October, 1998

Erik Brynjolfsson

MIT Sloan School of Management

erikb@stanford.edu

Lorin M. Hitt

The Wharton School

University of Pennsylvania

lhitt@wharton.upenn.edu

While many theories about the relationship between information technology (IT) and organizational design have been proposed, there is little empirical evidence on the issue. We examine the influence of organizational design on the demand for IT and the productivity of IT investments, using data from approximately 380 US firms. We find greater demand for IT in firms with greater decentralization of decision rights (especially the use of self-managing teams), and greater investments in human capital, including training and screening by education. In addition, IT has a greater contribution to output in firms that adopt a more decentralized and human capital-intensive work system. This relationship is robust to alternative measures of IT and of work systems, as well as alternative specifications for demand and for productivity. These findings lend support to the idea that organizational practices are important determinants of IT demand and productivity.

This research has been generously funded by the National Science Foundation (Grant IIS-9733877), The Organization for Economic Cooperation and Development, the Stanford Computer Industry Project, the Industrial Performance Center and the Center for Coordination Science at MIT and NationsBanc Montgomery Securities.

We are grateful to Susan Athey, Eric Clemons, Zvi Griliches, Bengt Holmstrom, Tom Hubbard, Maryellen Kelley, Thomas Kochan, Thomas Malone, Roy Radner, Jim Rebitzer, Kathryn Shaw, and Albert Wenger as well as to seminar participants at Carnegie-Mellon University, MIT, the National Bureau of Economic Research, New York University, Stanford University, and the Wharton School for valuable comments on earlier versions ofthis paper. We would also like to thank Incon Research, the Center for Survey Research, Computer Intelligence Infocorp and Informationweek magazine for providing or collecting essential data.

While many theories about the relationship between information technology (IT) and organizational design have been proposed, there is little empirical evidence on the issue. We examine the influence of organizational design on the demand for IT and the productivity of IT investments, using data from approximately 380 US firms. We find greater demand for IT in firms with greater decentralization of decision rights (especially the use of self-managing teams), and greater investments in human capital, including training and screening by education. In addition, IT has a greater contribution to output in firms that adopt a more decentralized and human capital-intensive work system. This relationship is robust to alternative measures of IT and of work systems, as well as alternative specifications for demand and for productivity. These findings lend support to the idea that organizational practices are important determinants of IT demand and productivity.

Technological innovation is rarely neutral in its effect on different organizational forms.

In the early 20^th century, small electric motors provided the greatest benefit to "horizontal" factory designs that distributed machinery and work (David, 1990). On a larger scale, the transition from handicraft work to machine production has been associated with a shift to a system of more hierarchical control and Taylorist work practices (Piore and Sabel, 1984).

The diffusion of information technology (IT) into the modern workplace may also benefit some types of organizations more than others. The nature of the relationship between IT and organizational design is of increasing concern to managers and numerous authors have developed theories and predictions. For instance, Malone, Yates, and Benjamin (1987) hypothesize that IT will favor more decentralized decision-making relative to centralized hierarchies. Milgrom and Roberts (1990) argue that the exogenous price decline of IT as a key driver that has favored "modern manufacturing" relative to "mass production." While the number of theories about IT’s organizational impacts continues to grow, there is much less empirical evidence on this question. Case studies of particular companies or industries have provided valuable insight into particular uses of IT, but we know surprisingly little about whether there are any more general patterns that relate IT adoption and organizational design.

As a general-purpose technology, IT has diffused broadly into many different industries and applications. The basic components of the technology are similar, regardless of where they are used, just as the task of information processing in organizations has many commonalities regardless of industry, product line, or location (Galbraith, 1977). However, some types of organizations may benefit disproportionately from technical advances in computing and communications. While energy-intensive methods of production are favored when the price of energy declines, so one would expect that "information intensive" methods of production would be favored when the price of (computer-generated) information declines. Organizations that choose to incur higher IT expenses presumably expect compensating benefits, such as the ability to respond more quickly and accurately to changes in their environment or to use less of other resources. As the price of IT declines, these trade-offs will increasingly favor strategies and structures that use IT more intensively. In particular, Malone, Yates, and Benjamin (1987) have specifically predicted that declines in the costs of IT will tend to favor decentralized "coordination-intensive" structures over more centralized ways of organizing, which economize in coordination costs at the expense of efficient resource use and flexibility. They argue that while all types of firms will benefit in absolute terms from a decline in the price of IT, firms in which IT represents a larger share of inputs have the most to gain.

However, we do not have a good understanding of what types of organizations are the most intensive users of IT. Empirical research has been hampered by a scarcity of meaningful, generalizable data on organizational design. As a result, theories about IT’s organizational impacts are often only loosely grounded in empirical evidence and, inevitably, different theories sometimes come to opposite conclusions about how IT affects organizations. For instance, the question of whether IT will tend to favor centralized decision-making or decentralization has been debated for almost 40 years (Leavitt and Whisler, 1958; George and King, 1991).

This paper explores the relationship between IT and organizational design using new data from a broad cross section of firms. We empirically examine the relationship between organizational design and IT use, and how the combination of IT and organizational design affects firm productivity. This is accomplished by matching detailed data on IT use and productivity with a new survey of organizational practices for approximately 380 large U.S. firms.

We find that organizational practices are predictive of both firms’ demand for IT and the contribution of IT investment to firm output. In particular, demand for IT is greatest in firms that delegate more decision-making to line workers, use self-managing teams more extensively, use group-oriented incentive systems, hire more educated workers, and invest more in training than other firms in the same industry. Furthermore, these organizational practices are all correlated with one another, suggesting that firms can usefully be characterized as choosing different points on a single-dimension that, at the risk of oversimplification, we label centralization-decentralization. We also find evidence that firms using more decentralized work systems derive higher benefits from their IT investments, and this result is robust to alternative measures of decentralization and changes in the productivity specification.

