Printer Friendly

IT and productivity in U.S. manufacturing: do computer networks matter?

I. INTRODUCTION

How computers affect productivity is a long-standing research question. Many recent studies argue that information technology (IT), particularly computers, is a significant source of U.S. productivity. The specific mechanism remains elusive. Detailed data on the use of computers and computer networks have been scarce. This article uses new plant-level data on computer networks collected by the U.S. Census Bureau to estimate the effect of computer networks on labor productivity across U.S. manufacturing plants. The Computer Network Use Supplement (CNUS) to the 1999 Annual Survey of Manufactures (ASM) focused on the use of computer networks, rather than the presence of computers alone. Linking the CNUS data to current and previous information for the same plants collected in the 1999 ASM, and the 1997 and 1992 Census of Manufactures (CM), allows us to examine the relationship between productivity and the use of computer networks.

Our work differs from others in several important aspects. First, this is the first study linking the use of computer networks to labor productivity, using data for approximately 30,000 plants representing the U.S. manufacturing sector. Most previous studies examining the link between productivity and computers or other IT in the United States focus on the presence of computers, using either data on the stock of computer capital or on current IT or computer investment as proxies for the computer stock. Only one previous study, McGuckin et al. (1998), examined the link between productivity and how computers were used. That study was limited to five U.S. two-digit manufacturing industries covered in the 1988 and 1993 Surveys of Manufacturing Technology (SMT) collected by the U.S. Census Bureau and did not separate the use of computer networks from other uses of computers and advanced technologies.

Second, we extend the existing model relating IT to productivity by including materials as a separate input. Our dependent variable is a gross-output measure of labor productivity. Although gross output is an appropriate measure of the theoretical output, particularly at the plant level, most previous plant-level studies for the United States use a value-added measure and exclude materials as a factor input, possibly making their results subject to omitted variable biases.

Third, we model the probability that a plant has a computer network as a function of its performance and conditions in prior periods. This probability is of interest in its own right. It also makes possible the fourth distinguishing feature of our work, testing for possible endogeneity problems associated with the computer network variable. If good plants are more likely to have computer networks, we must account for these characteristics to get a more accurate estimate of the effect of networks.

Our research has five principal findings. First, average labor productivity is higher in manufacturing plants with networks than in plants without networks. Second, computer networks have a positive and significant effect on labor productivity after controlling for other important factors, such as capital intensity and other plant characteristics. Third, the choice of theoretical model has empirical consequences. Previous studies using value-added models appear to overstate the effect of IT on productivity by factors of two to three. Fourth, plants with lower relative productivity in previous periods are more likely to have computer networks. Fifth, computer networks have a positive and significant effect on labor productivity even after taking account of possible endogeneity of the computer network variable.

II. BACKGROUND

Computers play an important role in the strong economic performance of the U.S. economy according to many recent studies. This role is particularly important during the surge of productivity growth in the late 1990s as discussed in Oliner and Sichel (2000), Jorgenson and Stiroh (2000), Jorgenson (2001), Stiroh (2001), and Triplett and Bosworth (2000). (1) For example, Jorgenson and Stiroh (2000) find relatively higher growth in total factor productivity (TFP) and average labor productivity between 1958 and 1996 in manufacturing industries. Within manufacturing, the annual growth rates of average labor productivity for computer-producing industries are far higher than for other industries: 4.1% for Industrial Machinery and Equipment (Standard Industrial Classification [SIC] 35) and 3.1% for Electronic and Electric Equipment (SIC 36), compared to 2.6% in the next highest industry, Instruments (SIC 38).

Triplett and Bosworth (2000) report similar findings for TFP and labor productivity growth over three periods between 1960 and 1997. Productivity growth by either measure is far higher in manufacturing than in other industries during the two most recent periods (1973-97 and 1987-97) and is particularly pronounced for Electronic and Electric Equipment. That industry's TFP growth of 7.3% between 1987 and 1997 far exceeds the rate of 2.4% for durable goods manufacturing; 2.4% (also) for all manufacturing industries; 0.5% for services; -0.5% for finance, insurance, and real estate; and 0.9% for the private sector as a whole. Several mechanisms might link computers and productivity.

Computers may affect productivity because they are a specific capital input to the production process. This is the approach taken in most existing studies, including both the national and industry-level studies just cited, as well studies at the plant or firm level, such as Brynjolfsson and Hitt (2000), Dunne et al. (2000), Stolarick (1999), and McGuckin et al. (1998). Consider a steel mill. Computers and automated processes are used to control production processes in modern steel mills. Many supporting business processes also can be computerized. For example, computers can be used to maintain a database of customers or shipments or to do accounting or payroll. Computers may substitute for paper-based systems without changing the underlying business processes.

But computers may also be used to organize or streamline the underlying business processes. When these computers are linked into networks, they facilitate standard business processes, such as order taking, inventory control, accounting services, and tracking product delivery, and become electronic business processes (e-business processes; see Atrostic et al., 2000). These e-business processes occur over internal or external computer networks that allow information from processes to be exchanged readily. Shipments may be tracked online, inventories may be automatically monitored, and suppliers may be notified when predetermined levels are reached.

Adopting e-business processes automates and connects existing business processes. It can also change the way companies conduct not only these processes but also their businesses. The surge of interest in supply chains exemplifies this potential for computers to affect productivity outside of the manufacturing industries that produce them. These effects are thought to occur through organizational change. Many core supply chain processes are widely cited as examples of successful e-business processes that in turn are expected to shift the location of the process among the participants in the supply chain. Brynjolfsson and Hitt (2000) argue that the effects of organizational changes may rival the effects of changes in the production process. Viewed this way, computer networks are an example of what Breshnahan and Trajtenberg (1995) term a productivity-enhancing general-purpose technology.

Although computers and computer networks clearly may have distinct effects on plant-level productivity, few previous microdata studies assess the effect of computer networks. Most evaluate the effect of computers alone, using either data on book values of computer capital, or current investment in IT or computers, as a proxy for the computer capital stock. Motohashi (2001) is one of the few studies that has firm-level data on computer networks. Although McGuckin et al. (1998) touch on the link between productivity and how computers are used in

selected U.S. manufacturing industries, their study does not separate the use of computer networks from other uses of computers and advanced technologies.

III. THE MODEL

We begin our examination of the effect of computer networks by specifying a model of how computer networks affect labor productivity. Understanding which plants are likely to have computer networks is also of interest, so we specify a model of the probability that a plant has a network in the current period, given conditions in prior periods.

