly insignificant details has proved more reliable than trying to understand which characteristics really matter. Going from the plant level to the firm level only complicates the imitator's task.

The difficulty in implementing organizational complements can explain the apparent quasi-rents earned by firms that have been fortunate or skillful enough to have them in place. In some cases these quasi-rents, when measured in a production function framework, may show up as higher coefficients on other inputs to production. For instance, Brynjolfsson and Hitt and Frank Lichtenberg, (21) among others, find that heavy use of IT is correlated with significantly higher levels and growth rates of measured productivity. These rents may arise because nonadopters have not tried, or have tried and failed, to implement complementary workplace or product innovations. Brynjolfsson and Hitt interpret their productivity results as pointing to the existence of large but unmeasured inputs to production that are correlated with measured IT. A related literature finds that certain work practices and human resource policies are correlated with higher levels of productivity and thus constitute another typically unmeasured "input" to production. (22)

Because effective work organization can be costly to develop and implement but yields a stream of cash flows over time, it is natural to think of it as a kind of asset. This asset has variously been called "organizational capital," "e-capital," and "structural capital," depending on the context. (23)

Case Studies

Although statistical data are very useful for hypothesis testing, our own understanding of the role of organizational capital has been shaped in an important way through visits and interviews with managers who have implemented information systems projects and by teaching case studies on such projects. (24) Some common themes in these cases are the following:

--Computers and software are just the tip of a much larger iceberg of implementation costs. Successful projects require enormous management attention, worker training, and changes in seemingly unrelated areas of the business and perhaps the entire industry. Successful chief information officers are now expected to combine knowledge of technology with an understanding of the firm's business opportunities and challenges.

--Many of the practices that matter most are also to be found at the level of the business culture and work content of individual workers, not just in sweeping visions on the part of the chief executive officer or the chief information officer. As a result, organizational capital is quasi-fixed in the short run.

--Information technology initiatives are difficult and often fail. By the same token, one of the reasons they can provide competitive advantage and quasi-rents is that they are not trivial for other firms to duplicate. Wal-Mart, with a recent market capitalization of $273 billion and net tangible assets of $25.5 billion, is an example. (25) Wal-Mart has spent over $4 billion on its "retail link" supply chain system, and it has been called "by far the commercial world's most influential purchaser and implementer of software and systems." (26) A recent McKinsey Global Institute report singles out Wal-Mart for playing a disproportionate role in the productivity revival in the 1990s:

 Productivity growth accelerated after 1995 because Wal-Mart's success
 forced competitors to improve their operations.... By the mid-1990s,
 [Wal-Mart's] productivity advantage widened to 48%. Competitors reacted by
 adopting many of Wal-Mart's innovations, including ... economies of scale
 in warehouse logistics and purchasing, electronic data interchange and
 wireless bar code scanning. (27)

A key point in that report is that "[IT] was often a necessary but not sufficient enabler of productivity gains. Business process changes were also necessary.... " (28) Or, as Robert Solow puts it, "The technology that went into what Wal-Mart did was not brand new and not especially at the technology frontiers, but when it was combined with the firm's managerial and organizational innovations, the impact was huge." Solow concludes, "we don't look enough at organizational innovation." (29)

Other highly visible, computer-enabled business changes happen on factory floors and in back offices. For instance, Dell combined new materials management software with a set of redesigned workflows to roughly halve the floor space required in its main server assembly plant, while increasing overall throughput and reducing work-in-process inventories. Had Dell instead built a second factory on the site, the additional real estate and capacity would have been duly recorded on its balance sheet. In contrast, the processes that doubled the effective size of its existing facility went unrecorded.

Similarly, a Johnson & Johnson factory producing adhesive bandages dramatically increased the variety of products it could manufacture and reduced costs after combining new, computer-based flexible machinery with nearly a dozen carefully defined work practices and principles, including changes in the allocation of certain decision rights, incentive systems, and job responsibilities. (30) The right combination of work practices was discovered only after a lengthy and costly period of experimentation and false starts. When a system found to be highly effective was ultimately implemented, management ordered the factory windows painted black to prevent competitors from quickly learning the details of its implementation.

Other firms are more eager to disseminate their discoveries about organizational complements to IT. For instance, Cisco Systems has identified a set of practices and attitudes that the company associates with increased productivity from the use of the Internet; it calls these practices and attitudes its "Internet Culture." The firm has established a culture that produces the results it is looking for, and it invests heavily in maintaining that culture. Cisco not only has a "Director of Internet Culture" but also issues plastic cards, which employees are asked to carry with them, that summarize the eleven key components of that culture. (31) Furthermore, as a leading provider of some of the basic hardware of the Internet, Cisco encourages other firms to understand and adopt these practices, which they believe make investments in Internet technologies more productive.

Accounting for Intangible Assets

In each of these cases, the basic technology was available to all interested parties. However, the truly valuable assets were the complementary business processes, work practices, and even culture, all of which were harder to identify and implement. In effect, these constituted an organizational asset with real value, although one not reflected on the firm's balance sheet.

Even many of the direct project costs of an IT project may elude documentation on a firm's balance sheet. For example, less than 20 percent of the typical $20 million installation cost of a SAP R/3 system (a widely used large-scale package designed to integrate different organizational processes) is for hardware and software, which is capitalized; by far the greater part of the investment is for hiring consultants to help develop requirements; evaluate, select, and customize the software; redesign organizational processes; and train the staff in the use of the new system. (32)

According to the American Institute of Certified Public Accountants, although the costs of software purchase and development should be capitalized if they exceed some threshold, (33) most of these other project costs must be expensed. (34) As firms devote more resources to various IT projects, this accounting policy drives a wedge between the market value of a firm's assets and their value on its balance sheet. (35) As noted by Lev and Paul Zarowin:

 restructuring costs, such as for employee training, production
 reengineering or organizational redesign, are immediately expensed, while
 the benefits of restructuring in the form of lower production costs and
 improved customer service, are recognized in later periods. Consequently,
 during restructuring, the financial statements reflect the cost of
 restructuring but not its benefits, and are therefore largely disconnected
 from market values which reflect the expected benefits along with the
 costs. (36)

The accounting policy of excluding many such intangible assets from a firm's balance sheet should not be taken as an implication that they have no economic value, or even that their economic value is unknown and unknowable. On the contrary, it partly reflects the different goals of accountants and economists. No single number is a correct reflection of the value of an asset in all states of the world. A creditor, when evaluating a piece of collateral, might care most about that asset's value in those states of the world where the debtor is unable or unwilling to make interest payments. In those circumstances collateral may need to be seized and sold for salvage value, and then the value of many intangible assets, such as organizational capital, is likely to be very low or zero. Accountants, to the extent they are responsible for providing useful information to creditors and potential creditors, might reasonably adopt a conservative valuation for :many assets, particular those that have little or no salvage value in "bad" states of the world. Furthermore, a financial accountant needs to provide, for outside parties, reliable numbers that are not easily subject to "earnings management" or other types of gaming by management, since the interests of management and creditors or potential creditors are not always aligned. According to the Financial Accounting Standards Board, more conservative rules for recognizing assets are called for when valuations are uncertain. (37)

Equity investors care less than other creditors about the value of assets in "bad" states of the world, and more about the expected cash flows those assets can generate across all states of the world. This is one reason that, when assets have different values in different states of the world, there may be a large, and perfectly sensible, gap between balance sheet assets and the market value of a firm. To the extent that an economist is interested in the expected value of a firm's assets, their market value as judged by equity investors may provide a more accurate guide than the balance sheet. In this paper we attempt to make use of these judgments.

Econometric Model and Data

Here we sketch the derivation of our stock market valuation model and describe the data used in the analysis.

Derivation of Model for Stock Market Valuations

Our model draws on standard finance theory and assumes that managers are rational in their investment decisions and that investors are rational when they make their judgments about the valuation of corporate securities. Of course, this does not mean that their decisions will always be correct ex post; uncertainty and imperfect information make that unlikely. In addition, bubbles and other anomalies in the valuation of financial assets can make the interpretation of the econometric results more difficult. We address these possibilities in a variety of ways.

The basic structure of the model follows the literature on the valuation of capital goods that relates the market value of a firm to the capital goods it owns. (38) Others have taken variations of this framework adapted for empirical use and applied them to the valuation of firms' R&D, (39) and the relationship between firm diversification and firm value, using firm-level data. (40)

The empirical use of Tobin's q to capture intangible organizational assets has been proposed by other authors and has been significantly advanced recently by Robert Hall, who states, "Firms produce productive capital by combining plant, equipment, new ideas and organization," and concludes, "The data suggest that U.S. corporations own substantial amounts of intangible capital not recorded in the sector's books or anywhere in government statistics." (41) Elsewhere Hall discusses the analogy between a flow of investment in reorganization and a flow of investment in physical capital. (42) Whereas he uses the label "e-capital" to describe all intangible assets revealed by the gap between the financial markets' value of firms and the value of their replacement assets in the 1990s, this paper seeks to identify more explicitly the role of computer capital and particular organizational practices. (43) Thus our paper is most closely related to that of Brynjolfsson and Yang, (44) who found evidence of high q values for IT but did not explicitly link them to organizational investments.

We assume that firms face a dynamic optimization problem in which managers make capital investments (I) in several different asset types and expenditures in variable costs (N) with the goal of maximizing the market value of the firm (V). In turn, V is equal to the present value of all future cash flows [pi](t) according to a discount function u(t). The accumulation of capital investment, less depreciation ([delta]), produces a vector of the capital stock (K, which includes different components of capital [K.sub.j], j = 1 ... J, where the js are computer capital, other physical capital, and so forth). The capital stock, along with the variable inputs, is used to produce output through a production function (F). This yields the following program:

Maximize with respect to I and N

(1) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

where

(2) [pi](t) = F(K, N, t) - N - I,

and the following holds:

(3) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

Under the assumptions that F(K, N) is a homogeneous function of degree 1 over K, N, and I (constant returns to scale) and is twice differentiable, one can solve for the market value of the firm that results from this optimization problem. If all assets can be documented and no adjustment costs are incurred in making them fully productive, buying a firm is equivalent to buying a collection of separate assets. (45) Thus the market value of a firm is simply equal to the current stock of its capital assets:

(4) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

Econometric Issues

The formulation in equation 4 suggests a basic estimating equation that relates the market value of firm i to the assets that the firm possesses, allowing for repeated observations over time t:

(5) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

If the vector of assets K for each firm contains all relevant capital assets and there are no other forms of specification or measurement error or adjustment costs, we would expect that [alpha] = 0 and [v.sub.j] = 1 for all j. However, v may deviate from 1 if adjustment costs are significant or if there are omitted variables that are correlated with the quantity of observed capital assets.

In the presence of adjustment costs, the shadow value of installed capital can exceed its acquisition costs. Capital that is installed may be more valuable than capital that is not yet installed. For example, if there are two types of capital, computers ([K.sub.c]) and other physical capital ([K.sub.p]), then ([v.sub.c] - 1) would represent the difference in value between computer capital that is fully integrated into the firm and otherwise identical computers that are available on the open market, and ([v.sub.p] - 1) would be the corresponding value for other types of capital. (46) As shown by Fumio Hayashi, (47) v'K can be made observable when there are constant returns to scale, because marginal and average q will be equal. In essence, the value of the firm will simply be a function of the capital price vector (v) and the capital quantity vector (K) of each of the types of capital it owns.

Similarly, the observed market value of each capital asset may also deviate from 1 when there are other capital assets that are not measured, such as omitted intangible assets, or shocks to market value that are correlated with the levels of observed capital assets. For instance, these intangible assets might include organizational capital that is complementary to certain observed capital assets or persistent, firm-specific components of value (such as management quality) that are correlated with capital quantity. Market value shocks include persistent errors in stock market valuation that are simultaneously correlated with capital assets (for example, the stock market over- or undervaluing high-technology firms), or short-run events such as an increase in a firm's market size or opportunities that raises stock market value and induces capital investment. These specification errors can be represented as a systematic omitted component of market valuation (Mi,) in the theoretical market value relationship:

(6) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

From standard omitted-variables arguments, (48) the presence of [M.sub.it] in the market valuation equation will alter the estimates of the value of capital assets ([v.sup.*], a vector) in a systematic way depending on the correlation of the observed capital assets (K, a matrix with rows representing different assets for each firm and columns representing different firm-year observations) with the omitted component of market value (M, a column vector with elements [M.sub.it]). Specifically, in the absence of adjustment costs,

(7) [v.sup.*] = [(K'K).sup.-1] K'M.