While a comprehensive theory of the role of information in organizations remains elusive, our empirical results are more supportive of some classes of theories than others. In particular, the results are most supportive of those theories that predict that lower information processing and communications costs will be especially beneficial for organizations that distribute decision-making. In contrast, theories that predict greater centralization can be reconciled with our results only if some additional assumptions are made about unobserved factors in the environment.

In the next section, we briefly review some of the relevant literature, including theories of the relationship between IT and organization. Section 3 describes the empirical methodology and data. Section 4 presents the results, and we conclude the paper with a brief summary and discussion in Section 5.

This research relates to a number of different areas of the management literature. The work systems we identify are similar to what has been termed "high performance work systems" (Ichniowski, Kochan, Levine, Olson, and Strauss, 1996); our findings may help explain both the increasing diffusion and performance impact of this type of work organization. This paper also builds on earlier work on information systems management that has found differential effects of IT on alternative systems for organizing work, and recent work that has found substantial productivity benefits from the use of IT at the firm level (Brynjolfsson and Hitt, 1995; 1996; 1997; Lichtenberg, 1995). There have also been several papers that examined the industry-level determinants of the demand for IT, mainly as a function of price and output (Brynjolfsson, 1996; Gurbaxani, 1992; Gurbaxani and Mendelson, 1990) and the demand for mainframes and client-server computing (Bresnahan and Greenstein, 1997; Greenstein, 1996). Finally, portions of the empirical approach draw on the economics of complementarities (Athey and Stern, 1997; Milgrom and Roberts, 1990; Holmstrom and Milgrom, 1994).

A number of empirical studies have examined the relationship between organizational practices and firm performance (Dunlop and Weil, 1993; Huselid, 1994; Ichniowski, Shaw, and Prunnushi, 1994; Lawler, Mohrman, and Ledford, 1995; MacDuffie, 1995). In particular, attention has been focused on a cluster of organizational design choices labeled "high performance work practices." These can include the use of team-based production, the allocation of decision authority to workers, higher skill levels and training, broader job responsibilities, use of quality circles, adoption of total quality management, team-oriented incentive systems, and pre-employment screening practices (Ichniowski et. al., 1996). Several of these studies have recently found large economic benefits from the use of clusters of these practices leading Osterman (1994) to question why they are not more widespread. Ichniowski et. al. suggest that the need to coordinate changes in multiple practices, including IT use, may be part of the difficulty.

However, while there is some general consistency among the studies, they differ substantially with respect to their theoretical justification, setting, and operationalization of work systems. A comparison of work systems components that have been used in prior work appears in Appendix B.

Management theorists and, to a lesser extent, economists, have also voiced a perception that the trade-offs among different organizational designs has changed in recent years. In particular, hierarchy, centralized control, and bureaucracy seem increasingly out of favor. For instance, Drucker (1988) argues that we are in the midst of a "third period of change: the shift from the command-and-control organization...to the information-based organization."

Similarly, Aoki (1990) writes

It is important to note that the emerging doctrine of broadly delegating decision-making differs strikingly from the Taylorist tradition that helped make Henry Ford’s factories successful earlier in this century. This tradition can be characterized by a rigid distinction made between the decision-maker and the workers, the "separation of conceptualization and execution" (Sabel at al., 1989). The philosophy is epitomized by the modestly titled "First Law of Scientific Management" laid down at the beginning of the century:

Holmstrom and Milgrom (1994) venture a possible explanation for the change in received wisdom:

A former CEO of General Motors links the change more specifically to new technologies:

However, there is no unanimity regarding the causes, or even the existence, of any new trade-offs among organizational designs, which suggests a role for more detailed data collection and analysis.

Driven by advances in microelectronics, the quality-adjusted price of IT hardware has dropped by 20-30% annually for the last thirty years (Berndt and Griliches, 1990; Gordon, 1990). This has led to over a hundredfold increase in share of computers in the real capital stock since 1970. The availability of the personal computer in the 1980s shifted the location of computing power from large centralized "utilities" to workers’ desktops. In 1987 there was a personal computer (PC) for every 30 employees in large firms; by 1994 there was one PC for every six. What may be more important is that many other employees interact directly or indirectly with computers through terminals, embedded computing technologies or by using computer-generated information and reports. At the same time, significant growth in "coordination technologies", such as local area networks, databases, and electronic mail has reduced the costs of communication and collaboration among workers.

Studies using firm level data typically have found unexpectedly strong correlations between IT and productivity. In these studies, the marginal benefit of the increase in output associated with an increase IT investment appears to be larger than the marginal costs of IT (Brynjolfsson and Hitt, 1995, 1996, 1997; Lichtenberg, 1995). This has led Gordon (1994) to ask "if IT has excess returns, what is the hidden force that prevents greater investments?"

As with high-performance work organization, one possible answer is the difficulties associated with coordinated changes in multiple components of a broader system (termed "adjustment costs" by economists). In fact, individual firm-effects appear to account for up to half of the apparent excess returns to IT (Brynjolfsson and Hitt, 1995), and the returns to IT are substantially greater in long differences (5 to 7 years) than in short differences (1 year) (Brynjolfsson and Hitt, 1997). This suggests that some hidden, slowly changing, firm characteristics are correlated with IT and have important effects on productivity but are not counted in the standard inputs used to estimate productivity. In fact, one can think of the adjustment costs incurred in integrating IT into a work system as investments in an intangible, yet nonetheless valuable, asset (Hayashi, 1982).

The importance of integrating technology and organization has become more apparent as computers move from their traditional role in the "back-office" to supporting the day-to-day activities of line workers and managers, particularly those involved in knowledge work. The development of decentralized computing technologies has also coincided with the increased attention to business process redesign, emphasizing the interaction of work organization and investments in information systems (Davenport and Short 1990; Hammer, 1990). However, redesigning organizations is apparently not easy; firms that require a greater degree of organizational redesign appear to adopt new types of IT more slowly (Bresnahan and Greenstein, 1997).

While the relationship between information and information processing capability and organization has a long history in economic thought, there is no consensus on a theoretical model that captures all of the relevant relationships. However, there is a substantial body of literature that provides different perspectives on how IT and organizational design might be related. By conducting empirical analyses, we may be able to demonstrate which of these perspectives are more promising in modeling actual firm behavior.