Labor Productivity and Computer Networks

We specify a three-factor Cobb-Douglas production function

(1) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII.]

where Q, K, L, and M denote output, capital, labor, and materials. The parameters [[alpha].sub.1], [[alpha].sub.2], and [[alpha].sub.3] represent output elasticities of capital, labor, and materials. A is the usual "technological change" term, which is specified as a function of computer networks (Network), such that

(2) A = exp([[beta].sub.0] + [[beta].sub.1] Network).

Equation (2) is based on the idea that a plant using a computer network is likely to produce a higher level of output than its counterpart that does not have a computer network. We assume that networks indicate "disembodied technical change" that is not captured by the empirical measures of K, L, and M. Computer networks of course could be considered part of the plant's capital, K, because they are an investment like any other. If data were available separately on the investment or service flows from investments in networks, Stolarick (1999) notes that separate capital factors could be created. However, we have no data on the separate kinds of capital stocks, only data on total capital and on the presence of computer networks. We incorporate our information on presence of networks into the technological change term A. This approach is also taken by researchers such as McGuckin et al. (1998), Motohashi (2001), and Greenan and Mairesse (1996). We expect that [[beta].sub.1] is positive because computer networks should have a positive effect on the technological change term, A.

Substituting (2) into (1), dividing both sides by L, performing some algebraic manipulation, and taking logarithms on both sides, we have the following equation:

(3) Log(Q/L) = [[beta].sub.0] + [[beta].sub.1] Network + [[alpha].sub.1]log(K/L) + [[alpha].sub.3]log(M/L) + ([[alpha].sub.1] + [[alpha].sub.2] + [[alpha].sub.3] - 1)log(L).

Equation (3) directly relates computer networks to log-labor productivity. Our parameter of interest in this formulation is [[beta].sub.1]. It can be interpreted as measuring the effect of computer networks on labor productivity, controlling for capital intensity (K/L), materials intensity (M/L), and total labor (L), which in turn can be considered as a proxy for plant size. Note that if [[alpha].sub.1] + [[alpha].sub.2] + [[alpha].sub.3] = 1 (or [[alpha].sub.1] + [[alpha].sub.2] + [[alpha].sub.3] - 1 = 0), we have constant returns to scale. If [[alpha].sub.1] + [[alpha].sub.2] + [[alpha].sub.3] is less (greater) than 1, we have decreasing (increasing) returns to scale.

Our model differs from those used in previous studies. Ours is a three-factor production function using gross output as a proxy for output and incorporating materials as a separate productive input. Previous studies on IT and productivity such as McGuckin et al. (1998), Greenan and Mairesse (1996), and Brynjolfsson and Hitt (2000) used value added as a proxy for output and excluded material input from their models.

Gross output is preferred to value added as a proxy for the theoretical output. Baily (1986) shows that using a value-added model yields systematically biased estimates of the theoretically correct TFP. Value-added productivity measures are common in estimations using aggregate data because they net out potential double counting in aggregate gross output measures. (2) It is generally accepted that particularly at the plant level, gross output is an appropriate measure of "output" (see McGuckin and Nguyen, 1993). Parameter estimates from value-added and gross-output specifications have predictable relationships, if neoclassical assumptions hold as Stiroh (2002) notes. We estimate the value-added empirical model in the next section to allow comparisons with previous studies.

Probability of Having a Computer Network

Consistent ordinary least squares (OLS) estimates of [[beta].sub.1] in equation (3), however, require that Network be uncorrelated with the error term of the equation. Such an assumption is unsatisfactory.

Previous research suggests both positive and negative correlations between Network and the error term. McGuckin et al. (1998) find that adopting advanced communication technologies, of which a computer network was one, is positively correlated to plants' performance. That is, good plants are most likely to be subject to endogeneity biases. In contrast, Stolarick (1999) finds that less productive plants appear to invest in computer networks to try to catch up. More generally, strong cross-section effects often become muted after controlling for prior conditions, as noted in Griliches and Mairesse (1995). Finally, Atrostic and Gates (2001) present simple tabulations showing that although networks are pervasive across manufacturing, diffusion among three-digit subsectors ranges from 30% to 70%.

To allow for possible endogeneity, we postulate that having a network in the current period is related to the plant's prior performance and conditions, such that

(4) [Network.sub.t] = [[gamma].sub.0] + [ X.sub.os] [delta].

Equation (4) relates computer network use in the current period (t) to a vector of predetermined variables, [X.sub.o], representing the plant's performance and conditions in an earlier period (s).

IV. EMPIRICAL SPECIFICATIONS

Implementing our theoretical model depends on making the best use of the available data. We account empirically for important plant characteristics that may significantly affect a plant's labor productivity but are not in our theoretical models. Our specification also reflects the fact that data on critical theoretical variables are not all available for the same year. Our final specification is given in equation (5):

(5) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII.]

SKILL is worker skill mix. MULTI denotes the plant that belongs to a multiunit firm, RLP is the plant's labor productivity in a prior period relative to its industry, and IND represents the plant's industry. Suffixes of 92, 97, and 99 denote whether data on the variable are available in 1992, 1997, or 1999. A dummy variable, NEW, denotes plants that did not exist in 1992.

In the estimating equation (5), L enters the denominator of the dependent variable. It also enters the denominator of two of the independent variables, K/L and M/L, and it enters by itself as an independent variable. If L is measured with error, the coefficient estimates of the equation will be biased. We shall discuss this issue further in the next section.

Theory and the empirical literature suggest that computer networks are likely to be endogenous. We address potential endogeneity through a two-step procedure that yields consistent estimates and describe that process in the first part of this section. Data and details of how we specify variables empirically are discussed in the final two parts of this section.

Econometric Methodology

We can estimate the effect of Network on plants' labor productivity by applying OLS to equation (5). However, the potential endogeneity of computer networks means that OLS estimates are likely to be subject to selection bias. To correct for this bias, following Heckman (1979), Maddala (1983), and Greene (2000), we use a two-stage estimation process that yields correct estimates. The probability that a plant has a computer network can be estimated using a reduced form probit regression:

(6) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII.]

Network is observed in 1999. The explanatory variables in (6) are measured in a prior period, 1992, for most plants. For plants new since 1992, the prior period is 1997.

In a nutshell, equation (6) postulates that having a computer network in 1999 depends on the plant's performance and conditions in the prior 1992 period. Performance is measured by the plant's productivity relative to its four-digit industry (RLP92). Prior conditions include its 1992 computer investment/labor ratio (COMP92), skill mix (SKILL92), industry (IND92), and whether or not the plant belongs to a multiunit firm (MULTI92).