This implies that [v.sup.*] is simply the vector of coefficients that would arise in a regression of the capital assets on the omitted market valuation component (M):

(8) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

Thus, in the absence of adjustment costs, a high value for a particular capital asset in the market value equation implies a correlation between that asset and M, for instance, a large stock of (valuable) intangible assets or investor errors in valuation that is correlated with the quantity of observed assets.

As noted by Stephen Bond and Jason Cummins, (49) distinguishing the exact nature of this relationship is more difficult. If we interpret M as arising from the omission of intangible assets such as organizational capital from the equation, we can write [M.sub.it] = [Q.sub.it] [K.sub.o,it], where [Q.sub.it] is the market-determined shadow price of organizational capital and [K.sub.o,it] the quantity of organizational capital. In general, it is difficult without making further assumptions to distinguish the price and the quantity of organizational capital individually, (50) but for purposes of identifying the value of organizational complements to computers we need only determine the value of [M.sub.it]. However, if there is a bubble in the valuation of corporate securities (an "error" in the markets' perception of [Q.sub.it]), this can alter estimates of [M.sub.it].

Instrumental variables techniques will be of little help insofar as they do not distinguish between true organizational complements and errors in valuation. Removing the influence of factors that are both unobserved and correlated with productivity would remove the very variables we seek to measure: intangible assets. If such a technique were successful, the coefficients on the observed capital assets would be driven close to their theoretical value of 1, but no light would be shed on the magnitude of the intangibles.

However, if data are available that allow one to measure some of the components of [M.sub.it], it may be possible to partially distinguish the relative contribution of market valuation errors and the relative contribution of the intangible assets that comprise [M.sub.it]. We may also be able to reduce the impact of identifiable sources of correlated shocks by means of other econometric adjustments such as control variables for time periods or industries.

Our analysis focuses on obtaining data and measuring the contribution of various aspects of organizational capital ([K.sub.o]) that, from our earlier discussion, represent potential components of M either alone or in combination with computer assets ([K.sub.c]). We assume that the level of [K.sub.o] is difficult to change (that is, quasi-fixed) in the short run and thus can be viewed as exogenous with respect to both computer asset levels and market valuations. (51)

There are several ways in which [K.sub.o] can influence market valuation. First, it can have a direct correlation with market value while being orthogonal to all other assets. Directly incorporating measures of [K.sub.o] in the regression will then improve the efficiency of the estimation while having no impact on the other coefficients.

Second, [K.sub.o] can influence market value through its correlation with other assets. In general, the degree of bias in the estimated components of v depends on the correlations among all capital assets. However, if [K.sub.o] is positively correlated with [K.sub.c] but orthogonal to the other capital components (including any components of M other than [K.sub.o]), from equation 7 this will reduce the coefficient on [K.sub.c] when direct measures of organizational capital are present in the regression. (52) The reduction is proportional to the correlation between [K.sub.c] and [K.sub.o].

Finally, [K.sub.o] may influence market valuation disproportionately when [K.sub.c] is also large, if there are additional, unobserved intangible assets that are correlated with the simultaneous presence of both [K.sub.c] and [K.sub.o]. Under standard q theory, the coefficient on an asset can be interpreted as a function of the adjustment cost of increased investment in that asset. Thus the presence of arbitrage opportunities normally requires that the market value correlated with an asset be the same regardless of what other assets are also present. However, this need not be the case if the value of otherwise unobserved intangible assets varies systematically with the relationship among observed assets, yielding an additional correlation with market value above their simultaneous direct correlation. That is, there may be a distinct intangible asset that is correlated with the combination of [K.sub.c] and [K.sub.o] (but not necessarily with these :inputs separately). In particular, the simultaneous presence of high values for both [K.sub.c] and [K.sub.o] may signal that a firm has successfully adopted a certain work system. If that work system is both valuable and costly to implement, even for firms that have already adopted [K.sub.c] or [K.sub.o] alone, firms with the combination of both inputs should be expected to have disproportionately higher market values. This interpretation is consistent with the findings of the literature on IT impacts. (53) Unlike the previous two relationships, which can be captured simply by introducing [K.sub.o] into the regression in levels, this relationship will be revealed by the interaction [K.sub.c] x [K.sub.o].

Collectively, these relationships can be captured by including both [K.sub.o] and [K.sub.c] x [K.sub.o] in the regression. In addition, we can also test the uniqueness of this organizational relationship to computers by estimating the correlations between the other components of K and [K.sub.o] as well as their interactions in the market value equation.

The estimation of the contribution of intangible assets relies on minimizing other omitted variables that are correlated with asset levels. To the extent that many of these types of variables are common across the entire economy (for example, changes in the price of new capital investment, or the rate of overall economic growth) or unique to particular industries (such as the introduction of a new production technology throughout an industry), they can be accounted for by controls for year and industry.

If instead the omitted variables are time-invariant factors that are specific to individual firms, they can be removed by estimating difference equations that remove the contribution of firm-specific effects. Moreover, differences measured over long intervals (long differences) may be robust to a variety of other short-run shocks to the extent that asset levels and market value have sufficient time within the difference interval to return to equilibrium levels following a shock. However, these types of techniques also remove at least some of the true organizational capital that we are looking for, to the extent that organizational practices differ across firms and are relatively slow changing.

Finally, some omitted variables, such as R&D investment or advertising, could be indicators of other assets. To the extent this is a problem, these variables can be directly incorporated into the estimating equation as additional covariates.

For purposes of estimation we divide assets into three categories: computers, other permanent physical assets (property, plant, and equipment, or PP&E), and other balance sheet assets (receivables, inventories, goodwill, and other assets). We deduct current cash balances both from market value and from other assets. We also include control variables: the ratio of R&D capital to sales, the ratio of advertising expense to sales, dummy variables to account for missing observations on R&D or advertising expenditure, industry dummies (usually at the two-digit Standard Industrial Classification, or SIC, level), and year dummies. (54)

Data Sources and Construction

The data set used for this analysis is a panel of computer capital and stock market valuation data for 1,216 firms over the 1987-97 period, matched to a cross-sectional survey of organizational practices conducted in 1995 and 1996. A brief description of each data source follows; the appendix provides additional detail.

COMPUTER TECHNOLOGY. The measures of computer use were derived from the Computer Intelligence Infocorp (CII) installation database, which details IT spending by site for Fortune 1000 companies. Data from approximately 25,000 sites were aggregated to form the measures for the 1,000 companies that represent the total population in any given year. This database is compiled from telephone surveys that gather detailed information about the ownership of computer equipment and related products. Most sites are updated at least annually, with more frequent sampling for larger sites. The year-end state of the database for each year from 1987 to 1997 was used for the computer measures. (55) From these data we obtained the total capital stock of computers (central processors, personal computers, and peripherals). The IT data do not include all types of information processing or communications equipment and are likely to miss some portion of computer equipment that is purchased by individuals or departments without the knowledge of information systems personnel. (56)

ORGANIZATIONAL PRACTICES. The organizational practices data in this analysis come from a series of surveys of large firms. These surveys adapted questions from previous surveys on human resource practices and workplace transformation. (57) The questions address the allocation of various types of decisionmaking authority, the use of self-managing teams, the breadth of job responsibilities, and other miscellaneous characteristics of the workplace (further detail appears in the results section). Organizational data were collected at the end of 1995 and early 1996, covering most of the Fortune 1000. This yielded a cross section of 416 firms, with a survey response rate of 49.7 percent. We detected no significant pattern of response bias when the sample was compared with the population of firms in the Fortune 1000. Of the 416 firms that responded to the survey in some way, we have complete IT, organizational, and financial data for a total of 272.

MARKET VALUATION AND OTHER DATA. Compustat data were used to construct stock market valuation metrics and provide additional firm information not covered by other sources. Measures were created for total market value (market value of equity plus book value of debt), PP&E, other assets, R&D expense, and advertising expense. For the productivity analysis we also compute constant-dollar value added, labor input, and the capital stock. We removed from the sample those firms for which the data were inconsistent from year to year, firms that principally produced computers or software, and firms in the communications sector (SIC 4813). The last two groups of firms were removed because the nature of computers and telecommunications equipment as both a production input and output makes these firms very different from the rest of the economy. (58)

The full data set comprises 7,564 observations over eleven years for market value and computer capital stock, with each of 1,216 firms represented by at least one observation. After matching these data to the organizational practices surveys, we had complete organizational and market value data for a subsample of 272 firms, for a total of 2,097 observations.

Results

We performed regression and correlation analyses to test our five hypotheses. First, we explored the basic relationship between IT and market value for our full sample of firms. We then used correlation analyses to examine the relationship between computer capital and the adoption of specific organizational practices, and we constructed a single variable, ORG, to capture a portion of the relevant variation in organization across firms. It is this variable that will represent our (noisy) measure of organizational capital. Third, we investigated the effect of ORG on firm market value. Fourth, we studied how the combination of ORG and computers affects market value. Finally, we examined how these variables affect output in a production function framework. We also performed a number of checks on the robustness of our analysis and considered alternative hypotheses.

Computers and Market Value

BASIC FINDINGS. We begin by replicating earlier work by Brynjolfsson and Yang with our slightly larger data set. (59) Table 1 reports results of regression analyses examining the relationship between computers and market value. This equation relates market value to the three types of assets identified above: computers, PP&E, and other assets (principally, accounts receivable, inventories, and liquid assets other than cash). (60) Because we are pooling multiple firms in multiple years, we include dummy variables for each year and two-digit SIC industry. With the exception of our regressions using least absolute deviation (LAD) techniques, we use Huber-White robust standard errors or random-effects models to account for multiple observations of the same firm over time. (61) We also include measures of firms' R&D-to-sales and advertising-to-sales ratios. The ordinary least squares (OLS) regression reported in table 1 finds that each dollar of installed PP&E is valued at about $1.47, somewhat larger than the theoretical value of $1 that would be expected if there were no adjustment costs or correlated intangible assets. The market value of each dollar of other assets is close to $1; apparently these assets are less subject to adjustment costs or omitted components of market value.

Strikingly, however, each dollar of computer capital is associated with about $12 of market value. This apparent excess valuation of computers suggests the presence of substantial intangible assets, adjustment costs, or other omitted components of market value correlated with computer assets. In these and all subsequent regressions, time and industry controls are jointly significant (results not shown). We are thus able to remove some of the temporal shocks and other omitted components of market value unique to time period and industry. Although we do not have capital stock values for R&D or for advertising, we do have the investment flows for some of the firms in our sample, and we include their input shares as controls; these are also significant in most specifications (not shown).

Table 1 also reports estimates of the same equation using an LAD regression technique, which minimizes the sum of absolute values of the residuals rather than the sum of the squared residuals as in an OLS regression. This technique not only minimizes the influence of outliers but also reduces the impact of heterogeneity in firm size in our sample. This approach produces a similar estimate for the coefficient on computer capital (now 11.88), which is still far greater than the theoretical baseline of $1. The coefficient on PP&E falls slightly (to 1.18), whereas the coefficient on other assets is essentially the same.

LONG-DIFFERENCE SPECIFICATIONS. Our earlier discussion suggests that certain types of organizational practices are likely to have a significant influence on the value of computer assets. One way to account for these practices without measuring them directly is by estimating a difference specification that eliminates the contribution of any time-invariant, firm-specific component of market value. To the extent that organizational assets can be viewed as quasi-fixed, at least over moderately long periods, this suggests that we may be able to examine the amount of computer value potentially attributable to these types of factors. Table 2 reports estimates of our basic specification (including year and industry dummy variables) in differences ranging from those at one year to those at ten years.

Although any of these differences would presumably remove all time-invariant, firm-specific characteristics, these alternative specifications may yield different results for at least two reasons. First, longer differences are much less subject to bias from measurement error in the independent variables. (62) Thus, if measurement error were the only concern, we would expect the longest differences to produce estimates closest to the "true" coefficient values. Second, longer differences allow for more time for market values or asset quantities affected by short-run shocks to return to equilibrium values. Thus, varying difference lengths may enable comparisons of short-run and long-run relationships. (63)

The first several columns of table 2 suggest that, in equations using short (one- and two-year) differences, changes in computer asset levels appear to have no significant correlation with changes in market value, whereas in equations with longer differences the relationship is substantial. The coefficients rise from essentially zero for one-year differences to around 10 for five-year differences, and stabilize beyond that. The point estimate for the longest difference possible in the sample (ten years) is considerably higher, but because of the small sample size it is very imprecisely estimated and not statistically different from the other long-difference coefficients.