Brynjolfsson and Mendelson (1993) stress that the information systems and decision-making structure of organizations are jointly determined. To the extent that computers lower the cost of lateral communication, enable firms to obtain more information to process, or increase the capabilities of line workers to perform information processing tasks, a variety of models would predict an association between computers and greater levels of decentralized authority. For example, Wyner and Malone (1996) present a model in which decreased communications costs enable firms to utilize decentralized organizational structures. Similarly, Radner’s (1993) finding that hierarchical decision-making minimizes communications costs suggests that the benefits of hierarchy diminish as communication becomes less expensive. In the literature on "bounded rationality," Van Zandt (1997) provides a model in which firms will optimally decentralize decision-making as the flow rate of new information increases relative to the capabilities of the central information processors. Mendelson and Pillai (forthcoming) argue that organizations in industries characterized by rapid change will invest in both technological and organizational factors to reduce information overload. Furthermore, the location of knowledge within the organization also may influence organizational design: if critical specific knowledge is held by line workers, organizations would optimally choose delegated decision making (Anand and Mendelson, 1997; Aoki, 1986), as long as the communications and coordination costs do not outweigh the benefits of more efficient utilization of knowledge and information.

However, studies that have emphasized vertical communication and the coordination advantages of centralized control often lead to the conclusion that information technology should lead to greater reliance on hierarchy and central planning. For instance, Bolton and Dewatripoint (1994) argue that lower communication costs make it easier for a central decision-maker to coordinate without having to delegate decision-making to middle managers. Others have also argued that technology may also improve the capabilities of central decision-makers (Leavitt and Whistler, 1958), enabling them to process more information without suffering from information overload.

Previous empirical studies have linked computers to shifts in employment toward higher skilled workers, increased employment in "knowledge work" and "information work" occupations, higher wages for computer-using employees, and increased income inequality (Wolff, 1996; Krueger, 1993; Autor, Katz and Krueger, 1997; Berman, Bound, and Griliches, 1994). Bresnahan (1997) hypothesizes that the economic effects of computerization reflect the emergence of a broader system of work; this system includes not only computers, but also new types of work organization and new levels of product and service quality. A growing body of case studies supports this view and there is also some evidence from studies of larger samples. For instance, Kelley (1994) found that computerized machinery was also disproportionately associated with modern work in metalworking plants. In an exploratory study of firm-level data, Hitt and Brynjolfsson (1997) report that decentralized work practices are often correlated with increased use of IT. However, Pinsonneault and Kraemer (1997) found that whether IT led to greater centralization or decentralization in a sample of government organizations depended significantly on organizational politics.

In this paper, we extend the earlier work in this area by 1) explicitly estimating demand equations at the firm-level, rather than correlations, 2) considering the performance effects of the IT-organization relationship at the firm-level, and 3) evaluating alternative explanations for the results. In part, our more detailed analysis is made possible because we employ a substantially larger firm-level data set than earlier studies and analyze a richer set of measures for IT and organizational design.

The relationships between IT and organizational design are explored using multi-industry, firm-level data on IT characteristics, work practices, and organizational performance. Production functions are used for performance measurement since they have a well-established theoretical framework and has been used in previous work on IT and firm performance (Brynjolfsson and Hitt, 1995; 1996; Lichtenberg, 1995).

The level of analysis was chosen to strike a balance between focused case studies and research on broader trends in the economy. Case studies can provide insight into the mechanisms by which IT and organizational design interact but lack generalizability, while economy-wide or industry-based studies have the opposite strengths and weaknesses.

The analysis focuses on firm level data, rather than plant-level data for a number of reasons. First, when comparing the costs and benefits of alternative work systems, it is important to define clearly the boundaries of the unit of analysis. For many of the relevant costs and benefits, it may not be meaningful to treat establishments within a firm as separate entities. IT networks often span multiple establishments within the same firms, as do managerial decision-making activities, yet neither is likely to be fully reflected in the accounting ledgers of the individual establishments. The common management and technology infrastructure of establishments within a firm may be associated with common human resource practices, work systems, and "corporate culture" (Milgrom and Roberts, 1992). More generally, incomplete contracts theory (Grossman and Hart, 1986; Hart, 1988) argues that because firm boundaries are set to address problems of contractual incompleteness, the presence of multiple establishments in a single integrated firm suggests some difficulty that prevents these establishments from operating on a stand-alone basis.

Partly as a result of these factors, substantially more data are publicly available on firms than on plants or business units. This information is important for various parts of this analysis, particularly for performance measurement. Unlike firms, individual establishments do not have audited financial statements and the data that are available at an unconsolidated level are subject to intra-firm reporting biases that can add substantial error (Kaplan, 1989).

A multi-industry approach is used because it makes it easier to generalize the findings. Previous case-based or industry-specific studies have left open the question of whether findings generalize to the broader economy (Ichniowski et al., 1996). The relationship between IT and organizational design may differ on a case by case basis, depending on unobserved, idiosyncratic factors in a particular company or industry. Broader trends, if any exist, may be more apparent in a more diverse sample.

The data set used for this analysis is a cross sectional survey of organizational practices conducted in 1995 and 1996 matched to a panel of IT spending and productivity metrics over the 1987-1994 time period. A brief description of each data source follows with additional detail in Appendix A.

Computer Technology: The measures of IT use were derived from the Computer Intelligence Infocorp installation database that details IT spending by site for companies in the Fortune 1000 (approximately 25,000 sites were aggregated to form the measures for the 1000 companies that represent the total population in any given year). This database is compiled from telephone surveys that detail the ownership for IT equipment and related products. Most sites are updated at least annually with greater sampling for larger sites. The year-end state of the database from 1987 to 1994 was used for the IT measures. These data include variables capturing the total capital stock of IT (central processors, personal computers [PCs], and peripherals) as well as measures of computing power and the number of PCs. The IT data do not include all types of information processing or communication equipment and are likely to miss that portion of computer equipment that is purchased by individuals or departments without the knowledge of information systems personnel.