We construct an instrumental variable, Pr(Network99), for the probability that a plant has a computer network in 1999. Pr(Network99) = R(-CNHA T99), where R is the cumulative density function for the standard normal variable and CNHAT99 is the fitted value of Network99. Using this instrumental variable in estimating equation (5) takes into account possible endogeneity problems associated with the Network99 variable.

Maddala (1983) points out that the standard errors of the estimates of the second-stage of this two-stage procedure are incorrect because an explanatory variable, Pr(Network) in this case, is estimated. To obtain the correct standard errors, we apply STATA's treatreg procedure with the robust option. The robust variance estimation procedure was developed by Huber (1967) and White (1980, 1982).

New Data on Computers and E-Business Processes in U.S. Manufacturing

The data used in this study are taken from the CNUS to the 1999 ASM. (3) The CNUS focuses on the use of computer networks. In June 2001, the U.S. Census Bureau released an analytical report based on the 1999 CNUS and the 1999 ASM (E-stats, available online at www.census.gov/estats). The E-stats report shows that manufacturing plants use networks for much more than online sales and orders. Only half of manufacturing plants reporting that they have a network also report that they accept and/or place orders online.

Atrostic and Gates (2001), using the new 1999 CNUS data, find computer networks widely diffused within manufacturing, with networks at 52% of plants. Plants with networks are slightly more common in nondurables (54% of plants) than in durables (51%), but the share of employment at plants with networks is almost identical--76% in nondurables and 75% in durables. Within each subsector, diffusion rates range from lows of 27.1% in Apparel and 35.3% in Furniture to highs of 71.1% in Chemicals and 72.2% in Electrical Equipment. These estimates are based on plant-level responses but calculated from data aggregated to a subsector level, and their analysis does not address productivity.

Because the data are only from respondents to the CNUS, and are not weighted (see the discussion at www.census.gov/estats), our results may apply only to responding plants. However, the plants included in our sample account for a substantial share of the U.S. manufacturing employment and output (about 50% and 60%) represented in the ASM.

Variables in the Productivity Equation

* Gross output (Q): The total value of shipments (TVS) is a proxy for Q.

* Value added (VA): Value added is defined as Q - M where M is material input as defined below.

* Labor (L): Labor is total employment (total number of employees in the plant).

* Materials (M): Materials are the sum of values of materials and parts, values of energy consumed (including electricity and fuels), and values of contract work.

* Labor productivity (LP): Both gross output (Q) and value added (VA) are used. Gross output labor productivity = Q/L, and value added labor productivity = VA/L.

* Computer Networks (Network): Network is the key explanatory variable in this study. It takes on a value of 1 if the plant reported having a computer network, and 0 otherwise. About 88% of the plants responding to the CNUS used computer networks (see Table 1).

* Capital (K): Because data on capital services (K) are not available, we use book values of gross capital stocks (including buildings and machinery assets, collected in the 1997 CM) as a proxy. (4) Book values are likely to be subject to important measurement errors. First, the data reported in book values do not accurately reflect the value of capital. Second, using a simple sum of buildings and machinery assets as a proxy for K assumes that these components of capital are homogenous. This assumption is obviously incorrect. Third, there is no adjustment for differences in the quality of capital. Fourth, there is no adjustment for intensity of use. Although we recognize these limitations, it is difficult to see how the problems could have been handled in the context of cross-sectional analysis. As a practical matter, we follow previous studies such as McGuckin et al. (1998) and Greenan et al. (2001) and use book values of the capital as a proxy for K. This implies that services are proportional to book value of capital. This assumption appears to be reasonable given that we control for plant characteristics in our regressions.

* Skill Mix (SKILL): Skill mix is defined as the number of nonproduction workers (OW) divided by total employment (TE) in the plant, as reported on the 1999 ASM. The expected sign may be positive or negative. Computer networks require highly skilled workers to develop and maintain them. Productivity might thus be higher at plants with a higher proportion of skilled labor because these workers are able to develop, use, and maintain advanced technologies, including computer networks. But applications such as expert systems may allow a function to be carried out by employees who have lower skill levels or with fewer employees.

Occupational detail would be desirable to test the relationship among productivity, networks, and the presence of such skilled occupations as computer programmers and systems support staff, as reported in Greenan et al. (2001) and Motohashi (2001). However, the ASM only collects information on the total numbers of production and nonproduction workers in the plant, with no further detail by process, function, or worker characteristic. Dunne and Schmitz (1992) found that plants in the 1988 SMT that used advanced technologies had higher ratios of nonproduction to total workers. As with many other plant-level studies, including McGuckin et al. (1998) and Dunne et al. (2000), we use this employment ratio to proxy for skill mix in our productivity estimates. Production workers accounted for about one-quarter (27%) of employment among CNUS respondents in the manufacturing sector. This share is similar to the shares that McGuckin et al. (1998) reported for the five two-digit U.S. SIC industries in the 1988 and 1993 SMTs.

* SIZE: Plant size is specified as total employment: L, the number of workers. This measure is consistent with our theoretical specification (3) and is used by such other researchers as Greenan and Mairesse (1996).

Because L enters the denominator of both dependent and independent variables, the parameter estimates of the equation would be biased if there were significant measurement error associated with L. However, L (the number of workers) is known to be one of the most accurately measured variables in any business data set collected by the Census Bureau. We therefore do not expect a significant measurement error in L relative to other variables, such as capital and materials. We addressed this potential concern in our preliminary work by estimating the model using different proxies for the size variable. All proxies yield similar results. We present the result based on L here because its estimated coefficient can be used to make inferences regarding returns to scale. (5)

* Multiunit firms' plants (MULTI): Many manufacturing plants are part of a multiunit firm, so employment size alone is an inadequate indicator of available resources, managerial expertise, and scale. We construct a dummy variable, MULTI, that takes on the value of 1 if the plant is part of a multiunit firm, and equals 0 otherwise. Nearly two-thirds of the plants in our sample are part of a mulitunit firm.

* Industries (IND): All previous studies of plant-level behavior note substantial heterogeneity among plants within detailed manufacturing industries, as well as between detailed industries. There are 21 three-digit NAICS manufacturing industry groups in our sample (NAICS codes 311-316, 321-327 and 331-337). Industry dummies (IND) are included in the empirical estimates to capture industry-specific effects on plant-level labor productivity. (6)

* New Plants (NEW): New plants are those plants did not exist in 1992. NEW is interacted with the other independent variables to allow for the possibility that the estimated impacts of these variables differ for new plants. (7)

Variables in the Probit Equation

The dependent variable in the network probit equation (6) is Network. We hypothesize that having a network in 1999 is a function of the plant's condition in prior periods. The prior period for our explanatory variables is 1992.