These results have several interpretations. First, they suggest that there may be considerable measurement error in the estimates of computer assets, biasing downward the short-difference more than the long-difference coefficients. This explanation also implies that the "true" estimate of the computer coefficient is more closely approximated by the longer-difference estimates, which are all considerably above 1. Second, the results may suggest that it may take several years of adjustment for computer assets to become valuable; this would be consistent with the nature of the complementary organizational changes discussed earlier. These results also provide some evidence against many types of correlated shocks, such as a new invention that immediately raises market value and requires additional investment in computer equipment for new production facilities. Presumably these types of shocks would act on a much shorter time scale, such as one year. These regressions also eliminate the bias due to omitted variables that are time invariant. However, these results do rule out the possibility that the computer estimates are affected by gradual adjustment over long time periods to firm-specific shocks.

YEAR-BY-YEAR ANALYSIS. Another way to examine the robustness of the results is to examine year-by-year cross sections of the results. If the results are biased upward by short-run shocks, some years will have disproportionately high values while others will be close to their equilibrium value. Table 3 presents regressions for each of the eleven individual years in our sample. Although there is some year-to-year variation in the computer asset coefficients, there is no particular time trend, and none of the estimates are statistically different from the estimate based on the pooled data of approximately 12. Although this does not rule out the possibility of considerable changes outside the 1987-97 time period or our sample of firms, it does show that, for our sample, computer asset values consistently show coefficient values of 9 or greater. (64)

Basic Findings Regarding the Role of Organizational Structure

Here we report correlations found between computer asset values and various measures of internal organization. (65) All correlations are Spearman rank-order correlations between various measures of computers and the organizational variables, controlling for firm size (employment), production worker occupation, and industry. (66) We used three different measures of a firm's IT: the total value of the IT installed base, total central processing power (in millions of instructions per second), (67) and the total number of personal computers. We used multiple measures because they capture slightly different aspects of computerization (for example, central processing power measures centralized computing assets, whereas the number of personal computers measures decentralized computing assets).

Table 4 presents correlations between each of these different measures of IT and four dimensions of organizational design: structural decentralization, individual decentralization, team incentives, and skill acquisition. Previous theoretical and empirical work has linked these types of practices to IT investment. (68) Consistent with our argument that IT and organizational practices are complementary, we confirm that, across multiple measures of IT and multiple measures of organization, firms that use more IT differ statistically from other firms: they tend to use more teams, have broader job responsibilities, and allocate greater authority to their workers, even after controlling for firm size and industry. These are only broad averages and do not apply to all firms in all circumstances: many successful IT users do not implement all or even any of these practices. In particular, computers have helped centralize a large subset of decisions involving aggregate data analysis (analyzing bar code data, for example), even as they have facilitated the decentralization of many decisions that require on-the-spot information, human relations, exception processing, and nonroutine inference.

In addition to being correlated with IT, these practices are all correlated with each other. Following Brynjolfsson and Hitt, (69) we constructed a composite variable (ORG) as the standardized (mean 0, variance 1) sum of the standardized individual work practice variables. This allowed us to capture, in a single construct, an organization's overall tendency to use this collection of work practices, which we could then use for further analysis. A principal components analysis (reported in table 5) showed that all components of this variable have high loadings on a single factor (which explains approximately 35 percent of the variance of these measures), and a scree plot (not shown) suggests that this is the only nonnoise factor. We interpret ORG as potentially capturing some aspects of the relevant organizational capital, [K.sub.o], as described above. We have no expectation that our simple measures will capture more than a small portion of such capital, but rather see it as an early attempt to demonstrate the existence and economic relevance of work practices in this kind of framework.

The composite variable ORG is highly correlated with computerization, consistent with our earlier discussion. In what follows we will explore the influence that this cluster of practices has on both the market value of the firm and the market value of computer capital.

Our earlier arguments suggested that these types of organizational practices should be strongly related to computer assets, but they are presumably less related to traditional physical capital (PP&E) or other assets. One way to investigate this relationship is to examine a simple conditional correlation equation that relates the logarithm of asset quantity to the price of the asset (proxied by time dummy variables), firm size (proxied by the logarithm of employment), and ORG. We also include controls for two-digit SIC industry to control for sample heterogeneity.

Table 6 shows the results of this analysis for the three types of assets we consider. ORG is correlated with greater use of computer assets (first column): firms with a value for ORG 1 standard deviation above the mean have 18 percent more computer assets on average (this result is statistically significant at the 1 percent level). However, PP&E and other assets have no significant relationship with ORG (second and third columns). In addition to confirming our hypothesized relationship between computers and ORG, this suggests that we can treat the other capital components as essentially orthogonal to ORG; thus any omitted-variables bias that is related to ORG is likely to affect the measured value of computer assets, but not that of the other assets. The correlation between IT and ORG implies that firms with high levels of ORG either receive greater benefits from investments in IT, or have lower costs of adoption, or both.

The Relationship between Organizational Structure and Market Value

ORGANIZATION VARIABLE IN THE MARKET VALUE EQUATION. Here we extend our basic estimating equation to include ORG and its interaction with computer assets, using various regression techniques. Regression equation 7-1 in table 7 is the baseline OLS regression (comparable to the first column of table 1) for market value for the subsample of 272 firms for which we have data for the ORG variable. These results are similar to our earlier baseline results from the full sample. Next we introduce the ORG variable into the regression in several ways. Because the ORG variable is an index (with a mean of 0 and a variance of 1), it must be appropriately scaled to be comparable to the other inputs. As ORG is essentially a concept relating to the organization of labor (education levels, training costs, allocation of certain decision rights to employees), it is natural to scale ORG by employment in considering its direct contribution to market value. The results of including the scaled ORG term in the regression are shown in column 7-2. Note that the coefficient on computer assets drops from 14.6 to 11.5 (approximately 20 percent), suggesting that at least some of the high observed value of computers is a result of computers serving as a proxy for organizational assets, as would be expected from our earlier discussion on omitted variables.

Of course, the simple measures that constitute ORG represent only a small fraction of the putative organizational capital of a typical firm. Furthermore, there are undoubtedly circumstances in which these particular practices are not beneficial. Nonetheless, we do find some evidence that investors, on average, may be treating the organizational practices that we measure much like more tangible types of capital by recognizing their contribution to the market value of a firm. All else equal, firms that are endowed with workers with greater education and training, and that have developed systems that support more decentralization of certain decision rights, appear to be valued at a slight premium to their industry peers. This relationship accounts for at least a part of the "excess" valuation of computers.

INTERACTION BETWEEN ORGANIZATION AND OTHER ASSETS. We next consider how ORG relates to the value of computer assets by including terms interacting ORG and each of the three capital components: computer assets, PP&E, and other assets. In computing these and other interactions that include terms already in the regression, we center both variables so that the existing linear terms in the regression have the same interpretation. (70) The results are shown using OLS regression (column 7-3) and alternative specifications (columns 7-5, 7-6, and 7-7).

First, we find that when the interaction term is included in the regression, the coefficient on IT drops substantially. In the OLS regression (column 7-3), the coefficient on IT drops by roughly 50 percent compared with the regression that includes only the scaled ORG term (column 7-2) and is only 40 percent of the original value in the baseline regression (column 7-1). A similar, perhaps slightly larger difference for the IT coefficient is found when we use a generalized least squares random effects model, which also yields slightly more precise estimates for the interaction terms (columns 7-4 and 7-5). This is notable because it suggests that much of the apparent excess value is located in firms that have high levels both of IT and of organizational assets. Note that, because we centered the interaction terms, this is not due to a simple correlation between IT and its interaction with ORG.

Second, the interaction of ORG with computer assets is significant in all specifications and is especially strong when we use a random-effects or a robust regression technique (LAD) to perform the estimates (columns 7-5 and 7-6).

Third, this relationship appears to be unique to computer assets. Although it appears that there is a small, positive interaction between ORG and PP&E, this relationship is relatively weak and statistically significant only in the LAD regression. The interaction between ORG and other assets is close to zero and inconsistent in sign.

As a final check on the results, we also perform a fixed-effects regression (column 7-7). Because ORG is time invariant, the interaction terms have to be interpreted as the relationship between changes in asset quantity (relative to the firm average) and changes in market value mediated by the level of ORG. Here we obtain similar coefficients on the interaction terms, suggesting that the intangible assets we are capturing through the interaction terms are not simply fixed firm effects but the result of factors that have varied, uniquely, at the firm level over time.

Taken together, these results suggest that firms that have large investments in computers have a disproportionately large amount of other intangible assets, and that these firms account for a substantial portion of the excess value of computers above and beyond the direct contributions of ORG. This finding is consistent with the case literature on the complementarities between IT and organizational structure. It is also consistent with the econometric finding of Bresnahan, Brynjolfsson, and Hitt, (71) using similar data, that firms that combine IT and ORG are disproportionately more productive than firms with high levels of only one or the other. Our interpretation is that, in addition to their direct effects on productivity, IT and ORG may be parts of a broader system of technologies and practices that the financial markets value but that is costly to implement. Even if this system does not appear on firms' balance sheets, it can be at least partly identified by the simultaneous presence of high levels of two observable assets: IT and ORG.

On the other hand, it is difficult to reconcile this finding with the hypothesis that the results are a manifestation of a bubble or similar persistent mispricing by investors. That would require not only that a persistent mispricing be correlated with IT and with ORG, but also that such a persistent mispricing be disproportionately severe for those firms specifically with high values of both variables simultaneously.

NONPARAMETRIC ESTIMATION. The above results show that each dollar of computer capital is associated with more intangible assets in high-ORG firms than in firms that invest less in human capital and the other organizational characteristics we identify. If the stock market is valuing these firms properly, this suggests that the benefits of computerization are likely to go disproportionately to firms that have adopted the organizational practices we identify.

Figure 1 captures this idea by plotting results from nonparametric regressions. The figure is a level plot of fitted values from a regression of market value on both the computer capital and the ORG variables, after netting out the influence of other variables. It presents a clear picture of the interaction effect between computer assets and ORG. Firms that are abundant in both computers and ORG have much higher market values than firms that have one without the other. Moreover, market valuation is disproportionately high in the quadrant where both asset levels are simultaneously above the median.

[FIGURE 1 OMITTED]

Intangible Assets and Productivity

One way to further distinguish the intangible asset hypothesis from alternatives such as investor mispricing is to look at production functions. If intangible assets really exist, they should be visible in the product markets as well as the financial markets. In particular, if indeed these assets are productive, firms possessing them should have higher future output. Previous work documented the importance of organizational assets in a production function framework and the existence of production complementarities between IT and organizational assets: firms with greater levels of ORG had both greater investments in IT and higher productivity of these investments. (72)

Using similar data on computers and output, but not considering the effects of organizational assets explicitly, Brynjolfsson and Hitt conducted a more detailed analysis of productivity growth and found that, as longer time periods were considered, the relationship between computer investment and productivity became stronger. (73) They found that the long-run productivity benefits are approximately five times the direct capital cost of computers, which would be consistent with a valuation of IT on the order of five times the valuation of ordinary capital. Both of these previous results support our assertion that the market value relationships we observe are consistent with the intangible assets hypothesis. However, neither of these studies considered the relationship among computer investment, organizational assets, and future output.

Table 8 reports production function estimates with IT and the interaction of IT with ORG lagged up to three periods. These regressions can be interpreted as identifying the relationship between current productivity and past investment in IT. Column 8-1 reports results from a regression that includes current-period IT, IT x ORG, and other production inputs (capital, labor) as well as time and industry controls (the latter at the 1 1/2-A-digit SIC level). (74) We find that both IT and ORG and the IT x ORG interaction make positive contributions to multifactor productivity in a manner similar to results found by Bresnahan, Brynjolfsson, and Hitt. (75) Moreover, the relationship among IT, ORG, and IT x ORG grows stronger as we consider effects on output further in the future. Although some of this rise is due to the change in the sample, a comparison with a current-period regression (not shown) using the same sample shows that the coefficients are similar at one- and two-year lags, and larger at three-year lags.

A natural interpretation of these results is that they reflect the net returns to a set of intangible assets, presumably some of the same intangible assets that investors are valuing. The fact that real output is higher for such firms suggests that the investors' beliefs have a reasonable economic basis.