Compustat. Compustat data was used to construct various performance metrics and provide additional firm information not covered by other sources. For the calculation of productivity, we followed the procedures in Hall (1990) and Brynjolfsson and Hitt (1997). Measures were created for output, capital, labor, and value added in order to calculate a production function relationship between value added and the various inputs.

Organizational Practices: There are no publicly available data on the organizational characteristics of interest and we therefore undertook a new survey. This survey was based on the organizational characteristics identified in prior literature and adapted questions from prior surveys on human resource practices and workplace transformation (Huselid, 1994; Ichniowski, Shaw and Prunnushi, 1997; Osterman, 1994). These questions address the allocation of various types of decision-making authority, the use of self-managing teams, investments in training and education, incentives and promotion criteria, the extent of computerization and other miscellaneous characteristics of the workplace (further detail appears in the Results section).

Administered to senior human resource managers or their designees, the survey asked questions about organizational practices at the most typical establishment. The approach of Osterman (1994) was followed in focusing on a single class of employee, termed "production employees" (which corresponds to Osterman’s "core employee"). A production employee was defined as "non-managerial, non-supervisory personnel directly involved in producing a firm’s product or delivering its service."

Organizational data were collected in three waves, covering most of the Fortune 1000. A total of 416 firms provided at least some data for the study, including 93 from the first survey, conducted in Summer, 1995, 138 from a survey administered in Fall, 1995 and 199 from a survey in Summer, 1996. Because some firms on the HR practices survey do not have complete matching data from CII and Compustat or have missing data on key questions on the survey, most analyses are conducted using a sample size of 379 firms. After the first wave of data was collected, the survey instrument was shortened to improve the response rate, and several of the survey questions were rephrased for clarity. The overall response rate was 52%. The most common explanations for non-response were "company policy" or "didn’t have time."

In interpreting the results, it is important to keep in mind the limitations of the study design. By matching cross-sectional data on organizational characteristics to panel data on productivity, this approach is consistent with the attempt to use workplace characteristics to explain long-run differences in productivity. However, it does limit the ability to address issues of causality. While it would have been ideal to have panel data on organizational characteristics as well, these data were unavailable historically, and extensive retrospective surveying was deemed too unreliable.

Summary statistics on the sample respondents are provided in Table 1 for firms that have a complete set of production inputs and outputs (value added, capital, labor, and IT capital) from CII and Compustat. The main difference between the firms in the sample and the rest of the Fortune 1000 is that included firms tend to be slightly larger and more capital-intensive. Approximately 55% of the sample are from the manufacturing, mining, or construction sectors and 45% are in services.

One potential difficulty of our sampling approach is that the practices reported by the respondent may not be representative of the work practices across the entire firm. To address this issue, the revised instrument contained questions about how representative production workers were in terms of total employment and the uniformity of work practices for this category of workers. Overall, for the average firm in the second survey subsample, production workers account for about two-thirds of total employment and organizational practices are found to be fairly uniform: 65% of respondents said that all workers have the same work practices. Furthermore, 82% reported that at least 80% of workers had the same work practices, and 92% reported that at least 40% had the same work practices.

Another concern is that, even if production workers are important and fairly uniform within the firm, our survey instrument may not accurately capture the important organizational design characteristics. Unlike employment or wages, there are no clear metrics for measuring concepts like "decentralization." However, our results appear to be consistent whether we measure decentralization as a system of work practices, the use of specific work practices (e.g., teams), or examine the specific delegation of authority for different types of tasks. Moreover, these alternative measures of work systems are all highly correlated, suggesting internal validity of our survey instrument. Given that the various work practices that are complementary to IT (both the ones we measure and those that we do not) are likely to be complements themselves, it is probably better to interpret our measures as capturing a piece of a broader work system that includes these components rather than focusing on the measured practices themselves.

What types of organizations are most likely to gain from future declines in the cost of IT? Clearly, a given percentage decrease in the cost of IT will benefit firms that are more IT-intensive more than firms that rely less heavily on IT. We find that decentralized organizations tend to be more IT-intensive, regardless of whether we look at the demand for IT or IT’s output elasticity (roughly, the increase in output associated with a 1% increase in IT).

In this section, we examine the correlation between IT and decentralization and related practices. All correlational analysis is done using Spearman rank order correlations between various measures of IT and the work system variables, controlling for firm size (employment), production worker occupation, and industry. Six different measures of IT are used. Three are from CII, which include total value of IT installed base (ITCAP), total central processing power in millions of instructions per second (MIPS), and number of personal computers (TOTPC), which captures decentralized computing. Three additional measures are obtained from the organizational practices survey: the percentage of production workers that use general purpose computers, the percentage of the work force that uses electronic mail, and a five-point scale of the computerization of the workplace (COMP). Multiple measures are employed because they capture slightly different aspects of computerization (for example, MIPS measures centralized computing, while TOTPC measures decentralized computing).

The three measures of IT taken from the organizational practices survey have the advantage that they focus on the technology in use by the same production workers covered by the work systems data, but the disadvantage that some respondents might let their answers to some questions influence their answers to other questions. On the other hand, the three IT measures derived from CII data have the advantage that they are collected by a separate census of different individuals (with different job titles), but the disadvantage that they apply to the whole firm, not necessarily the specific portion described by the work practices survey. Thus, some additional noise is likely to be introduced when the organizational practice variables are matched to the CII-based measures.

As discussed in Section 2, there is an ongoing debate about whether IT is associated with centralized decision-making, decentralization, both, or neither. While there are case studies that bolster each of these views, there has been no broader evidence about whether IT is more strongly correlated with centralization or decentralization. The survey addresses several aspects of this question, including the use of structures that decentralize authority, such as self-managing teams, employee involvement groups, and broader duties assigned to line workers. It also asks directly about the allocation of individual decisions on various aspects of the production process, such as the pace or method of work and the locus of responsibility for various types of decisions. Finally, we consider practices associated with decentralization, such as team-oriented incentives and measures to increase human capital (training, pre-employment screening).