Requiring data from prior periods reduces to 10,496 the number of observations in the network probability estimates. This is about 30% of the 29,840 observations in the 1999 cross-section gross-output productivity estimates. Computer expenditures data are not available for plants that opened after 1992, and a number of plants that existed in 1992 did not report their expenditure on computers. Stolarick (1999) reports a similar drop in sample for 1992 because plants did not respond to the computer expenditure question.

One facet of the plant's prior condition is its performance. We specify a relative performance measure, RLP92, that is the plant's labor productivity in 1992 relative to the average for its four-digit SIC industry. Using a relative productivity measure means we do not need to develop deflators for each plant for the 1992-99 period.

Previous studies suggest that capital intensity and skill mixes are correlated with use of computers. We measure capital intensity as K/L92, the plant's capital/labor ratio in 1992, and skill mix as the share of nonproduction workers in the plant's total employment in 1992. Spending on computers in previous periods is an important component of its prior condition, and affects the likelihood that it has computers to network. We specify prior computer spending, COMP92, as computer expenditures per employee in 1992. We also control for whether the plant was part of a multiunit firm, MULTI92, and its two-digit SIC industry, in 1992.

V. EMPIRICAL FINDINGS

Average labor productivity is higher in plants with computer networks. Table 1 shows that average labor productivity is about 30% higher in plants with computer networks. The 30% differential holds for both gross output and value-added labor productivity measures. The productivity differential varies among industries, according to estimates not reported in Table 1. However, for any industry, the differential is about the same for both productivity measures.

The literature provides few points of comparison. McGuckin et al. (1998) find average labor productivity is higher in the five SMT manufacturing industries for plants using any of 17 advanced technologies (including local area networks and intercompany computer networks). Differentials range from about 20% for plants using between one and three technologies, compared to plants using none, to 80%-100% for plants using 13-17 technologies. However, differentials are not reported separately for any of the technologies.

Conclusions cannot be based on simple averages because these statistics do not control for other factors. Econometric estimates of equations (5) and (6) allow us to assess the effect of computer networks on productivity while controlling for other important factors, such as capital intensity, skill mix, and industry, and the potential endogeneity of the computer network variable.

The Effect of Computer Networks on Plants' Labor Productivity: OLS Estimates

Column 1, Table 2, reports OLS estimates of equation (5). Labor productivity is higher in plants using computer networks, controlling for skill mix, capital intensity, materials intensity, being part of a multiunit firm, and industries. The coefficient of 0.037 is significant at the 1% level, indicating that labor productivity in plants with computer networks is about 3.7% higher than that in plants without computer networks.

Our finding of a positive and significant relationship between computer networks and plant-level productivity in U.S. manufacturing is consistent with the few other studies addressing this relationship in the United States or other countries. McGuckin et al. (1998) use 1988 and 1993 SMT data to examine the relationship between the use of advanced technologies and labor productivity in five U.S. manufacturing industries. (8) They find that productivity is higher at plants using advanced technologies, even after controlling for multiple economic characteristics of the plant. The relationship holds both in terms of the number of technologies used and in the intensity of that use. Most relevant to our study is their finding that labor productivity in 1993 is about 12% higher for plants using one group of technologies, computer networks and other communication and control equipment, controlling for other plant characteristics and use of other technologies.

Related studies include Greenan and Mairesse (1996) and Motohashi (2001). Greenan and Mairesse analyze the effect of using computers (networked or not) in French manufacturing and services firms in 1987, 1991, and 1993. They find that an effect of about 20% might be conservative. Motohashi (2001) is one of the few studies with firm-level data on computer networks. That study analyzes the effect of computer networks in the manufacturing, wholesale, and retail sectors in Japan in 1991. For firms with networks, the estimated effects on productivity vary with the type of network and the e-business processes in which it is used. Motohashi (2001), like Brynjolfsson and Hitt (2000), finds that IT affects TFP only in firms with higher human capital and flatter workforce organization.

Expected relationships with many other explanatory variables hold in our estimates. Capital intensity, material intensity, and skill mix all are positively and significantly related to labor productivity, with coefficients of about 0.08, 0.45, and 0.03. The positive relationship with skill mix supports the hypothesis that productivity is linked to the use of new production processes, including use of computer networks, which require skilled workers.

The relationship of firm and plant size to productivity is more complex. Being part of a multiunit firm matters. Productivity in plants that are part of multi-unit firms is about 9% higher than in single-unit plants, controlling for networks, skill mix, capital intensity, materials intensity, and size (column 2). However, controlling for being part of a multiunit firm, we find a negative but only weakly statistically significant estimate for the size variable, with a value of -0.005. This implies a coefficient of returns to plant scale of 0.995, slightly less than 1. (9)

Plants that performed well in 1992 and survived to 1999 have much higher labor productivity. The coefficient of 0.276 is significant at the 1% level.

The coefficient of the NEW dummy variable and interactions between NEW and selected explanatory variables (which allow their estimated impacts to differ for the new plants) are of interest in their own right. The vintage capital model suggests that new plants open with the newest technologies and management processes, and existing plants exit when their productivity becomes too low relative to the new entrants. The empirical literature surveyed in Bartlesman and Doms (2000) yields mixed findings about this model.

The coefficient of NEW, by itself, supports the vintage model. The coefficient, 0.203, is positive and significant at the 1% level. However, most of the interaction terms between NEW and other variables, when they are statistically significant, are negative. This suggests a complex relationship among plant age, its use of technology, and productivity.

Probability of Having A Computer Network: Probit Estimates

The sparse empirical literature contains mixed results about the link between performance in prior periods and the use of a computer network in the current period. Our probit estimates, reported in Table 3, support the hypothesis that less productive plants are more likely to adopt a computer network in attempts to boost their productivity. These estimates are consistent with Stolarick (1999). The marginal effect of plants' prior relative labor productivity (RLP92), -0.016, is significant at the 1% level.

The marginal effect of skill mix in the previous period is not statistically significant at the 10% level. However, our proxy for skill mix (the ratio of nonproduction workers to total employment) may not accurately reflect the true skill mix. Indeed, nonproduction workers in the CM data include all types of white-collar workers, such as managers, engineers, and technical workers, as well as other office workers.