Discussion and Conclusions

Our results suggest that the organizational complements to firms' installed computer capital are treated by investors as intangible assets. The financial markets treat the organizational assets associated with IT much like other assets that increase long-term output and profits. By covering several hundred firms over a period of eleven years, the analysis helps to document and explain the extent to which computerization is associated with both direct and indirect measures of intangible assets. Furthermore, this approach helps reveal the pattern of interactions among IT, organizational practices, and market valuations. If these assets are in fact becoming more important in modern economies, in part because of the information revolution engendered by computers and telecommunications, it is incumbent upon us to understand not only particular cases, but also any broader relationships and patterns that exist in the data.

Our main results are consistent with each of the testable implications about complementarities between computers and organizational design described in the introduction:

--The financial markets put a higher value on firms with more installed computer capital. The increase in market value associated with each dollar of IT substantially exceeds the valuation placed on other types of capital.

--Computer-intensive firms tend to have measurably different organizational characteristics, involving teams, more broadly defined jobs, and greater decentralization of certain types of decisionmaking.

--Firms with these organizational characteristics have higher market valuations than their competitors, even when all their other measured assets are the same.

--Firms with higher levels of both computer investment and these organizational characteristics have disproportionately higher market valuations than firms that invest heavily on only one or the other dimension.

--Firms with higher levels of IT, these organizational characteristics, or both have higher measured productivity in subsequent years.

Taken together, these results provide evidence that the combination of computers and organizational structures creates more value than the simple sum of their separate contributions. The evidence is consistent with the perception among many managers that IT is a catalyst for a broad set of organizational changes. It is also consistent with the econometric evidence that IT is more productive when combined with the specific cluster of organizational practices that we measure.

Our interpretation has focused on the assumption that the stock market is approximately correct in the way it values IT and other capital investments. Although econometrics alone cannot rule out all forms of potential omitted-variable bias, the evidence is not particularly consistent with alternative explanations such as a stock market bubble during the 1987-97 period. The fact that our results apply to a broad cross section of the economy over the ups and downs of a full business cycle suggests that fads, industry idiosyncrasies, and investor errors are not solely driving the results. Moreover, year-by-year estimation showed a consistently high valuation of computer capital throughout the 1987-97 period, although there is some evidence that valuations increased at the very end of the sample. Our analysis also predates the large increase in the market value of technology stocks in the late 1990s, and our sample is primarily composed of large, established firms (from the Fortune 1000) rather than new high-technology entrants. Indeed, we explicitly excluded information technology-producing firms because the role of computers in software or computer services firms is clearly different from the role they play in the computer-using sectors. For this reason also, our results are less likely to be sensitive to a high-technology stock bubble.

The fact that these intangible assets that are correlated with computers also appear to affect real production supports the hypothesis that the intangible assets are important. We find that our organizational assets variable is also correlated with increased subsequent output. These results appear to vindicate investors' beliefs that these assets are valuable. Nonetheless, it is also possible that some degree of mispricing has crept into the estimates. Furthermore, our results do not necessarily apply to other time periods or other types of firms or assets.

A particularly interesting finding is the significance of the interaction terms in our market value regressions. This result is difficult to explain in the traditional framework where the market values associated with assets simply reflect their shadow costs. Instead, it seems likely that we have identified a cluster of firms that, through skill, foresight, or luck, are endowed with a valuable form of business organization. Since the premium they command does not seem to have been quickly dissipated by competition, it must be difficult to imitate the specific technologies and practices they have implemented. Simply buying a lot of computers and implementing a stock set of practices such as we identify in our ORG variable is presumably not enough. Although both these types of assets are correlated with success as a business organization, much of the iceberg remains unidentified, and there is much that managers do not yet fully understand or, at least, cannot easily implement.

Our results give some support to one view of the productivity slowdown after 1973 and its revival in the 1990s. As several other researchers have argued, (76) the productivity slowdown may be explained by the changes in the value of assets and reorganization accompanying the transformation from a capital-intensive industrial economy to a computer-intensive, information-based economy. This view is reinforced by Robert Hall's interpretation of investment in reorganization. (77) In particular, the evidence is consistent with the view that certain work practices are complementary investments to computers, and the overall economic impact of these practices is being recognized by both managers and investors.

APPENDIX

Data Sources and Description

THE VARIABLES VSED for this analysis were constructed as follows.

Computer Assets

Direct measurements of the current market value of computer equipment by firm are from Computer Intelligence Corp. Computer Intelligence constructs a table of market values for each model of computer and uses this to calculate the current market value (replacement cost) of all computers owned by each firm.

PP&E

We considered two options to construct the physical capital variable, using data from Standard & Poor's Compustat Annual Dataset. The first, following the method in B. Hall (1990), takes the gross book value of physical capital stock (Compustat item 7: total gross PP&E) and deflates it by the GDP implicit price deflator for fixed investment. The deflator can be applied at the calculated average age of the capital stock, based on the three-year average of the ratio of total accumulated depreciation (calculated from Compustat item 8: total net PP&E) to current depreciation (Compustat item 14: depreciation and amortization).

The other, simpler method uses the net physical stock depreciation (calculated from Compustat item 8). According to the productivity literature, the first method should be used, but to conduct the market value estimation we adopted the second approach to ensure consistency with market value and other assets, which are measured in current dollars. The dollar value of IT capital (as calculated above) was subtracted from this result.

Other Assets

The "other assets" variable was constructed as total assets (Compustat annual data item 6) minus physical capital, as constructed above. This item includes receivables, inventories, cash, and other accounting assets such as goodwill reported by companies.

Market Value

Market value was calculated as the value of common stock at the end of the fiscal year plus the value of preferred stock plus total debt. In Compustat mnemonic code, this is MKVALF + PSTK + DT, which represents the total worth of a firm as assessed by the financial markets.

R&D-to-Assets Ratio

The R&D-to-assets ratio was constructed from R&D expenses (Compustat annual item 46). Interestingly, this item includes software expenses and amortization of software investment. R&D stock was constructed using the same rule as in B. Hall (1993a, 1993b). Fewer than half of the firms in our sample reported R&D expenses. The missing values were filled in using the averages for the same SIC four-digit industry.

Advertising-to-Assets Ratio

We constructed the advertising-to-assets ratio from advertising expenses (Compustat annual item 45). Fewer than 20 percent of our sample of firms reported the item. We applied the same rule as for the R&D-to-assets ratio.

Organization Variable

We constructed the ORG variable from items from a survey conducted in 1995 and 1996. The construction procedure using principal components analysis is described in the text. This variable captures the degree to which firms have adopted new organizational practices as identified by Osterman (1994), MacDuffie (1995), and Huselid (1995).

Value Added

This variable was calculated as constant-dollar sales less constant-dollar materials. Sales (Compustat annual item 12) were deflated by price indexes for each two-digit industry (from Gross Output and Related Series published by the Bureau of Economic Analysis). When an industry deflator was not available, we used the sector-level producer price index for intermediate materials, supplies, and components (from the 1996 Economic Report of the President). The value of materials was calculated by subtracting undeflated labor expenses (calculated above) from total expense and then deflating by the industry-level output deflator. Total expense was computed as the difference between operating income before depreciation (Compustat annual item 13) and net sales (Compustat annual item 12).

Ordinary Capital

This variable was computed from the total book value of capital (equipment, structures, and all other capital) following the method in B. Hall (1990). The gross book value of the capital stock (Compustat annual item 7) was deflated by the GDP implicit price deflator for fixed investment. The deflator was applied at the calculated average age of the capital stock, based on the three-year average of the ratio of total accumulated depreciation (calculated from Compustat annual item 8) to current depreciation (Compustat annual item 14). This calculation of average age differs slightly from the method in B. Hall (1990), who made a further adjustment for current depreciation. The constant-dollar value of IT capital (as calculated above) was subtracted from this result. Thus the sum of ordinary capital and IT capital equals the total capital stock.

Labor Expense

Labor expense was either taken directly from Compustat (annual item 42) or, when this was not available, calculated as a sector-average labor cost per employee multiplied by total employees (Compustat annual item 29). The average labor expense per employee was taken from BLS data on the hourly cost (including benefits) of workers for ten sectors of the economy. For firms whose labor expense as directly reported by Compustat did not include benefits, we adjusted the labor figure by multiplying the reported labor expense by the ratio of total compensation to wages for its sector as reported by the BLS.

1 1/2-Digit Industry Controls

The industry controls used in some of our analyses correspond to an intermediate level between one- and two-digit Standard Industrial Classification (SIC) codes. Based on the firm's reported primary SIC code on Compustat, we constructed the following control variables: mining and construction, SIC 11 to 20; process manufacturing, SIC 26, 28, and 29; other nondurable manufacturing, SIC 20 to 23 and 27; high-technology manufacturing, SIC 36 to 38 and 3571 (computers); other durable manufacturing, SIC 24 to 25, 30 to 35 (except 3571), and 39; transportation, SIC 40 to 47; utilities, SIC 48 to 49; trade, SIC 50 to 59; finance, SIC 60 to 69; and other services, SIC 70 to 79.

Table 1. Regressions of Market Value on Asset Quantities, 1987-97 (a)

 Estimation method

Independent variable OLS LAD

Computer assets 11.947 11.882
 (4.025) (0.361)
PP&E 1.474 1.181
 (0.088) (0.004)
Other assets (b) 1.064 1.039
 (0.012) (0.001)
[R.sup.2] 0.950 n.a.

Source: Authors' regressions. See appendix for description of data
sources.

(a.) The dependent variable is firms' market value. The sample contains
7,564 observations from 1,216 firms. All regressions control for year,
the ratio of R&D to sales, the ratio of advertising expenditure to
sales, and SIC industry; the OLS regression includes two-digit
controls, while the LAD regression includes 1 1/2-digit controls.
Standard errors are reported in parentheses; for the OLS regression
they are Huber-White robust standard errors.

(b.) Includes accounts receivable, inventories, and noncash liquid
assets.
Table 2. Long-Difference OLS Regressions of Market Value on Asset
Quantities, 1987-97 (a)

 Number of years differenced

Independent variable 1 2 3

Computer assets -0.020 -1.053 3.276
 (1.214) (1.748) (2.687)
PP&E 1.101 1.305 1.413
 (0.111) (0.120) (0.139)
Other assets 1.096 1.104 1.101
 (0.016) (0.012) (0.012)

Summary statistic
No. of observations 6,218 5,186 4,325
[R.sup.2] 0.789 0.829 0.861

 Number of years differenced

Independent variable 4 5 6

Computer assets 8.396 10.435 9.525
 (3.895) (4.279) (4.504)
PP&E 1.547 1.709 1.895
 (0.159) (0.181) (0.212)
Other assets 1.095 1.096 1.103
 (0.011) (0.011) (0.012)

Summary statistic
No. of observations 3,645 3,015 2,417
[R.sup.2] 0.898 0.921 0.928

 Number of years differenced

Independent variable 7 8 9 10

Computer assets 10.986 11.006 11.452 18.211
 (5.443) (6.185) (8.586) (16.331)
PP&E 1.971 2.043 2.026 2.072
 (0.240) (0.274) (0.275) (0.287)
Other assets 1.114 1.124 1.140 1.156
 (0.013) (0.016) (0.018) (0.020)

Summary statistic
No. of observations 1,857 1,325 847 413
[R.sup.2] 0.922 0.910 0.906 0.907

Source: Authors' regressions. See appendix for description of data
sources.

(a.) The dependent variable is firms' market value. All regressions
include controls for year and two-digit SIC industry. Huber-White
robust standard errors are reported in parentheses.
Table 3. Year-by-Year OLS Regressions of Market Value on Asset
Quantities, 1987-97 (a)

Independent variable 1987 1988 1989 1990

Computer assets 19.156 14.329 10.696 9.175
 (5.936) (4.585) (5.064) (8.001)
PP&E 1.018 1.109 1.283 1.286
 (0.063) (0.057) (0.068) (0.079)
Other assets 0.979 0.979 1.008 0.979
 (0.008) (0.009) (0.018) (0.016)

Summary statistic
No. of observations 651 646 641 641
[R.sup.2] 0.974 0.981 0.964 0.961

Independent variable 1991 1992 1993 1994

Computer assets 11.989 12.718 16.264 21.798
 (10.218) (9.556) (6.744) (7.154)
PP&E 1.389 1.449 1.547 1.457
 (0.080) (0.091) (0.095) (0.093)
Other assets 1.020 1.040 1.027 1.012
 (0.026) (0.021) (0.016) (0.014)

Summary statistic
No. of observations 666 676 702 706
[R.sup.2] 0.928 0.952 0.973 0.974

Independent variable 1995 1996 1997

Computer assets 9.450 18.809 30.038
 (3.804) (6.692) (13.612)
PP&E 1.581 1.667 1.822
 (0.123) (0.155) (0.207)
Other assets 1.066 1.081 1.114
 (0.014) (0.011) (0.012)

Summary statistic
No. of observations 747 742 746
[R.sup.2] 0.956 0.949 0.942

Source: Authors' regressions. See appendix for description of data
sources.