The correlations of these measures with IT are shown in Table 2a. In terms of structural decentralization, the results show strong correlations between the use of self-managing teams and IT as well as some evidence that high IT firms employ broader job classifications and make greater use of employee involvement groups. There is also evidence that firms use supporting incentives for team-based organization, either through promotion incentives or team building exercises. Finally, IT-intensive firms also train a larger fraction of their staff and screen new employees more extensively for education.

The individual decentralization measures also show a substantial correlation with IT. Firms that use more IT are also more likely to give their line workers discretion over the methods used in their work. For the second and third survey, we expanded the individual decision authority scale to cover seven distinct types of decisions and broadened the scale from 3 points to 5. Of the 42 pairs of IT measure and decision-making type, 41 show positive correlations and 74% of the correlations are significant at p<.05 (Table 2b). Firms in which workers, not managers, are responsible for deciding the "best way to perform tasks" are especially likely to have higher demand for IT. This suggests that IT is less of a tool for monitoring compliance with management decisions than for enabling workers to make better decisions independently.

As a summary statistic, we created a composite decentralization scale by summing the standardized values of the seven decision authority variables (Cronbach’s alpha=.73). We find a consistent positive correlation between IT and decentralization, significant at p<.01 for four of the six IT measures. Finally, an alternative measure that simply asks directly for "the extent to which workers are provided individual control and decision making responsibility" shows similar results when correlated with IT.

On balance, there is substantial evidence that IT is broadly related to decentralized authority, both in terms of teams as well as individual decision rights. The results are consistent across IT measures, although the IT measures derived from the organizational practices survey show particularly strong correlations.

The results indicate that IT is correlated with a collection of practices that includes team-based organization, decentralization of decision rights, and acquisition of human capital. If one identifies the firm with a system of complementarity practices (Holmstrom and Milgrom, 1994), then these practices will not be chosen in isolation. Instead, all of these factors are likely to be correlated, suggesting that a single measure can be constructed that captures the adoption of these practices as a system.

To examine whether these practices tend to be adopted together as the theory would predict, a principal components analysis of the eight relevant measures of work practices was conducted (Table 2d). A Scree plot (Figure 1) suggests that the first principal component is the only non-noise factor. The factor loadings are consistent with the work system arguments and the weights on the practices are .4 or higher. Based on this analysis, a proxy for organizational decentralization is constructed (org) by summing the sum of the sign-corrected standardized values of all variables included in the factor analysis. The resulting Cronbach’s alpha is .71, which indicates adequate reliability. To facilitate interpretation when this measure is used in regression analyses, this variable is standardized to a mean of zero and a variance of 1, with higher values indicating firms that are more "decentralized."

We find that all measures of IT are significantly correlated with the org variable (Table 2e). It is possible that this relationship is due to work force composition or industry rather than a direct relationship with IT. Firms that employ a disproportionate number of professionals are likely candidates for both increased use of IT as well as the adoption of decentralized work structures.

To explore this possibility, the correlations are repeated first without controls and then successively controlling for firm size (employment), production worker occupation, and industry and then the measures of human capital: skills, education, and work force composition. The correlations tend to decline somewhat as additional control variables are added. However, considering the risk of "overcontrolling" for factors that should be free to vary with org, the correlations remain fairly high. This suggests that the results do not appear to be driven only by differences in work force composition or human capital across firm, or at least not to the extent that we are able to measure such differences.

As an alternative, we also examine results when we chose a single measure as the representative of the work system scale. Use of a single variable avoids assumptions required for aggregating non-metric scales. The variable "self-managing teams" had the highest factor loading in the principal components analysis and was selected as the representative. As reported in Table 2a, self-managing teams is highly correlated with all measures of IT after controlling for size, industry, and production worker occupation. In addition, we find that the correlations between the six measures of computerization and self-managing teams are positive and generally significant, even when the full set of controls are added, including those for skills, education, and work force composition. Four of the six correlations are significant at p<.05 and five are significant at p<.1.

A more structured way to analyze these relationships is to estimate demand equations for IT. We relate the logarithm of IT input quantity to changes in the price of IT and the prices of other factors, and demand shifters, such as industry, year, organizational design, and scale. This simple demand structure is consistent with prior work on IT demand (Brynjolfsson, 1994; Gurbaxani, 1992) and yields coefficients that can be easily interpreted in terms of demand elasticities. In addition, results are similar when we use the factor demand structure implied by a generalized Leontief or trancendental logarithmic cost function. Because the prices paid for IT and other inputs are essentially invariant across firms, the time dummy variables will capture both price effects and any other economy-wide changes in demand.

Our previous arguments have suggested that IT and ORG are complements, and in the long run we would expect that they influence each other. However, in the short run it is much easier to adjust the level of IT investment than it is to make major organizational changes. Thus, for our econometric analyses we make the assumption that ORG is not influenced by changes in IT over our sample frame, and therefore exogeneous relative to IT investment. This assumption allows us to identify the effect of ORG on IT demand. In a later section, we probe this assumption using other survey measures and instrumental variables regression and generally find that although there is evidence that ORG has changed slightly over our sample period for some firms, these changes do not appear to affect the conclusions of the paper.

Using the demand structure outline above and our identifying assumption about ORG being fixed relative to IT investment, the basic estimating equation is:

where: org is the measure of decentralization, VA is value added, IT is IT capital, and T is a set of time dummy variables.

Given the earlier results, it is not surprising that we find that the more decentralized firms have a higher demand for IT (Table 3). Interestingly, the effect of org on IT demand is roughly the same magnitude when we control for industry and overall work force composition. An examination of the industry dummy variables indicates that demand for IT is highest in finance and high-tech manufacturing, which is consistent with previous estimates using industry-level data (Gurbaxani, 1992).

The effect of org on the demand for IT is similar when the industry controls are replaced with variables representing the proportion of the firm in each industry rather than a pure 0-1 dummy variable for the firm’s primary industry (Table 3, column 5). Thus, even though the firms in our sample span multiple industries, this also does not appear to substantially alter the results. This provides further evidence that the relationship between IT and organization is capturing a system of work practices that is widely applicable across production settings and industries. We also examined the influence of whether organizational decentralization is related to the use of decentralized computer architectures and find that the results are similar irrespective of what types of computing are measured.