Plants belonging to a multiunit firm in the prior period are more likely to have a computer network than a single-unit firm. The marginal effect, 0.067, is significant at the 1% level. Investing in computers in the prior period does not necessarily imply that the plant will have a network. The coefficient for the computer expenditures variable is positive, but it is not statistically significant at the 5% level.

Robustness of the Results

Our OLS estimates assume that computer networks are exogenous and that we specified our theoretical models and chose our samples correctly. This section assesses each of these assumptions in turn. We show that our estimates are robust to potential endogeneity, model misspecification, and sample selectivity.

Endogeneity. Equation (5) assumes that Network is exogenous. The potential endogeneity of Network means that OLS estimates may not be consistent and may be subject to endogeneity bias. As already discussed, we address potential endogeneity through two-stage estimations.

We first reestimate our model, replacing the Network dummy variable in equation (5) with Pr(Network), the probability estimated from a standard two-stage estimate of (6). The results are reported in column 4 of Table 2. Both equations show that labor productivity is higher in plants that have or are predicted to have computer networks. The OLS estimate of Network, 0.037 (column 1, Table 2), is significant at the 1% level. The two-stage coefficient for Pr(Network), 0.669 (column 4, Table 2), also is significant at the 1% level. This two-stage estimate suggests that, ceteris paribus, a 1% increase in the probability of a plant having a computer network would increase its labor productivity by 0.669%.

Although this two-stage estimate accounts for potential endogeneity, Greene (2000) shows that the standard errors using conventional estimation techniques may not be correct. We report in column 5, Table 2, the results of the two-stage model using STATA's robust estimation procedure that corrects the standard errors. The corrected standard errors are larger, and associated t-statistics are smaller, as expected. However, the larger standard errors do not change the significance levels of the associated estimates.

Impact of Networks in Two-Stage Estimates. We emphasize that the impacts estimated from OLS and two-stage estimates are not comparable. The variable Network in the OLS regressions is a dummy variable whose value is either 0 or 1, whereas the Pr(Network) variable is continuous and has values between 0 and 1. Interpreting the OLS estimates is straightforward: Productivity is about 4% higher in plants with computer networks. Interpreting the two-stage estimates is more complex because the estimated effects depend on which pair of plants we compare.

One way to interpret the two-stage estimates is to compare the productivity impacts of computer networks on plants at two points in the predicted probability of having a computer network. We report the results of a series of such comparisons in Table 4. To illustrate, consider an extreme case comparing plants in the 1st and 99th percentiles of the probability distribution. The plant in the 99th percentile outperforms the plant in the 1st percentile by 16%. An example closer to our data compares plants at the 10th and 90th percentiles of the estimated probability of having a computer network. (About 12% of the plants in our sample do not have a computer network.) The estimated probabilities of these plants adopting a computer network are 0.8658 and 0.9733. Using the estimated coefficient for the Pr(Network) of 0.669 from the probit regression (column 4 of Table 2), we can calculate the expected productivity difference between the two plants: 0.669(0.9733 - 0.8658) = 0.072. The productivity difference means that a plant moving from the 10th percentile (less likely to have a computer network) to the 90th percentile (more likely to have a computer network) would increase its labor productivity by 7.2%.

Empirical experience leads us to expect a smaller network effect in the two-stage estimates. We find instead that it is higher. The two-stage estimate of 7.2% is about 3.5 percentage points higher than and is nearly twice as large as the OLS estimate of 3.7%.

Choice of Theoretical Model. Our specification differs from many other studies because it uses the theoretically preferable gross-output measure of labor productivity and includes materials as a separate input. Specification differences could matter empirically because materials make up a large share of total inputs. To assess the importance of specification, we estimate a value-added specification similar to those in the literature and report the results in column 2, Table 2.

We find that differences in theoretical specifications matter empirically. The estimated impact of computer networks on labor productivity is twice as high in the value-added model. The estimated coefficient for Network is 0.075 in the value-added model, compared to 0.037 in the gross-output model. This coefficient of 0.075 is at the lower end of the range of estimates reported in previous studies using the value-added model.

Our value-added estimates are consistent with others in the literature, particularly those using U.S. Census Bureau plant-level data. The value-added estimates we report in column 2, Table 2, for U.S. manufacturing in 1999, are similar to those of McGuckin et al. (1998, table 7), for five two-digit U.S. manufacturing industries in 1993. The specification most similar to ours lacks interactions with plant age. They find that computer networks and other communication and control technologies increase labor productivity by about 12% in 1993. Our estimate is about 7.5%. Their elasticity of capital intensity is 0.144; ours is 0.130. They estimate the skill elasticity to be 0.078, and our estimate is 0.071.

Our results strongly suggest that a value-added model is subject to omitted variable bias. The coefficients of materials intensity, log(M/L), are significant and their magnitudes are virtually invariant across specifications and sample sizes. In addition, the explanatory power of the gross-output model ([R.sup.2]) is about twice that obtained from the two-factor model (0.8133 versus 0.4159). The [R.sup.2] of the gross-output model is stable across models and sample sizes (0.8113 for the OLS model on the full sample, and 0.7811 for the OLS model on the sample used in the two-stage model). These results suggest that previous studies using the value-added model overestimated the effects of computer networks and IT on productivity. The overestimation reflects the simple fact that the value of materials is a large share of the value of shipments: In our data, value of materials is 60% of the value of shipments. Stiroh (2002) obtains similar differences between gross output and value-added specifications. He notes that the capital and labor elasticities of a gross output production function should equal the elasticity from a value-added production function multiplied by one minus the material share, if the neoclassical assumptions hold. Basu and Fernald (1997) also show that value-added estimates may suffer from an important omitted variable bias. Although the differences in estimates from the models may be predictable, they are too rarely noted.

Sample Bias. Plants that responded to the CNUS are substantially larger than the typical manufacturing plant. Average employment is 223 employees in our data, compared to an average of about 45 employees for the entire manufacturing sector. Although our sample appears to be biased toward large plants, these plants account for large shares of manufacturing employment and shipments.

There may also be sample bias in the reduced CNUS sample we use to estimate the two-stage model. Because we must observe plants in both 1992 and 1999, the sample for the two-stage model may include only stable, "good" plants that stay in business. In addition, only half of the plants in 1992 reported computer expenditures in 1992. If response was not random, this second source of sample reduction may cause bias.

We assess the potential effects of longevity and response bias by reestimating the OLS regressions reported in column 1, Table 2, using the 10,496 plants that existed in 1992 and reported computer expenditures and all other variables in 1992. We report the results in column 3, Table 2. The OLS estimate for Network with the reduced sample is 0.038 and statistically significant at the 1% level. This estimate is virtually identical to the coefficient of 0.037 based on the larger sample and reported in column 1.