(a.) The dependent variable is firms' market value. All regressions
include controls for the ratio of R&D to sales, the ratio of
advertising spending to sales, and two-digit SIC industry. Huber-White
robust standard errors are reported in parentheses.
Table 4. Correlations between Organizational Structure and Alternative
IT Measures (a)

 Measure of IT investment

 Installed
Measure of organizational structure IT base

Structural decentralization (b)
Degree of team self-management 0.17 ***
Employee involvement groups 0.07
Diversity of job responsibilities 0.07

Individual decentralization (b)
Who determines pace of work 0.04
Who determines method of work 0.16 ***
Degree of individual control 0.11 *

Team incentives (c)
Degree of team building 0.15 ***
Workers promoted for teamwork 0.02

Skill acquisition
Percent of workers that receive
 off-the-job training 0.14 **
Degree of screening prospective
 hires for education (d) 0.16 ***
ORG composite variable 0.24 ***

 Measure of IT investment

 Total central
 processing
Measure of organizational structure power

Structural decentralization (b)
Degree of team self-management 0.22 ***
Employee involvement groups 0.08
Diversity of job responsibilities 0.12 **

Individual decentralization (b)
Who determines pace of work 0.06
Who determines method of work 0.20 ***
Degree of individual control 0.15 **

Team incentives (c)
Degree of team building 0.19 ***
Workers promoted for teamwork 0.10 *

Skill acquisition
Percent of workers that receive
 off-the-job training 0.15 ***
Degree of screening prospective
 hires for education (d) 0.18 ***
ORG composite variable 0.30 ***

 Measure of IT investment

 No. of
 personal
Measure of organizational structure computers

Structural decentralization (b)
Degree of team self-management 0.20 ***
Employee involvement groups 0.08
Diversity of job responsibilities 0.10 *

Individual decentralization (b)
Who determines pace of work 0.02
Who determines method of work 0.15 ***
Degree of individual control 0.15 **

Team incentives (c)
Degree of team building 0.18 ***
Workers promoted for teamwork 0.00

Skill acquisition
Percent of workers that receive
 off-the-job training 0.14 **
Degree of screening prospective
 hires for education (d) 0.21 ***
ORG composite variable 0.25 ***

Source: Authors' calculations. See appendix for description of data
sources.

(a.) Table reports Spearman partial rank-order correlations,
controlling for firm size, worker occupation, and 1 1/2-digit SIC
industry. Sample size varies from 300 to 372, depending on data
availability. * indicates significance at the 10 percent level, ** at
the 5 percent level, and *** at the 1 percent level.

(b.) Scale of 1 to 5, where 5 represents greatest decentralization
(most extensive use of self-managing teams or employee involvement
groups, greatest diversity of job responsibilities, greatest worker
control of pace or method of work, or greatest degree of individual
control).

(c.) Scale of 1 to 5, where 5 represents greatest use of team
incentives (projects requiring team efforts, or promotion of workers
for teamwork).

(d.) Scale of 1 to 5, where 5 indicates most extensive use of
screening.
Table 5. Unrotated Principal Components Used to Construct the ORG
Variable (a)

 Loading on first Loading on second
 principal principal
Work practices component component

Degree of team self-management 0.751 0.006
Employee involvement groups 0.707 0.176
Who determines pace of work 0.528 -0.628
Who determines method of work 0.572 -0.456
Degree of team building 0.747 0.250
Workers promoted for teamwork 0.401 0.367
Percent of workers who receive
 off-the-job training 0.425 0.408
Degree of screening prospective
 hires for education 0.466 -0.095
Memorandum: percent of
 variance explained 34.8 12.6

Source: Authors' calculations. See appendix for description of data
sources.

(a.) See table 4 for definitions and scale measures.
Table 6. OLSD Regressions of Computer Assets, PP&E, and Other Assets
on Organizational Assets, 1987-97 (a)

 Dependent variable

 Computer
Independent variable assets PP&E Other assets

ORG 0.179 *** 0.033 0.009
 (0.055) (0.050) (0.048)
Employment 0.736 *** 0.801 *** 0.800 ***
 (0.102) (0.091) (0.084)
[R.sup.2] 0.470 0.817 0.803

Source: Authors' regressions. See appendix for description of data
sources.

(a.) All variables except ORG are expressed in natural logarithms. The
sample contains 2,444 observations. All regressions include controls
for year and two-digit SIC industry. Huber-White robust standard errors
are reported in parentheses. *** indicates significance at the 1
percent level.
Table 7. Regressions of Market Value on Organizational Assets and
Related Interaction Terms, 1987-97 (a)

 Estimation method

 OLS

Independent variable 7-1 7-2 7-3

Computer assets 14.593 11.452 5.387
 (6.094) (5.116) (5.409)
PP&E 1.313 1.318 1.313
 (0.079) (0.078) (0.079)
Other assets 1.066 1.067 1.075
 (0.023) (0.023) (0.017)
ORG x employment 18.939 11.315
 (8.431) (7.661)
ORG x computer assets 13.089
 (6.443)
ORG x PP&E 0.016
 (0.059)
ORG x other assets -0.029
 (0.031)
[R.sup.2] 0.928 0.928 0.929

 Estimation method

 Random effects

Independent variable 7-4 7-5

Computer assets 9.949 2.229
 (2.566) (3.220)
PP&E 1.564 1.527
 (0.064) (0.070)
Other assets 1.085 1.086
 (0.009) (0.010)
ORG x employment 8.880
 (8.451)
ORG x computer assets 13.352
 (3.654)
ORG x PP&E 0.060
 (0.065)
ORG x other assets 0.026
 (0.017)
[R.sup.2] n.a. n.a.

 Estimation method

 LAD Fixed effects
Independent variable 7-6 7-7

Computer assets 9.742 1.888
 (0.768) (3.322)
PP&E 1.103 1.661
 (0.008) (0.089)
Other assets 1.023 1.087
 (0.002) (0.010)
ORG x employment -0.934 5.262
 (0.945) (12.085)
ORG x computer assets 4.264 12.596
 (0.894) (3.745)
ORG x PP&E 0.026 0.163
 (0.007) (0.095)
ORG x other assets -0.024 0.049
 (0.003) (0.019)
[R.sup.2] n.a. 0.887

Source: Authors' regressions. See appendix for description of data
sources.

(a.) The dependent variable is firms' market value. The sample contains
2,444 observations from 272 firms with reported organizational assets.
All regressions include controls for year, the ratio of R&D to sales,
the ratio of advertising spending to sales, and two-digit SIC industry.
Standard errors are reported in parentheses; for OLS regressions they
are Huber-White robust standard errors.
Table 8. OLS Regressions of Current Production on Past Computer
and Organizational Assets, 1987-2000 (a)

 Dependent variable

Independent variable t t + 1 t + 2 t + 3

Computer [assets.sub.t-1] 0.045 0.047 0.043 0.045
 (0.020) (0.025) (0.029) (0.036)
ORG x computer [assets.sub.t] 0.024 0.020 0.030 0.058
 (0.014) (0.014) (0.015) (0.019)
ORG 0.004 0.010 0.014 0.026
 (0.022) (0.022) (0.023) (0.026)
[Capital.sub.t] (b) 0.255 0.248 0.253 0.244
 (0.023) (0.025) (0.025) (0.028)
[Labor.sub.t] (c) 0.624 0.635 0.632 0.622
 (0.035) (0.039) (0.049) (0.060)

Summary statistic
No. of observations 2,091 1,605 1,182 791
[R.sup.2] 0.835 0.836 0.849 0.839

Source: Authors' regressions. See appendix for description of data
sources.

(a.) The dependent variable is value added in constant dollars. All
variables except ORG are expressed in natural logarithms. The sample
is restricted to 272 firms that have reported organizational assets.
All regressions include controls for year and 1 1/2-digit SIC industry.
Huber-White robust standard errors are reported in parentheses.

(b.) Ordinary capital stock in constant dollars, computed by the method
described in B. Hall (1990). See appendix for details.

(c.) Labor expense in constant dollars. See appendix for details.

Comments and Discussion

Martin Neil Baily: This paper is the latest chapter in a series of important papers, by Erik Brynjolfsson, Lorin Hitt, and Shinkyu Yang and other coauthors, exploring the relationship between computers, on the one hand, and economic performance (as measured by productivity and firm market value), on the other. The basic message of this body of work is that computers are very important, but that the success of companies that use them depends on how they use them and on what other aspects of the production process have changed.

This is a good story. It may very well be the right story, and it is a very different story from the one that we heard in the 1980s, when productivity growth was weak, and everybody was wondering why computers were not contributing to the economy. For example, a student at MIT's Sloan School of Management, Gary Loveman, found that computer capital had had little or no effect on productivity. (1) In production function estimates, computer capital had a coefficient smaller than those on other kinds of capital.

Despite my sympathy with the overall story, I will raise some concerns about the results in this paper. The paper comes at a time when economists, business decisionmakers, and policymakers are reassessing the impact of information technology and its potential to fuel future economic growth and increases in equity valuation. After the computer paradox of the 1980s, there was an extravagant enthusiasm for IT in the 1990s, which in turn was followed by a technology bust, led by the collapse of the dot-coms and followed by the collapse of spending on computers and communications equipment.

The basic motivation of the paper, presented in the authors' table 1, is

their earlier finding that $1 worth of computer capital appears to contribute $15 or so to a company's market value. As the authors say, however, this cannot be a structural result. Since computers are not rationed, this coefficient, if taken literally, implies that companies could boost their market value enormously just by buying computers. In the late 1990s some companies may have actually thought that buying computers would boost their market value, because indeed they bought them like there was no tomorrow. But more seriously, this is not a sensible implication and motivates attempts to explain what must be a nonstructural coefficient. Computer capital must be proxying for another variable, an unobserved characteristic of the companies.

One possibility is that the market did not provide a rational evaluation of the companies in this sample. And the market's overvaluation may have been associated with companies that were doing interesting things that involved buying a lot of computers. Computer capital may have been a proxy for irrational investor enthusiasm for the earnings growth potential of some companies. The authors argue that this cannot be the case because their data end in 1997, whereas the technology bubble mostly occurred after 1997. But the possibility of irrationality in the market was there even earlier. Federal Reserve Chairman Alan Greenspan warned of irrational exuberance when the Dow Jones Industrial Average was only at 6,000. It would be surprising if irrational valuation could explain the authors' whole set of results, but it may explain part of the puzzle. I wonder how computer capital would show up in a similar regression on company valuation today.

The authors make the case that the missing variable is organizational capital. This ties in with their overall conclusion that it is not the computers themselves but what you do with them that counts. The computer is a building block of a much larger asset of the firm, namely, the whole organizational structure that it uses to produce and market its products or services. The authors therefore construct a measure of organizational capital based on observed characteristics of the companies in their sample.

Simply adding this variable to the valuation regression lowers the coefficient on computer capital from 14.6 to 11.5 (see their table 7, columns 7-1 and 7-2). But that is not a large change. The authors argue that the key to further reducing the anomalously large coefficient on computer capital is to include an interaction term: organizational capital times computer capital. When this term is added, the coefficient on computer capital moves into a range from around 2 to 10, depending on other aspects of the regression. That is a big improvement and an encouraging sign for their hypothesis. But it still leaves something of a puzzle and a concern about the interpretation of the results.

For one thing, the puzzle of the too-large coefficient on computer assets also extends to other types of capital. (2) The coefficient on PP&E assets is consistently and significantly above 1, and a company's stock of such capital is likely to far exceed that of its computer capital and may have a much larger impact on its overall market value. This makes me uneasy about the authors' interpretation. Until we know what is driving the anomalies in capital valuation, we cannot be sure that the true structural relations have been captured and that computers and their related organizational capital are as important as the authors conclude.

One possibility is that the causality is going the other way: Those companies that the market values highly may, for other reasons, buy a lot of computers, and may indeed buy a lot of other capital as well. I have argued elsewhere that some reverse causality may be at work in the productivity literature. The standard growth accounting framework attributes a very large fraction of the acceleration of productivity in the late 1990s to investment in IT. But the fact that the surge in productivity coincided with the surge in IT investment does not prove causality. In the late 1990s, buying IT became the thing for companies to do, and the surge in output, cash flow, and productivity encouraged them to do it. Within the context of this paper, those companies with high market valuations presumably found it very easy to issue equity, float bonds, and raise funds. The cost of capital was very low, and the companies bought a lot of computers and other capital, believing that this would pay off.