When we examine the demand for IT over time, we find that not only do decentralized firms have higher demand, but they also increase their demand for IT at a faster rate than centralized firms. This is illustrated in Figure 2, which plots "demand curves" for two hypothetical organizations, one which is one standard deviation above the mean in ORG and one which is one standard deviation below the mean. Both the intercept and the slope of the demand curve are allowed to vary by organizational type, but changes are constrained to be constant for all years. As in the specification in which slopes did not vary, we find that high org firms have about a 20% higher demand than low org firms (t = 8.6). In addition, we find that their elasticity of demand is 9.4% higher (t = 2.25), although as discussed above, this may simply reflect a better matching of the org variable to IT at the end of the sample period than at the beginning of the sample period.

It has long been recognized that differences in profitability can be a major determinant of the rate of diffusion of a technology (Griliches, 1957), and so one might expect that new investments in IT would occur first where they are most profitable. Thus, one explanation for the correlation we find between IT and decentralized work systems is that firms that decentralized enjoy greater benefits from adopting IT.

However, several other explanations are also possible. A second possibility is that decentralized firms might be more likely to duplicate IT investments or make wasteful IT investments. According to this argument, centralized firms may spend less on computing, but they are not necessarily getting any fewer benefits.

A third explanation is that decentralized firms may have lower adjustment costs. If firms are not all at their optimal level of investment, then the firms with lower adjustment costs may temporarily have greater adoption of IT. In fact, decentralized firms do tend to have more educated workers, and past research (e.g., Bartel and Lichtenberg, 1987) has indicated that educated workers may have an advantage in implementing new technologies.

A fourth possibility is that there may be systematic mismeasurement that leads us to overestimate the amount of IT in decentralized firms relative to centralized firms. However, the consistency of the results, regardless of which measure of IT was used, is not very supportive of this interpretation.

Finally, the correlations can also be explained by the existence of a variety of third factors that might simultaneously lead firms to use more IT and adopt decentralized structures without there being any direct link between the two. For instance, firms in a turbulent environment might choose to increase IT and also to rely more on decentralized decision-making, while those in static environments might have less demand for IT and choose greater centralization. In general, if exogenous factors that influence IT use (except for decentralization) and exogenous factors that influence decentralization (except for IT) are affiliated, then IT and decentralization can show positive covariation without any relationship between the two (Athey and Stern, 1997; Holmstrom and Milgrom, 1994).

Under any of the first three explanations, a continuing decline in the costs of IT will tend to benefit decentralized firms disproportionately. In particular, the demand equations indicate that a firm with one standard deviation above average in org will have about 20% higher demand for IT at any given price. This implies that they will get about 20% more "manna" every time the price of IT declines, as long as this does not increase their other costs in an offsetting way. If their demand elasticity is higher as well, as some evidence suggests, then this benefit will grow even larger over time.

Examining the productivity impact of various combinations of decentralization and IT can help one to distinguish among some of these explanations. In particular, if decentralized firms’ higher IT investments are indicative of their actual benefits, then the output elasticity of IT should also be higher for these firms. In contrast, the hypotheses that decentralized firms’ higher IT investments are symptoms of mismeasurement, wasteful duplication, or lower adjustment costs each imply that the output elasticity of IT will not be proportionately higher in decentralized firms.

Differences in output elasticity will not necessarily rule out the existence of unmeasured, third factors. It is possible that a third factor, which is correlated with productivity, may also be disproportionately associated with firms that adopt both IT and decentralization. However, it is not sufficient that they simply adopt more of everything; for this result to be consistent with our analysis, high productivity in some firms must also be correlated with decreased adoption of IT and increasing which is a much more difficult set of conditions. Second, investing in more centralized computers (e.g. mainframes) and undergoing wrenching organizational changes does not appear to be a rational way to consumer excess cash flow which makes stories of managerial agency less plausible as a source of spurious correlation (see e.g. Brynjolfsson, Bresnehan and Hitt, 1998).

Ultimately, statistical evidence regarding patterns of IT adoption and productivity must be combined with evidence from other sources, such as cases and theory, to build a comprehensive understanding of IT’s effects.

To examine whether the measured output of IT varies with organizational structure, we adopt the standard approach employed to study the effects of inputs, such as R&D on firm performance (Griliches, 1986). This framework has also been applied in previous work on the relationship between IT and output (Brynjolfsson and Hitt, 1995; 1996; Lichtenberg, 1995).

Several variations of a production function were estimated relating firm (i) value added (VA) to three inputs: IT Capital (C), Non-IT Capital (K) and Labor (L), industry (j) and time (t). Assuming that the Cobb-Douglas form can suitably capture the production relationship, the equation is:

From this specification, we can derive an estimating equation by taking logarithms and appending an error term with the usual OLS properties:

Similar to our earlier assumption, we assume that work systems are not easily changed in the short run and thus is not influenced by unexpectedly high or low productivity in any year; thus ORG can be considered exogeneous in this analysis. Work systems will be incorporated in two ways. First, to examine the overall effect of the work system on productivity, the work system variable can simply be entered additively into the equation. Second, an IT-org interaction variable can also be added to the equation. To enable the coefficient on the IT-org interaction to be interpreted as an increment to the baseline IT elasticity, we center IT in the interaction (dropping time and firm subscripts):

A comparison of a baseline production function regression and regressions adding org variables is shown in Table 4a. In these analyses, the cross-sectional data on org and other organizational factors is matched to an eight-year panel of productivity data. In the basic regression, the output elasticity of IT is clearly positive (t=5.9), the magnitudes of all coefficients are comparable with previous estimates using different a firm-level data set (Brynjolfsson and Hitt, 1996), and the overall regression appears to explain a substantial amount of the variance in output (although much of this explanatory power is due to the labor input term). More importantly, the interaction between IT and organization is significant and in the expected direction. When the org variable is entered, it has a positive effect on productivity, and the direct effect of IT drops by about 10%. This suggests that decentralized firms tend to be somewhat more productive, and that a portion of the output elasticity attributed to IT in the standard production function may in fact be due to organizational factors.