VI. CONCLUDING REMARKS

This article explores the relationship between computer networks and labor productivity using data on nearly 30,000 plants representing the U.S. manufacturing sector. Our work makes four important contributions. First, ours is the first study linking the use of computer networks to plant-level labor productivity. Second, we extend the existing model relating computers to productivity by including materials as an explicit factor input. Third, we model and predict the probability that a plant has a computer network. Fourth, we test for possible sample and endogeneity biases associated with the computer network variable.

Our results show that computer networks have a significant effect on plant labor productivity. This finding is robust, holding up for two measures of labor productivity and several alternative model specifications, including models that take into account endogeneity. It is also consistent with the few other studies in the literature that look at the use of advanced technologies in the United States, such as McGuckin et al. (1998) and computer networks in other countries, such as Motohashi (2001).

Estimates of the network based on our preferred three-factor, gross-output, specification are half the size of estimates based on the two-factor, value-added model used in many previous studies of the effect of IT on productivity. Because the three-factor model is theoretically superior to the two-factor model and has empirically superior explanatory power, our result provides strong empirical evidence that the value-added model is subject to omitted variable bias and overestimates the effects of computer networks and IT on productivity.

Although we think that the empirical work in this article goes in the right direction, our results should be considered suggestive rather than conclusive. There are two important reasons for this remark. First, due to the lack of data, we are not able to treat computer capital as a separate input in this study. The second reason is that our study covers only a single year, 1999, because data on computer networks are available for the first time in that year. Our result nevertheless strongly suggests that computer networks have a significant effect on U.S. manufacturing plants' labor productivity.

ABBREVIATIONS

ASM: Annual Survey of Manufactures

CM: Census of Manufactures

CNUS: Computer Network Use Survey

IT: Information Technology

OLS: Ordinary Least Squares

SIC: Standard Industrial Classification

SMT: Survey of Manufacturing Technologies

TFP: Total Factor Productivity
TABLE 1
Definitions and Means of Variables for Plants in Sample

 Average for Plants

 With Without
Variable Definition Networks Networks

Labor productivity TVS/TE 284.79 222.39
Labor productivity VA/TE 133.65 103.29
Employment TE 235.70 118.64

 All Plants

Labor productivity TVS/TE 277.34
Labor productivity VA/TE 130.03
Employment TE 221.72
Materials M 333.22
Network Network = 1 if plant uses 0.88
 computer network
Capital intensity Capital/labor ratio in 1997 107.50
 (K97/TE97)
Materials intensity M/TE 147.31
Skill mix OW/TE 0.27
Multiunit Multi = 1 if plant owned by 0.64
 multiplant firm
 0 < TE < 50 0.29
 50 [less than or equal to] 0.19
 TE 99
 100 [less than or equal to] 0.28
 TE < 250
 250 [less than or equal to] 0.14
 TE < 499
 500 [less than or equal to] 0.07
 TE < 999
Size TE [greater than or equal 0.03
 to] 1000
 Three-digit NAICS
 subsectors
Industry 311-316;321-327;331-337 NA

Variable definitions: TVS: total value of shipments; TE: total
employment (total number of production and nonproduction workers);
VA: TVS minus materials, energy, contract work; K: total asset value
(book value of buildings and machinery) in 1997; OW: number of
nonproduction workers; M. value of materials and parts, energy
consumed, and contract work.

TABLE 2
Labor Productivity Regression Results

 OLS Estimates

Dependent Variable

 Gross Output Value-Added (a)
Independent Variables (1) (2)

Intercept 2.948 ** (114.95) 3.946 ** (103.56)
Network99 0.037 ** (3.00) 0.075 ** (3.56)
Pr(Network99) (--) (--) (--) (--)
Log(K/L97) 0.078 ** (24.19) 0.130 ** (24.53)
Log(M/L99) 0.451 ** (118.96) (--) (--)
Log(L99) -0.005 (+) (1.78) 0.003 (0.52)
Log(RLP92) 0.276 ** (32.80) 0.508 ** (66.24)
Log(Ski1199) 0.035 ** (7.25) 0.071 ** (8.46)
MULTI99 0.088 ** (10.04) 0.152 ** (10.32)
NEW 0.203 ** (7.15) -0.120 ** (2.84)
NEW x Network99 -0.008 (0.56) -0.015 (0.61)
NEW x Pr(Network99) (--) (--) (--) (--)
NEW x log(K/L97) -0.018 ** (4.53) -0.024 ** (3.80)
NEW x log(M/L99) -0.032 ** (7.04) (--) (--)
NEW x log(L99) 0.004 (1.01) 0.005 (0.76)
NEW x log(RLP92) -0.075 ** (8.40) 0.508 ** (66.24)
NEW x log(Ski1199) 0.005 (0.91) 0.017 (+) (1.65)
NEW x MULTI99 -0.020 (+) (1.85) -0.060 ** (3.32)
[R.sup.2] 0.8133 0.4159
Number of plants 29,840 29,758

 Two-Stage
 OLS Estimates Estimates

Dependent Variable Gross Output

 Gross Output (b) Standard (b)
Independent Variables (3) (4)

Intercept 2.92 ** (90.85) 2.362 ** (17.23)
Network99 0.038 ** (2.76) (--) (--)
Pr(Network99) (--) (--) 0.669 ** (4.39)
Log(K/L97) 0.083 ** (22.42) 0.082 ** (22.10)
Log(M/L99) 0.458 ** (105.21) 0.459 ** (105.39)
Log(L99) -0.004 * (-1.25) -0.003 (-0.94)
Log(RLP92) 0.277 ** (29.91) 0.289 ** (29.86)
Log(Ski1199) 0.032 ** (5.83) 0.034 ** (6.27)
MULTI99 0.082 (8.55) 0.04 ** (2.85)
NEW (--) (--) (--) (--)
NEW x Network99 (--) (--) (--) (--)
NEW x Pr(Network99) (--) (--) (--) (--)
NEW x log(K/L97) (--) (--) (--) (--)
NEW x log(M/L99) (--) (--) (--) (--)
NEW x log(L99) (--) (--) (--) (--)
NEW x log(RLP92) (--) (--) (--) (--)
NEW x log(Ski1199) (--) (--) (--) (--)
NEW x MULTI99 (--) (--) (--) (--)
[R.sup.2] 0.7811 0.7724
Number of plants 10,496 10,496