In the authors' defense, they do find that their results hold up in regressions from 1987 and 1988, when buying computers was not such a big deal. Nevertheless, I suspect that a bit of reverse causality may be taking place. In the authors' table 3, which presents year-by-year regressions, the coefficient on computer capital averages 13.3 over 1987-90, compared with 19.3 over 1993-97. The coefficient on PP&E rises almost monotonically over the period.

Another possibility is that there remain one or more omitted variables in the regression even after the organizational capital has been accounted for. The companies may have been doing something that showed up in both market value and capital accumulation (computer capital and otherwise). Perhaps the omitted variable is just business opportunities or innovations created at certain companies. The market picked up these companies' prospects for future earnings growth. The companies needed to make substantial investments to realize these opportunities, and they bought capital. Adding to computer capital was particularly likely given the nature of the new or expanded business opportunities that were appearing during this period. After all, computers have become part of the backbone of the economy.

To defend the authors once again, they test for this problem by running difference regressions to remove firm-specific effects. However, the results from that analysis are a bit mixed. They find (table 2) that the coefficient on computer capital falls to zero for one- and two-year differences (and the coefficient on PP&E falls closer to 1). As the time difference increases, however, the coefficients on computer capital and other capital rise way back up again. I do not know quite what to make of that. One could argue that, because of creative destruction, the potential market value of an innovation or a new business opportunity could be fairly ephemeral. A business opportunity comes along, market value goes up, the company buys a bunch of computers, and then competitors come in and the opportunity to generate excess returns is competed away. The fact that the computer capital coefficient comes back so strongly after such a short time difference makes that a somewhat harder story to tell. But the bottom line is that the first-difference equations do seem a red flag that part of what is picked up here is a firm-specific characteristic that happens to be correlated with computer buying.

I turn now to a different issue. The thrust of this paper is that computer capital is central to strong business performance, as shown by market value or by productivity. There is some literature that questions this hypothesis on the basis of findings from industry case studies--in particular, the work on IT and productivity by the McKinsey Global Institute (MGI). (3)

That report concludes that the surge of IT spending after 1995 was not the main reason for the surge in productivity growth. Productivity and profitability are not the same. It is possible that the companies that invested heavily in IT were able to gain a competitive advantage that raised their profits and market value, even though industry productivity was not greatly enhanced. But that is not what the MGI report finds. On the contrary, it finds that the companies that invested heavily in IT often achieved no better performance than those that failed to invest.

For example, the MGI case study of retailing suggested that a lot of what Wal-Mart did was not related directly to IT. Wal-Mart and other retailers made many operational improvements that did not depend on IT, notably the shift to big-box stores, which provided scale economies. In the retail sector more broadly, the increase in consumers' income over time allowed companies to increase the value added of the products they sell, and this increased their measured productivity.

Again in fairness to the authors of this paper, there are different perspectives on what is happening at the industry and the firm level, and at Wal-Mart in particular. Wal-Mart's is a very IT-intensive operation, and the company's organizational structure allows it to take advantage of the knowledge base the IT generates. But the MGI study, along with the work of some other New Economy skeptics, does raise a serious challenge. How central is IT to successful business performance? Some amount of IT is surely essential to function in a modern economy. But the jury is still out on whether IT is the key element for success, or an enabling technology that requires other changes and innovations and may or may not pay off, or indeed a distraction or hindrance as companies overinvested.

I would vote for the second proposition, and I assume the authors of this paper would, too. But I have been struck by conversations I have had with some consultants who serve IT vendors. In many or even a majority of cases, the expected performance benefits the vendors promised their customers have not yet materialized. The authors of this paper mention enterprise resource planning (ERP) software as an example of how the total cost of installing a complete IT system greatly exceeds the cost of the hardware alone. Indeed, I have heard stories about customers who signed up for ERP software and ended up paying and paying and never really achieving the promised cost savings. There is a joke in the software industry that installing certain ERP software is like pouring wet cement into a business' operations. The cement hardens and destroys flexibility and innovation.

To conclude, the authors are going down the right road. They are asking how IT interacts with operational factors and intangible capital, and what combination of these things enhances market value. I entirely agree with their strategy for future research: they just need to go farther down that road, to increase our confidence that they are really capturing a structural relation between computers and market valuation.

Robert E. Hall: A theme of recent work on wage differentials is that workers who know how to use computers earn a lot more than those who do not. This interesting paper by Brynjolfsson, Hitt, and Yang illustrates the same principle for firms: those that know how to harness modern information technology within their business models are worth a lot more than those that do not. Although this finding itself is not surprising, the magnitude is: a dollar of extra investment in IT brings a $15 increase in the investing firm's stock market value. The market infers the firm's use of' modern, efficient business processes from the presence of extra IT investments on its balance sheet. The authors argue strenuously--and successfully in my view--that this inference is rational and not the result of a bubble in the market for shares of computer-intensive firms. The authors also make it clear that they do not see the high marginal valuation of IT capital as a failure of the principle that a firm should continue to invest in all forms of capital until the marginal dollar of investment adds just a dollar to market value. Rather, the spectacular coefficient of 15 on IT capital is an iceberg ratio: for every dollar of visible capital, there is $14 worth of invisible investments in business processes.

The divergent experiences of the two biggest players in big-box retailing--Wal-Mart and Kmart--are the perfect example of the paper's theme. The stock market values Wal-Mart, the leader in modern computerized retailing, at six times the book value of its assets. Yet Kmart, with virtually the same retailing model, is in bankruptcy. Thanks to the advanced IT systems Wal-Mart has implemented, a manager at Wal-Mart's headquarters knows vastly more about what is on the shelves of a Wal-Mart store in another part of the country than a local Kmart manager knows about what is on the shelves of his or her own store. Wal-Mart's systems make possible a completely Stalinesque organization involving an astonishing 1.4 million employees. It is wonderfully ironic that the dreams of the Soviet central planners have come tree--thanks to modern information technology--in a worldwide empire run from a small town in Arkansas rather than from Moscow.

As the authors stress, their paper is about Wal-Mart and the other users of modern technology, not about the producers of that technology. Their studies exclude the technology-producing companies altogether. It is worth recording, however, that the foundation of modern IT is a pair of general-purpose technologies. One is the generic server--basically a souped-up personal computer. Because servers are the product of a competitive computer hardware industry, their prices have fallen to essentially zero by comparison with computer prices of a decade or so ago. The other technology is database software: generic software that can keep track of practically infinite information about a practical infinity of objects. Wal-Mart could never manage its empire without the capabilities of modern database management.

The basic finding that motivates this paper--the extremely high value that the stock market places on an incremental dollar of IT investment--has been reported in a series of earlier papers by Brynjolfsson and his coauthors. The starting point is Martin Baily's value revelation principle: that the stock market (or, more accurately, the totality of financial claims) should measure the value of the real assets of a firm. Absent a correlation between the unmeasured components of a firm's real portfolio and the measured components, each measured component should receive a value of 1 in a regression of total value on the components. When the observed coefficient is different from 1, we conclude that there is a correlation with an unmeasured component. The big news from the earlier papers was the extreme size of that unmeasured component.

The value added by this paper is in bringing in new information that bears directly on the unmeasured component. This information relates to organizational design: use of teams and team-based incentives, individual decisionmaking authority, and investment in skills and education. The authors blend the information into a single measure, which they call ORG. This measure is standardized to have a zero mean and unit variance. Thus it lacks the appropriate conceptual basis for inclusion in the regressions of market value on value components, because it is not a value at all. Data are not available to measure the cost of firms' investments in organizational design or even the investments that such designs call for, such as training. All of these costs are expensed along with normal operating costs and cannot be recovered from any standard financial accounts.

The paper uses ORG to explain the surprising findings of the basic value regressions. Interpretation of these regressions requires attention to the nature of the ORG variable and to the underlying framework of the regression. First, the left-hand-side variable is the value of the firm in millions of dollars. All of the right-hand-side variables need to be scaled suitably to take account of the huge range in values. The capital value variables of the basic regression are so scaled--they, too, are in millions of dollars--and the authors scale ORG by the number of employees. The addition of the scaled ORG variable lowers the coefficient on IT from 15 to 11, suggesting that some of the correlation of value with IT is better captured by ORG.

The authors' tables 7 and 8 include versions of ORG scaled by various types of capital. One could place these results into the basic framework by hypothesizing that each firm's organizational investment per dollar of IT investment is proportional to its ORG value. The coefficient on IT drops to 5 when the various proxies for organizational capital are included. In this interpretation, about two-thirds of the iceberg portion of the IT capital coefficient is attributable to the fact that ORG scaled by various types of capital serves as a proxy for organizational capital. That is an impressive finding.

The authors, however, do not rest with this interpretation of their findings. Instead, they go on to write, "This finding is consistent with the case literature on the complementarities between IT and organizational structure." I think this is a misunderstanding of the basic framework. Baily's value revelation principle applies to capital inputs irrespective of how they enter the production function. A competitive, optimizing firm has a market value equal to the sum of the values of its different kinds of capital, whether those kinds of capital are substitutes, complements, or something in between. No conclusion about complementarities can follow from the approach taken in this paper--to do so would require estimation of the technology, not a study of value.

The results show that ORG does not capture some of the important features of what I suspect accounts for the extra $14 or $15 valuation of a dollar of IT investment. In particular, it is not clear to me that Wal-Mart is a high-ORG firm or that the principles that go into Wal-Mart's stunning success have much to do with the factors measured by ORG. Here I will take advantage of the fact that I am one of the few economists interested in this subject who shops extensively at Wal-Mart. The success of Wal-Mart, I believe, comes from two main factors. First, its logistical performance is phenomenal. Its shelves are always well stocked. Even at the height of the busiest seasons, what you want is on the shelf. This performance is directly related to the centralized database system that lies at the heart of the Wal-Mart business process. Second, Wal-Mart has mastered the art of inducing ordinary people to do their jobs well. I believe much of this is also centralized: Wal-Mart has developed managerial procedures for employee accountability that the company applies relentlessly in all its stores. As a result, whereas Kmart's shelves are disorganized and bereft of the things people really want, and the employees are sullen and unhelpful, Wal-Mart has figured out how to make 1.4 million people work together to put what people want on the shelves and help them find it.

One of the interesting puzzles about modern business is the role of centralization and decentralization. Modern business processes based on IT have created amazingly successful large, centralized organizations. Wal-Mart is the leader, of course, but Dell Computer and Southwest Airlines are other prominent examples. On the other hand, variables measuring structural and individual decentralization are key ingredients of ORG and of discussions of successful modern firms. Success seems to require the adroit combination of the benefits of centralization of data management with decentralization of certain aspects of work.

It will come as no surprise that I am completely in agreement with the authors' conclusion that their valuation results are not just another reflection of a high-technology bubble in the stock market. To illustrate, my figure 1 shows the history of Wal-Mart's share price since 1987. The shaded area is the period most often considered that of a high-technology bubble, from the beginning of 1999 through March 2000. Wal-Mart did enjoy a sharp appreciation during the high-technology boom, but as the market sorted out the winners and losers, Wal-Mart plainly emerged as a winner. As of April 2002, the company enjoyed a stock market value quite close to its transitory peak during the high-tech frenzy of 1999.

[FIGURE 1 OMITTED]

The authors are candid in stating that they see this paper as only a first installment on the fascinating topic of understanding and explaining the iceberg of IT valuation. I hope that ultimately they are able to develop good direct measures of other types of capital, which, when included in their value regressions, will drive the coefficient on IT capital down to its proper value of 1.

General discussion: Several panel members expressed skepticism about the authors' interpretation of the large coefficients on computer assets and on the term interacting computer assets with organizational capital in their valuation equations. The authors claimed that these results show that computer capital is associated with significant intangible investment in organizational capital. Christopher Sims observed that these right-hand-side variables could just as well be regarded as dependent variables, "explained" by market valuation. For example, the authors' estimates of the effects of computer assets could be interpreted as saying that firms valued highly by the market, for any of a wide variety of reasons, take advantage of the resulting low cost of capital to purchase computers. Kevin Murphy concurred, remarking that the authors' analysis was like running a regression on house prices in which the number of bathrooms, the number of bedrooms, and so on are included as explanatory variables, and finding that the coefficient on a Jacuzzi variable was $120,000. No one would interpret that coefficient as implying that a Jacuzzi adds that much to a house's value.