We are particularly interested in the way organizational design affects the output elasticity of IT. A key finding of the production function analyses is that when org is interacted with IT and added as a regressor, the effect is economically large and statistically significant (t=2.8). To interpret this regression, note that the org variable is standardized to mean zero and standard deviation one, while the IT term in the interaction is not standardized. Thus, the coefficient on the ITxorg term can be interpreted as the incremental IT elasticity of IT for a one standard deviation increase in org. We find that a one standard deviation increase in org increases the elasticity of IT by about 33%. As in the demand analysis, results are similar if we use industry share controls rather than industry dummies: the coefficient rises even further to .0157 (44% above the base elasticity).

This indicates that firms adopting both high levels of IT use and more decentralized work systems have disproportionately greater output than firms that just adopt one or the other. Centralized firms do not increase their output as much as decentralized firms do when they boost IT levels.

This provides further evidence that decentralized firms are more IT-intensive than centralized firms and suggests an explanation for this empirical result. The increase in IT’s output elasticity is not consistent with the hypothesis that the higher IT levels that we found in decentralized firms is a measurement artifact or represents wasteful duplication. The results are also not supportive of the hypothesis that decentralized firms have invested more in IT because of lower adjustment costs – there is no evidence that the marginal product of IT is lower in decentralized firms. If anything, the marginal product of IT appears to be higher in decentralized firms since the output elasticity increases more than the factor share with decentralization (33% vs. 9% for a one standard deviation increase in org). This is consistent with the hypothesis that decentralized firms receive greater benefits from IT per unit of investment, rather than simply being larger IT users. Alternatively, the observation that the output elasticity increases even more rapidly than the IT factor share for decentralized firms may be evidence that these firms tend to have more unmeasured, complementary investments associated with each dollar of measured IT capital. Under either interpretation, the high measured marginal product of IT, particularly in high org firms, is not necessarily inconsistent with equilibrium.

In each of the years between 1987 and 1994, a disproportionate number of the more decentralized firms also have high IT intensity, and vice-versa. Interestingly, in 1987, when the productivity differences are most pronounced, we see only slight evidence that firms are disproportionately locating on the diagonal, based on a Pearson chi-square test of association. However, by 1994, firms are more clearly lining up on the diagonal, but the productivity differences are less substantial. One interpretation of this shift is that between 1987 and 1994 many firms adjusted their IT levels to better match organizational designs, thereby capturing additional productivity benefits.

To test whether this positive interaction of org is unique to IT or is present for all input factors, the model is reestimated, including interaction terms with ordinary capital and labor. Both of these terms are insignificant and negative, while the IT interaction rises noticeably (Table 4a, column 4). This supports the hypothesis that the effects of work systems are unique to the IT term. However, because of multicollinearity between the interaction terms, the point estimates should be interpreted with caution.

The fact that multiple waves of the survey were sent out and initially analyzed separately can provide a second robustness check and help guard against data-snooping biases. Several seminar presentations of the main results in this paper were given before the third wave of data was received and reported similar reasoning and results. More formally, when additional terms are added to capture survey differences (org x Survey ID and IT x org x Survey ID) the two added terms are not significant individually or jointly [F(2,2233)=1.22, p<.30]. This suggests that the results are equally strong "out of sample" and that differences among the surveys have not influenced the results.

As noted above, the team variable had the highest factor loading in the org construct. Thus, as a third robustness check, we replaced the org variable in the production function by three dummy variables representing no use of teams, slight to moderate team use, and heavy use of teams as a proxy for decentralized work systems. In our sample, 39% report no use of teams, 50% report slight to moderate use or teams, and 11% report heavy use of teams. Estimating the equation using multiple dummy variables avoids the econometric assumptions involved in aggregating discrete variables to form a scale, such as org, although at the cost of some loss in statistical power. The results are broadly similar to those with the org variable, although it appears that teams by themselves may have a negative impact on productivity before the interaction with IT is considered (Table 4b). As with the org variable, the interaction between teams and IT is positive and significant. The elasticity of IT increases monotonically with increased adoption of teams, with the two ends of the scale being statistically different, t = 2.6).

As a fourth robustness check, we consider the possibility that different types of information technology may have a disproportionate effect on output. As with our correlational and demand analysis, we find that the type of IT investment considered (PCs versus mainframes) has no effect on our results.

One difficulty of using the three input Cobb-Douglas production function is that the overall explanatory power in the regression is dominated by the input terms, particularly labor. Furthermore, it is quite possible that labor is at least partially endogenous in this regression since firms with unexpectedly high or low demand may be able to adjust their labor input within a single year. This would lead to a correlation between labor and the error term in the productivity regressions and possible unknown biases in the other estimates. To correct for this possibility, we estimate a partial productivity specification that fixes the elasticity of labor at its theoretical value as well as a "semi-reduced form" regression where labor is dropped entirely (Griliches and Mairesse, 1984). The results of this analysis are similar to those reported earlier although somewhat stronger in terms of statistical significance.

As a final robustness check, we examine the key assumption that a firm’s organizational characteristics remain fixed over our sample period, or that at least any biases introduced by changes in work systems over time are small. Approximately 29% of the sample reported that some sort of internal restructuring occurred within our sample period. However, based on content analysis of change descriptions, 8% (33 firms) reported a change toward team-based organization and only one firm reported a move toward centralization. Most of the other changes were more localized, such as changing an executive. This suggests that while restructuring was occurring during our sample period, work system change is a gradual process.