 Two-Stage Estimates

Dependent Variable Gross Output

 Corrected Errors (b)
Independent Variables (5)

Intercept 2.363 ** (14.68)
Network99 (--) (--)
Pr(Network99) 0.669 ** (3.88)
Log(K/L97) 0.082 ** (17.52)
Log(M/L99) 0.459 ** (52.21)
Log(L99) -0.003 (-0.082)
Log(RLP92) 0.289 ** (21.47)
Log(Ski1199) 0.040 ** (2.74)
MULTI99 0.0482 ** (3.63)
NEW (--) (--)
NEW x Network99 (--) (--)
NEW x Pr(Network99) (--) (--)
NEW x log(K/L97) (--) (--)
NEW x log(M/L99) (--) (--)
NEW x log(L99) (--) (--)
NEW x log(RLP92) (--) (--)
NEW x log(Ski1199) (--) (--)
NEW x MULTI99 (--) (--)
[R.sup.2] 0.7819
Number of plants 10,496

Notes: All estimates include controls for three-digit NAICS industry.
t-statistics are in parentheses.

** Significant at the 1% level.

* Significant at the 5% level.

(+) Significant at the 10% level.

(a) The number of observations in column 2 is smaller than that in
column 1 because a number of plants have value-added equal to zero.

(b) The number of observations in columns 3, 4, and 5 is smaller
than that in columns 1 and 2 for several reasons. Estimating the
probit in the first stage of the two-stage estimates reported in
columns 4 and 5 required variables from prior periods that are not
used in the OLS estimates. One of these variables, computer
expenditures, is reported by only about half of all plants.
Additionally, many plants are new since the prior period, 1992.
The OLS regression reported in column 3 uses the same reduced
sample that is used in the two-stage estimates.

TABLE 3
Probit Regression Results (Dependent Variable: Plant has Computer
Network in 1999 [1, 0])

 Marginal
Independent Variables in Initial Period Effects z

Log(RLP92) Log of labor productivity -0.016 -2.86 **
 relative to the plant's
 four-digit SIC industry
Log(SKILL92) Log of skill mix -0.002 -0.61
Log(COMP92) Log of computer 0.030 1.49
 expenditures/total
 employment
MULTI92 Part of multiunit firm 0.067 10.66 **
IND92 Two-digit SIC industry Yes
 group
Number of plants 10,801
Chi-square 368.57 **
Pseudo [r.sup.2] 0.0636
Log likelihood -2712.8

Notes: Initial period is 1992 for plants in existence in 1992, 1997
for new plants.

** Significant at the 1% level.

TABLE 4
The Effect of Computer Networks on Plant
Labor Productivity

 % Increase
 in Labor
Percentiles (%) of Pr(Network) Productivity (a)
(1) (2)

1% (0.7467) versus 99% (0.9816) 15.72%
5% (0.8214) versus 99% (0.9816) 10.72%
1% (0.7467) versus 10% (0.8658) 7.97%
10% (0.8658) versus 90% (0.9733) 7.20%
10% (0.8658) versus 50% (0.9426) 5.14%
25% (0.8957) versus 75% (0.9613) 4.39%

Note: Based on two-stage estimates with robust standard
errors.

(a) The estimated increases in labor productivity in
column 2 are calculated by comparing plants at different
points in the distribution of predicted probabilities of
having a computer network. For example, the first row
compares plants and the 1st and 99th percentiles of the
predicted probability of having a network. The entry in
column 2 is calculated as 0.6694(0.9816 - 0.7467) = 0.1572
(15.72%), where 0.6694 is the estimated coefficient.


(1.) Gullickson and Harper (1999) discuss possible sources of measurement bias in aggregate productivity growth.

(2.) However, using Domar aggregation takes care of potential double counting, allowing gross output to be used.

(3.) The 1999 CNUS was mailed to the plants in the ASM sample. The supplement asked about the presence of computer networks, and the kind of network (EDI, Internet, both). It also collected information about manufacturers' e-commerce activities and use of e-business processes. The questionnaire asked if the plant allowed online ordering and the percentage of total shipments ordered online. Information on online purchases was also asked. In addition, information was collected about the plant's current and planned use of about 25 business processes conducted over computer network (such as procurement, payroll, inventory, etc.) and whether the plant shared information online with vendors, customers, and other plants within the company. The ASM is designed to produce estimates of the manufacturing sector of the economy. The manufacturing universe consists of approximately 365,000 plants. Data are collected annually from a probability sample of approximately 50,000 of the 200,000 manufacturing plants with five or more employees. Data for the remaining 165,000 plants with fewer than five employees are imputed using information obtained from administrative sources. Approximately 83% of the plants responded to this supplement. All CNUS data are on the NAICS basis.

(4.) Weuse 1997 data on capital intensity (K/L) because data on total capital stock are not available in 1999, which is not an Economic Census year.

(5.) We estimated equation (5) using size class variables as proxies for size. That is, we classified plants into six standard employment size groups: fewer than 50 employees, 50 to 99, 100 to 249, 250 to 499, 500 and 999, and 1000 or more. We also estimated equation (5) using another set of proxies for size that assigns a value of 1 for group 1, a value of 2 for group 2, etc. Because the three proxies yield similar estimates, we report only the results based on the measure consistent with our theoretical specification.

(6.) Ideally, one would want to allow for heterogeneity among industries, and estimate separate industry regressions, rather than simply including industry fixed effects. However, when we separate the sample into three-digit industries, where the assumption of a common production function is more tractable, few such industries have enough variation in the presence of computer networks to allow meaningful regression estimates. For several industries for which we have sufficient data, we found that the estimated coefficients for Network are positive and significant.

(7.) We are grateful to the anonymous referee for pointing out this possibility.

(8.) The SMT collected data only in industries thought to be primary users of such technology: Fabricated metal products (SIC 34), Industrial machinery and equipment (SIC 35), Electronic and other electric equipment (SIC 36), Transportation equipment (SIC 37), and Instruments and related products (SIC 38). Plants were asked about their use of 17 advanced technologies.

(9.) In our preliminary work, we found larger negative coefficients for the size variable in several alternative specifications using the CNUS data that do not control for plant age (NEW). Returns to scale are also less than one in McGuckin et al. (1998) for the SMT data, in coefficients using a different set of U.S. manufacturing panel data from a much earlier period, and in the Nguyen and Reznek (1991) study of returns to scale in selected U.S. manufacturing industries.