Sims noted that the interpretation of the coefficient on the interacted variable, ORG x computer assets, is subject to the same problems as that of computer assets alone. This variable might be picking up a factor that influences how investments are financed, rather than any effect of technology. For example, a firm's ability to secure bank loans might be compromised if the firm is making unobservable investments in training and organizational restructuring, which bankers do not readily accept as collateral, as opposed to investing in traditional physical capital. If this were the case, the interacted variable might reflect the degree to which a firm had financed itself with equity capital instead of debt. The extraordinary performance of the stock market during the recent period would then result in large coefficient estimates. Sims further noted that if the skills embodied in the organization variable are scarce, the consultants and managers who have those skills--not the equity holders--should be collecting the rents. He commented that the contemporaneous nature of the regressions makes it difficult to identify causality, and increases in equity value may actually precede investment in computers and organization. He acknowledged that it may not be practical to specify and estimate a model that includes equations for all the endogenous variables, or to find satisfactory instruments for computer investment. But he thought there should be more discussion of the potential biases in the results and of alternative interpretations.

William Brainard observed that coefficients on physical capital significantly above 1 call into question the underlying assumption of the analysis, namely, that market values accurately reflect the value of a firm's constituent parts. He noted that in the authors' year-by-year regressions the coefficient estimates on physical plant and equipment increased systematically over time, from about 1.0 in 1987 to more than 1.8 in 1997. The coefficient on PP&E is approximately 1.5 in the pooled regressions, which include time dummies. He noted that the dramatic rise in overall market q during the period made it essential to distinguish the valuation effects of cross-sectional variation from those of time variation in computer assets. Brainard was not satisfied with the authors' attempt to make this distinction by using time dummies in pooled regressions. With the regressions estimated in levels, increases in overall market value during the period show up as the same dollar addition to the value of every firm independent of the size of the firm. Hence the growth in overall market q will be embedded in the coefficients on PP&E and on computer assets. The coefficients cannot be interpreted as arising only from the cross-sectional variation in asset composition. Nor can one say how much of the coefficient on computer assets simply reflects their growing importance in the later periods when market valuations in general were high.

Several panelists discussed the quantity of intangible capital implied by the authors' estimates and interpretation. The authors reported the stock of this capital as being between $1.5 trillion and $2 trillion. William Nordhaus observed that these numbers are consistent with Dale Jorgenson's finding of no surge in total factor productivity once investment in information technology is taken into account. Both Nordhaus and Sims urged the authors to extend their analysis to include earnings as well as market valuations, noting that their results would be strengthened if the story told in the paper were reflected in earnings. Nordhaus noted that although one could view the results as showing that supernormal returns result from a synergistic interaction between computers and organization, if instead they reflect intangible assets with normal returns, as the authors believe, there should be a particular time pattern of earnings. Earnings should be lower at the beginning of a period of intangible investment and higher later, as the intangible capital accumulates. He suggested that the authors examine earnings to see whether this pattern emerges.

(1.) Loveman (1986).

(2.) William Brainard made this point in the general discussion of the paper.

(3.) McKinsey Global Institute (2001).

(1.) Jorgenson and Fraumeni (1995).

(2.) B. Hall (1993a); Griliches (1981); Lev and Sougiannis (1996).

(3.) Bresnahan, Brynjolfsson, and Hitt (2000, 2002).

(4.) Chan, Lakonishok, and Sougiannis (1999); R. Hall (2001b).

(5.) R. Hall (2000a); Brynjolfsson and Yang (1999).

(6.) On the impact of computers on organizational redesign see R. Hall (2000a), Brynjolfsson and Hitt (2000), Brynjolfsson, Renshaw, and Van Alstyne (1997), Black and Lynch (2001, forthcoming), and Milgrom and Roberts (1990); on their impact on skill mix see Autor, Levy, and Murnane (2000) and Bresnahan, Brynjolfsson, and Hitt (2000, 2002).

(7.) See Brynjolfsson and Hitt (2000) for a review.

(8.) David (1990).

(9.) Bresnahan, Brynjolfsson, and Hitt (2000); R. Hall (2000a); Sauer and Yetton (1997).

(10.) See, for example, Osterman (1994), Huselid (1995), Ichniowski, Shaw, and Prennushi (1997), and Bresnahan, Brynjolfsson, and Hitt (2000). This is a welcome change. As Alfred Sloan, former chief executive of General Motors (1964, p. 50), noted, "The principles of organization got more attention among us than they did then in universities. If what follows seems academic, I assure you that we did not think it so."

(11.) Barth, Elliott, and Finn (1997).

(12.) Kemerer and Sosa (1991) provide examples.

(13.) For example, Baily (1981).

(14.) See, for example, Milgrom and Roberts (1990) and Radner (1993).

(15.) Brynjolfsson and Mendelson (1993).

(16.) Bresnahan, Brynjolfsson, and Hitt (2000, p. 184).

(17.) See, for example, Applegate, Cash, and Mills (1988), Autor, Levy, and Murnane (2000), Barras (1990), Bresnahan and Greenstein (1997), Brynjolfsson, Renshaw, and Van Alstyne (1997), Davenport and Short (1990), David (1990), Hitt and Brynjolfsson (1997), Ito (1996), Malone and Rockart (1991), Milgrom and Roberts (1990), Orlikowski (1992), Scott Morton (1991), and Woerner (2001). Brynjolfsson and Hitt (2000) provide a review of the literature.

(18.) See, for example, Kemerer and Sosa (1991).

(19.) Ito (1996); Bresnahan and Greenstein (1997); Bresnahan (1999).

(20.) Brynjolfsson, Renshaw, and Van Alstyne (1997).

(21.) Brynjolfsson and Hitt (1995, 2001); Lichtenberg (1995).

(22.) See, for example, Ichniowski, Shaw, and Prennushi (1997).

(23.) Brynjolfsson and Yang (1999), R. Hall (2000a), and Lev (2001), respectively.

(24.) Some of the following examples, as well as the broader econometric studies, are described in more detail in Brynjolfsson and Hitt (2000).

(25.) Figures are as reported by Yahoo! Finance on March 29, 2002.

(26.) Martin Schrage, "Wal-Mart Trumps Moore's Law," Technology Review, March 2002, p. 21.

(27.) McKinsey Global Institute (2001, p. 2).

(28.) McKinsey Global Institute (2001, p. 4).

(29.) Martin Schrage, "Wal-Mart Trumps Moore's Law," Technology Review, March 2002, p. 21.

(30.) Brynjolfsson, Renshaw, and Van Alstyne (1997).

(31.) Woerner (2001).

(32.) Gormley and others (1998).

(33.) AICPA Statement of Position 98-1. This does not apply for "small" software purchases or development projects. Firms have some discretion as to the exact threshold: at FleetBoston Financial, for example, software development projects smaller than $500,000 are normally entirely expensed, according to Cherie Arruda, technology controller at FleetBoston (personal communication with the authors, March 22, 2002).

(34.) Specifically, "Examples of the kind of work that must be expensed include: Alternatives development and evaluation, development of requirements, training, data conversion, evaluation of technology, and choosing one of the alternatives being proposed" (Colenso, 2000, p. 3).

(35.) Before the adoption of Statement of Position 98-1, firms typically expensed even more of their internal software development costs. See, for example, the notes to the consolidated financial statements in Lucent Technologies' 2001 annual report (www.lucent.com/investor/ annual/01/pdf/notes.pdf). The exception is firms that produced software for sale, which began more aggressive capitalization in 1985 as a result of FASB Statement 86.

(36.) Lev and Zarowin (1999, p. 20).

(37). FASB Concept No. 6, 1985, paragraph 175.

(38.) See, for example, Tobin (1969), Hayashi (1982), Naik (1994), Yang (1994), Bond and Cummins (2000), and R. Hall (2000a, 2001a, 2001b).

(39.) Griliches (1981); Griliches and Cockburn (1988); B. Hall (1993a, 1993b, 1999).

(40.) Montgomery and Wernerfelt (1988).

(41.) R. Hall (200la, p. 1186).

(42.) R. Hall (2000a).

(43.) Interestingly, in related work using similar data on computers and organizational investments but a different framework, Bresnahan, Brynjolfsson, and Hitt (2000, 2002) also find empirical support for the second main claim of R. Hall (2000a) regarding the organizational drivers of the increased demand for college-educated workers.

(44.) Brynjolfsson and Yang (1999).

(45.) Baily (1981).

(46.) See, for example, Abel (1990).

(47.) Hayashi (1982).

(48.) Greene (1993, p. 246).

(49.) Bond and Cummins (2000).

(50.) See R. Hall (2001b) for one approach.

(51.) See the related discussion in Bresnahan, Brynjolfsson, and Hitt (2002).

(52.) To apply equation 7, replace K and M with their values conditional on [K.sub.o].

(53.) See, for example, Bresnahan, Brynjolfsson, and Hitt (2000) and Brynjolfsson and Hitt (2001).

(54.) Advertising and R&D are other types of nonstandard "assets" that have been considered in prior work. Because no capitalized value is reported for them, we simply include them as ratios in the reported regression. This can be thought of as treating current spending on these assets as a noisy indicator of their capital stock values (B. Hall, 1993a, 1993b; see also Brynjolfsson and Yang, 1999, for a more detailed analysis of these assets in this context). Because R&D is available only for about half the sample, and advertising for only about a third, we set the values of these variables to zero when they are missing and include a dummy variable to capture the mean contribution of these variables when the data are not available.

(55.) There was a change in the valuation methodology in the CII database in 1994. Thereafter the market value of central processors was no longer calculated at the equipment level. However, CII did continue to obtain the market value information going forward, and thus comparable measures could be constructed by multiplying the aggregate number of units (personal computers, mainframes, workstations, and so on) by the average value for the category. Year-by-year regressions do not suggest the presence of any structural change in the data.

(56.) Another potential source of error in this regard is the outsourcing of computer facilities. Fortunately, to the extent that the computers reside on the client site, they will still be properly counted by CII's census. To the extent that these facilities are located at a third-party site, they will not be properly counted.

(57.) Huselid (1995); Ichniowski, Shaw, and Prennushi (1997); Osterman (1994).

(58.) For most of the economy, computers are a complement to other production assets. However, in the computer and software industries, computers are the principal production asset. Moreover, because these firms often use the technology assets they produce, they may face very different effective prices for these assets. Communications industries were excluded because of the difficulty in separating out corporate computer use from telephone switchgear, which is largely computer based.

(59.) Brynjolfsson and Yang (1999).

(60.) An alternative specification would include the book value of some of these other assets in the computation of market value and remove them from the list of independent variables. Leaving them in the regression allows us to test, rather than assume, that their market valuation is equal to their book value.

(61.) The LAD regressions, being nonlinear regression procedures, do not have analogous panel data corrections. Therefore the standard errors for analyses using these procedures may be understated by as much as a factor of 3.3 (the square root of the number of time-series observations), although in practice the error is well below this bound.

(62.) Griliches and Hausman (1986).

(63.) See, for example, Bartelsman, Caballero, and Lyons (1994) or Brynjolfsson and Hitt (2001) for further discussion.

(64.) Nothing in the theory requires the installed price of computer capital to be invariant as technology evolves and investor expectations change, any more than the value of an oil company's proven reserves need be invariant.

(65.) These results build on earlier work reported in Hitt and Brynjolfsson (1997) and Bresnahan, Brynjolfsson, and Hitt (2002).

(66.) Results are similar when probit or ordered probit regression techniques are used. We report Spearman rank-order correlations because they are easier to interpret given the nonmetric nature of most of our work system variables. Included in the regressions are separate controls for 1 1/2-digit SIC industries (see the appendix for details).

(67.) Not including the processing power of personal computers.

(68.) See Bresnahan, Brynjolfsson, and Hitt (2002), Bresnahan (1997), and Hitt and Brynjolfsson (1997). A survey of related work appears in Brynjolfsson and Hitt (2000).

(69.) Hitt and Brynjolfsson (1997).

(70.) If the variables are not centered, the interpretation of the direct coefficients is changed. Consider a hypothetical regression equation V = [[gamma].sub.o]O + [[gamma].sub.I] I + [epsilon] (model A) compared with another equation representing the "true" model of the underlying relationship V = [[gamma]'.sub.o]O + [[gamma]'.sub.I]I + [[gamma]'.sub.OI] (O x I) + [epsilon] (model B). When the interaction term is not centered, [[gamma].sub.I] = [[gamma].sub.I] + [bar]O[[gamma]'.sub.OI], where [bar] O is the mean of O. Centering the interaction term removes the [bar] O[[gamma].sub.OI] term and thus preserves the interpretation of [[gamma].sub.I] as the contribution of the linear term. However, in the centered regression it is possible that the linear term changes if the true underlying data relationship is not precisely the linear interaction model described by model B (for example, if there are additional unobserved assets).