To examine the effect of work system changes econometrically, we compare firms that report a restructuring of any sort during our sample frame in three ways. First, the analysis is conducted separately for firms that restructured and those that did not. Second, the two samples are pooled and analyzed with separate coefficients on the IT-org interaction for the restructuring and non-restructuring group. Finally, the panel structure of our data is used to refine the estimate by adding two interaction terms, one that includes all the points post-restructuring (and zero for firms that did not restructure or for points in restructuring firms before the event) and a second interaction that includes all the pre-restructuring points (and zero following restructuring). Thus, the sample is split into two parts, but the split point (in time) potentially occurs at a different point for each firm. In all three analyses, the results are consistent. The magnitudes of the interaction terms do not differ significantly from the earlier estimates in any analysis and are not significantly different from each other. Again, this suggests that the results do not appear to be substantially biased by within-sample changes.

In order for the production function analyses to have any statistical power, it must be possible to observe firms with different combinations of IT and work practices. There are two possible reasons why we could expect to find different combinations in practice: some firms are not fully optimizing, or firms face different conditions.

There is reason to believe both effects are present in practice. Case studies have shown that neither practitioners nor researchers fully understand the links between IT and organization. This can lead firms to be late adopters or make incorrect adoption choices. Many of the technologies and applications are new and have complex organizational implications and so their effects are not well understood. As one CEO succinctly stated, "The benefits of information technology occur in many unforeseen directions."

Firms may also face very different costs of organizational change, so some adopt these changes faster than others. As discussed above, firms that have highly skilled, educated, or white collar work forces may be able to adapt better to organizational change (Bartel, Lichtenberg and Vaughan, 1989). From our earlier analysis, we found that existing levels of human capital are correlated with IT use.

Accordingly, we treat IT and org as endogenous and use skill levels, education, and work force composition as instruments. In addition, time and industry are included as well as their interaction with skills and education. Interestingly, the IV regressions (Table 4c), the coefficients on both IT and its interaction with org rise substantially. This effect appears to be present whether or not capital is assumed to be exogenous or endogenous and whether or not we use lagged input levels as instruments. While it is still possible that these instruments are also correlated with unobserved drivers of IT investment and possibly ORG, we have no evidence that these biases, if present, would change the conclusions of our analyses.

This paper argued that the diffusion of low-cost IT may be related to a substantial shift in the economics of organizations toward greater use of teams and more decentralized decision-making. Using data on organizational practices for approximately 380 large firms, we find two types of support for this hypothesis: 1) firms that report the use of a decentralized work system have higher demand for IT; and 2) firms that use decentralized work systems show greater output from their IT investments. These findings are robust to alternative work system measures, IT measures, demand specifications, and productivity specifications.

A parsimonious explanation of our results is that IT is most effective in decentralized work systems. However, several alternative explanations are also possible. One possibility is that unobserved factors, such as "management" or "business opportunities" are the true sources of the differences in demand and output elasticity that we find. For instance, suppose that "good" managers are of two types, those with a taste for both IT and decentralization and those with a taste for neither. If this were the case, then it is possible that IT and org could covary in the data and also exhibit the observed effect on output, but they would not necessarily be complements.

Another possibility is that the IT is most valuable in the presence of other unmeasured organizational characteristics, such as information assets or worker expertise (Brynjolfsson and Yang, 1997). In turn, these hidden organizational characteristics may be complementary to decentralized work systems. In this case, it would be inappropriate to attribute all of the impact of such a system (IT, decentralization, information assets) to the subset of components that are measured in this study. However, the underlying argument and impacts are very similar: certain types of firms (e.g., those rich in information assets) benefit disproportionately from the fall in IT price and these firms tend to use decentralized work systems.

In summary, this study provides some of the first large-sample evidence of a broad relationship between decentralized work systems and IT use. The empirical facts uncovered by this study are exploratory rather than confirmatory. They can provide both guidance and constraints for theories about the interaction of IT and organizational design, but do not offer definitive "tests." For a "general purpose technology," such as IT, even more than for most other inputs to production, the specific circumstances of implementation and the idiosyncrasies of particular firms may significantly affect the way the IT is used and its effects on the organization of work. However, the data do provide more support for those theories that predict that lower information processing and communications costs will typically be especially beneficial for organizations that distribute decision-making.

While our study focused on how organizational factors influence the adoption and productivity of IT, our results also imply that organizational design can and should change to fit with the technology. In particular, certain types of organizational structures were found to use IT more intensively than others. Because the quality-adjusted price of IT is likely to decrease by an order of magnitude over the next a decade, IT-intensive organizational forms can be expected to benefit disproportionately in the future.

Table 1: Summary Statistics (Year End 1994) - Sample Means

Note: Dollar figures are in millions.

Table 2a: Correlations between IT and Decision Authority

Table 2b: Detailed Questions on Decision Rights

Question: "Who ...? Would you say exclusively managers, mostly managers, equally workers and managers, mostly workers or exclusively workers?"

Spearman partial rank order correlations controlling for industry, employment and production worker occupation. N=270-276, depending on data availability.

Table 2d: Unrotated Principal Components for org Variable Construction

Table 2e: Correlation of org with various measures of IT, varying control variables

Spearman rank-order correlation. All coefficients significant at p<.01.

Table 3: IT Demand Equations varying industry and human capital controls

Key: * - p<.1, ** - p<.05, *** - p<.01

Dependent variable is log(IT); ORG and Production Worker Skill standardized to mean 0, standard deviation 1.

Table 4a: Productivity Estimates

Key: * - p<.1, **- p<.05, *** - p<.01

Table 4b: Productivity: Organization as Self-Managing Teams

(Reference point is firms that do not use teams)

Key: * - p<.1, **- p<.05, *** - p<.01

Table 4c: Productivity Estimates: Instrumental Variables

Key: * - p<.1, **- p<.05, *** - p<.01

Table 5a: 2x2 Grid of IT (1987) vs. Organization: Productivity Comparisons

Standard Error in parenthesis

Table 5b: 2x2 Grid of IT (1994) vs. Organization: Productivity Comparisons

Standard Error in parenthesis

Table 5c: 2x2 Grid of IT (1994) vs. Organization: Productivity Comparisons

Matched Sample to 1987

Standard Error in parenthesis

Figure 1: Scree plot of work systems principal components

Figure 2: IT-demand by organizational type