REFERENCES

Atrostic, B. K., and J. Gates. "U.S. Productivity and Electronic Business Processes in Manufacturing." Working Paper CES-01-11, U.S. Bureau of the Census, 2001.

Atrostic, B. K., J. Gates, and R. Jarmin. "Measuring the Electronic Economy: Current Status and Next Steps." Working Paper CES-00-10, U.S. Bureau of the Census, 2000.

Baily, M. N. "Productivity Growth and Materials Use in U.S. Manufacturing." Quarterly Journal of Economics, February 1986, 185-95.

Bartlesman, E., and M. Doms. "Understanding Productivity: Lessons from Longitudinal Microdata." Journal of Economic Literature, September 2000, 569-94.

Basu, S., and J. Fernald. "Returns to Scale in U.S. Production: Estimates and Implications." Journal of Political Economy, April 1997, 249-83.

Breshnahan, T., and M. Trajtenberg. "General Purpose Technologies: 'Engines of Growth'?" Journal of Econometrics, 65, 1995, 83-108.

Brynjolfsson, E., and L. M. Hitt. "Beyond Computation: Information Technology, Organizational Transformation and Business Performance." Journal of Economic Perspectives, Fall 2000, 23-48.

Dunne, T., L. Foster, J. Haltiwanger, and K. Troske. "Wage and Productivity Dispersion in U.S. Manufacturing: The Role of Computer Investment." NBER Working Paper 7465, 2000.

Dunne, T., and J. Schmitz. "Wages, Employer Size-Wage Premia and Employment Structure: Their Relationship to Advanced-Technology Usage at U.S. Manufacturing Establishments." CES Working Paper CES 92-15, U.S. Bureau of the Census, 1992.

Greenan, N., and J. Mairesse. "Computers and Productivity in France: Some Evidence." NBER Working Paper No. 5836, 1996.

Greenan, N., J. Mairesse, and A. Topiol-Bensaid. "Information Technology and Research and Development Impacts on Productivity and Skills: Looking for Correlations on French Firm Level Data." NBER Working Paper 8075, 2001.

Greene, W. Econometric Analysis, 4th ed. New Jersey: Prentice Hall, 2000.

Griliches, Z., and J. Mairesse. "Production Functions: The Search for Identification." NBER Working Paper 5067, 1995.

Gullickson, W., and M. J. Harper. "Possible Measurement Bias in Aggregate Productivity Growth." Monthly Labor Review, 122(2), 1999, 47-67.

Heckman, J. "Sample Selection Bias as a Specification Error." Econometrica, 47(1), 1979, 153-62.

Huber, P. "The Behavior of Maximum Likelihood Estimates Under Nonstandard Conditions," in Proceedings of the 5th Berkeley Symposium in Mathematical Statistics, vol. 1. Berkeley: University of California Press, 1967.

Jorgenson, D. "Information Technology and the U.S. Economy." American Economic Review, March 2001, 1-32.

Jorgenson, D., and K. J. Stiroh. "Industry-Level Productivity and Competitiveness between Canada and the United States." American Economic Review, May 2000, 161-67.

Maddala, G. S. Limited-Dependent and Qualitative Variables in Econometrics. Cambridge: Cambridge University Press, 1983.

McGuckin, R. H., and S. V. Nguyen. "Post-Reform Industrial Productivity Performance of China: New Evidence from the 1985 Industrial Census Data." Economic Inquiry, 31, July 1993, 323-41.

McGuckin, R. H., M. L. Streitwieser, and M. E. Doms. "The Effect of Technology Use on Productivity Growth." Economic Innovation and New Technology Journal 7, October 1998, 1-26.

Motohashi, K. "Economic Analysis of Information Network Use: Organizational and Productivity Impacts on Japanese Firms." METI Working Paper, Tokyo, 2001.

Nguyen, S., and A. Reznek. "Returns to Scale in Small and Large U.S. Manufacturing Establishments." Small Business Economics, 1, 1991, 29-46.

Oliner, S. D., and D. E. Sichel. "The Resurgence of Growth in the Late 1990s: Is Information Technology the Story?" Journal of Economic Perspectives, Fall 2000, 3-22.

Stiroh, K. "Information Technology and the U.S. Productivity Revival: What Do the Industry Data Say?" Federal Reserve Bank of New York Staff Reports, number 115, January 2001.

Stiroh, K. "Reassessing the Impact of IT in the Production Function: A Meta-Analysis." Federal Reserve Bank of New York, November 2002.

Stolarick K. M. "Are Some Firms Better at IT? Differing Relationships between Productivity and IT Spending," Working Paper CES99-13, U.S. Census Bureau, 1999.

Triplett, J. E., and B. Bosworth. "Productivity in the Services Sector." Presentation at the American Economic Association meetings, January 9, 2000.

White, H. "A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity." Econometrica, 48, 1980, 817-830.

White, H. "Maximum Likelihood Estimation of Misspecified Models." Econometrica, 50, 1982, 1-25.

B. K. ATROSTIC and SANG V. NGUYEN *

* This article reports the results of research and analysis by the authors. It has undergone a more limited review than official publications. Opinions expressed are those of the authors and do not necessarily represent the official position of the U.S. Census Bureau. We have benefited from comments by Randy Becker, Seong Hoon Lee, Jacques Mairesse, Jack Triplett, and an anonymous referee, but all errors of course are our own.

Atrostic: Senior Economist, Center for Economic Studies, U.S. Census Bureau, 4700 Silver Hill Road, Stop 6300, Washington, DC, 20233. Phone 1-301-763-6442, Fax 1-301-455-1235, E-mail batrostic@ces.census.gov

Nguyen: Senior Economist, Center for Economic Studies, U.S. Census Bureau, 4700 Silver Hill Road, Stop 6300, Washington, DC, 20233. Phone 1-301-763-1882, Fax 1-301-455-1235, E-mail sgnuyen@ces.census.gov
COPYRIGHT 2005 Western Economic Association International
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2005 Gale, Cengage Learning. All rights reserved.

 Reader Opinion

Title:

Comment:



 

Article Details
Printer friendly Cite/link Email Feedback
Title Annotation:information technology
Author:Atrostic, B.K.; Nguyen, Sang V.
Publication:Economic Inquiry
Geographic Code:1USA
Date:Jul 1, 2005
Words:9454
Previous Article:The role of expectations in adolescent schooling choices: do youths respond to economic incentives?
Next Article:Inefficient standard adoption: inertia and momentum revisited.
Topics:

Terms of use | Copyright © 2014 Farlex, Inc. | Feedback | For webmasters