(71.) Bresnahan, Brynjolfsson, and Hitt (2002).

(72.) Bresnahan, Brynjolfsson, and Hitt (2002).

(73.) Brynjolfsson and Hitt (2001).

(74.) Our 1 1/2-digit controls divide the economy into ten sectors: high-technology manufacturing, process manufacturing, other nondurable manufacturing, other durable manufacturing, mining and construction, trade, transportation, utilities, finance, and other services. See the appendix for details.

(75.) Bresnahan, Brynjolfsson, and Hitt (2002).

(76.) Yorukoglu (1998); Greenwood (1997); Greenwood and Jovanovic (1999); Hobijn and Jovanovic (2001).

(77.) R. Hall (2000b).

References

Abel, Andrew B. 1990. "Consumption and Investment." In Handbook of Monetary Economics, vol. 2, edited by Benjamin M. Friedman and Frank H. Hahn. North-Holland.

Appelgate, Lynda M., James I. Cash, and D. Quinn Mills 1988. "Information Technology and Tomorrow's Manager." Harvard Business Review 66(6): 128-36.

Autor, David H., Frank Levy, and Richard Murnane. 2000. "Upstairs, Downstairs: Computer-Skill Complementarity and Computer-Labor Substitution on Two Floors of a Large Bank." Working Paper 7890. Cambridge, Mass.: National Bureau of Economic Research (September).

Baily, Martin Neil. 1981. "Productivity and the Services of Capital and Labor." BPEA, 1:1981, 1-50.

Barras, Richard. 1990. "Interactive Innovation in Financial and Business Services: The Vanguard of the Service Revolution." Research Policy 19(3): 215-37.

Bartelsman, Eric J., Ricardo J. Caballero, and Richard K. Lyons. 1994. "Customer- and Supplier-Driven Externalities." American Economic Review 84(4): 1075-84.

Barth, Mary E., John A. Elliott, and Mark W. Finn. 1997. "Market Rewards Associated with Patterns of Increasing Earnings." Research Paper 1423R. Stanford Graduate School of Business (November).

Black, Sandra E., and Lisa M. Lynch. 2001. "How to Compete: The Impact of Workplace Practices and Information Technology on Productivity." Review of Economics and Statistics 83(3): 434-45.

--. Forthcoming. "Measuring Organizational Capital in the New Economy." In Measuring Capital in the New Economy, edited by Carol Corrado, John Haltiwanger, and Dan Sichel. University of Chicago Press.

Bond, Stephen R., and Jason G. Cummins. 2000. "The Stock Market and Investment in the New Economy: Some Tangible Facts and Intangible Fictions." BPEA, 1:2000, 61-108.

Bresnahan, Timothy F. 1997. "Computerization and Wage Dispersion: An Analytical Reinterpretation." Working Paper 31. Stanford University (August).

--. 1999. "Computing." In U.S. Industry in 2000: Studies in Competitive Performance, edited by David C. Mowrey. Washington: National Academy Press.

Bresnahan, Timothy F., Erik Brynjolfsson, and Lorin M. Hitt. 2000. "Technology, Organization, and the Demand for Skilled Labor." In The New Relationship: Human Capital in the American Corporation, edited by Margaret M. Blair and Thomas A. Kochan. Brookings.

--. 2002. "Information Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence." Quarterly Journal of Economics 117(1): 339-76.

Bresnahan, Timothy F., and Shane Greenstein. 1997. "Technical Progress and Co-invention in Computing and in the Uses of Computers." BPEA, Microeconomics, 1-77.

Brynjolfsson, Erik, and Lorin M. Hitt. 1995. "Information Technology as a Factor of Production: The Role of Differences among Firms." Economics of Innovation and New Technology 3(3-4): 183-99.

--. 2000. "Beyond Computation: Information Technology, Organizational Transformation and Business Performance." Journal of Economic Perspectives 14(4): 24-48.

--. 2001. "Computers and Productivity Growth: Firm-Level Evidence." Unpublished paper. Massachusetts Institute of Technology and University of Pennsylvania (September).

Brynjolfsson, Erik, and Haim Mendelson. 1993 "Information Systems and the Organization of Modern Enterprise." Journal of Organizational Computing 3(3): 245-55.

Brynjolfsson, Erik, Amy Autin Renshaw, and Marshall Van Alstyne. 1997. "The Matrix of Change." Sloan Management Review 38(2): 37-54.

Brynjolfsson, Erik, and Shinku Yang. 1999. "Intangible Benefits and Costs of Computer Investments: Evidence from the Financial Markets." Unpublished paper. Massachusetts Institute of Technology (December).

Chan, Louis K. C., Josef Lakonishok, and Theodore Sougiannis. 1999. "The Stock Market Valuation of Research and Development Expenditures." Working Paper 7223. Cambridge, Mass.: National Bureau of Economic Research (July).

Colenso, Kim. 2000. "Satisfying SOP 98-1, Accounting for the Costs of Computer Software Developed or Obtained for Internal Use." Unpublished paper. Boulder, Colo.: Artemis Management Systems.

Davenport, Thomas H., and James E. Short. 1990. "The New Industrial Engineering: Information Technology and Business Process Redesign." Sloan Management Review 31(4): 11-27.

David, Paul A. 1990. "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox." American Economic Review, Papers and Proceedings 80(2): 355-61.

Ernst & Young. 1998. Accounting for the Costs of Computer Software Developed or Obtained for Internal Use--Statement Position 98-1. Assurance & Advisory Business Services, November.

Gormley, J., and others. 1998. "The Runaway Costs of Packaged Applications." Forrester Report 3(5): 1-15.

Greene, William H. 1993. Econometric Analysis, 2nd edition. Macmillan.

Greenwood, Jeremy. 1997. The Third Industrial Revolution: Technology, Productivity, and Income Inequality. Washington: AEI Press.

Greenwood, Jeremy, and Boyan Jovanovic. 1999. "The Information Technology Revolution and the Stock Market." American Economic Review, Papers and Proceedings 89(2): 116-22.

Griliches, Zvi. 1981. "Market Value, R&D, and Patents." Economic Letters 7(2): 183-87.

Griliches, Zvi, and Iain Cockburn. 1988. "Industry Effects and Appropriability Measures in the Stock Market's Valuation of R&D and Patents." American Economic Review, Papers and Proceedings 78(2): 419-23.

Griliches, Zvi, and Jerry A. Hausman. 1986. "Errors in Variables in Panel Data." Journal of Econometrics 31(1): 93-118.

Hall, Bronwyn H. 1990. "The Manufacturing Sector Master File: 1959-1987." Working Paper 3366. Cambridge, Mass.: National Bureau of Economic Research (May).

--. 1993a. "The Stock Market's Valuation of R&D Investment during the 1980s." American Economic Review 83(2): 259-64.

--. 1993b. "Industrial Research during the 1980s: Did the Rate of Return Fall?" BPEA, Microeconomics 2:1993, 289-330.

--. 1999. "Innovation and Market Value." Working Paper 6984. Cambridge, Mass.: National Bureau of Economic Research (February).

Hall, Robert E. 2000a. "E-Capital: The Link between the Stock Market and the Labor Market in the 1990s." BPEA 2:2000, 73-102.

--. 2000b. "Reorganization." Carnegie-Rochester Conference Series on Public Policy 52: 1-22.

--. 2001a. "Struggling to Understand the Stock Market." American Economic Review, Papers and Proceedings 91(2): 1-11.

--. 2001b. "The Stock Market and Capital Accumulation." American Economic Review 91(5): 1185-1202.

Hayashi, Fumio. 1982. "Tobin's Marginal q and Average q: A Neoclassical Interpretation." Econometrica 50(1): 213-24.

Hitt, Lorin M., and Erik Brynjolfsson. 1997. "Information Technology and Internal Firm Organization: An Exploratory Analysis." Journal of Management Information Systems 14(2): 81-101.

Hobijn, Bart, and Boyan Jovanovic. 2001. "The Information Technology Revolution and the Stock Market: Evidence." American Economic Review 91(5): 1203-20.

Huselid, Mark A. 1995. "The Impact of Human Resource Management Practices on Turnover, Productivity, and Corporate Financial Performance." Academy of Management Review 38(3): 635-72.

Ichniowski, Casey, Kathryn Shaw, and Giovanna Prennushi. 1997. "The Effects of Human Resource Management Practices on Productivity: A Study of Steel Fishing Lines." American Economic Review 87(3): 291-313.

Ito, Harumi. 1996. "Essays on Investment Adjustment Costs." Ph.D. dissertation, Stanford University (September).

Jorgenson, Dale W., and Barbara M. Fraumeni. 1995. "Investment in Education and U.S. Economic Growth." In Productivity, Volume 1: Postwar U.S. Economic Growth, edited by Dale W. Jorgenson. MIT Press.

Kemerer, Chris F., and Glenn L. Sosa. 1991. "Systems Development Risks in Strategic Information Systems." Information and Software Technology 33(3): 212-23.

Lev, Baruch. 2001. Intangibles: Management, Measurement, and Reporting. Brookings.

Lev, Baruch, and Theodore Sougiannis. 1996. "The Capitalization, Amortization, and Value-Relevance of R&D." Journal of Accounting and Economics 21: 107-38.

Lev, Baruch, and Paul Zarowin. 1999. "The Boundaries of Financial Reporting and How to Extend Them." Unpublished paper. New York University (February).

Lichtenberg, Frank R. 1995. "The Output Contributions of Computer Equipment and Personnel: A Firm-Level Analysis." Economics of Innovation and New Technology 3(3): 201-17.

Loveman, Gary W. 1986. "The Productivity of Information Technology Capital: An Econometric Analysis." Unpublished paper. Massachusetts Institute of Technology (January).

MacDuffie, John Paul. 1995. "Human Resource Bundles and Manufacturing Performance: Organizational Logic and Flexible Production Systems in the World Auto Industry." Industrial and Labor Relations Review 48(2): 197-221.

Malone, Thomas, and John Rockart. 1991. "Computers, Networks and the Corporation." Scientific American 265(3): 92-99.

McKinsey Global Institute. 2001. "U.S. Productivity Growth 1995-2000." Washington: McKinsey & Company.

Milgrom, Paul, and John M. Roberts. 1990. "The Economics of Modern Manufacturing: Technology, Strategy and Organization." American Economic Review 80(3): 511-28.

Montgomery, Cynthia A., and Birger Wernerfelt. 11988. "Diversification, Ricardian Rents, and Tobin's q." Rand Journal of Economics 19(4): 623-32.

Naik, Vasanttilak. 1994. "Asset Prices in Dynamic Production Economies with Time-Varying Risk." Review of Financial Studies 7(4): 781-801.

Orlikowski, W. J. 1992. "Learning from Notes: Organizational Issues in Groupware Implementation." Conference on Computer Supported Cooperative Work, Toronto, Canada.

Osterman, Paul. 1994. "How Common Is Workplace Transformation and Who Adopts It?" Industrial and Labor Relations Review 47(2): 173-88.

Radner, Roy. 1993. "The Organization of Decentralized Information Processing." Econometrica 61(5): 1109-46.

Sauer, Christopher, and Philip W. Yetton. 1997. Steps to the Future: Fresh Thinking on the Management of IT-Based Organizational Transformation. San Francisco: Jossey-Bass.

Scott Morton, Michael S., ed. 1991. The Corporation of the 1990s: Information Technology and Organizational Transformation. Oxford University Press.

Sloan, Alfred P. 1964. My Years with General Motors. Doubleday.

Tobin, James. 1969. "A General Equilibrium Approach to Monetary Theory." Journal of Money, Credit and Banking 1(1): 15-29.

Woerner, Stephanie L. 2001. "Networked at Cisco." Center for eBusiness Teaching Case 1. MIT Sloan School of Management (July).

Yang, Shinkyu. 1994. "The Relationship between IT Investment and Market Value of Firms." Master's thesis, MIT Sloan School of Management.

Yorukoglu, Mehmet. 1998. "The Information Technology Productivity Paradox." Review of Economic Dynamics 1(2): 551-92.

ERIK BRYNJOLFSSON
Massachusetts Institute of Technology

LORIN M. HITT
University of Pennsylvania

SHINKYU YANG
New York University