JOB HOPPING, INFORMATION TECHNOLOGY SPILLOVERS, AND PRODUCTIVITY GROWTH

Transcription

1 JOB HOPPING, INFORMATION TECHNOLOGY SPILLOVERS, AND PRODUCTIVITY GROWTH Prasanna Tambe Stern School of Business New York University Lorin M. Hitt The Wharton School University of Pennsylvania DRAFT November 2010 Abstract We show that regional differences in returns to information technology (IT) investments by US firms are attributable in part to IT spillovers generated by the flow of IT workers among firms. We use a newly developed source of employee micro-data with employer identifiers and location information to model IT workers mobility patterns. Access to an external IT pool that is one standard deviation larger than the mean is associated with a substantial increase in the output elasticity of own IT investment. The effects of spatial proximity disappear after controlling directly for employee mobility. Our estimates suggest, for instance, that the output elasticity of IT investment for non-durable goods manufacturing firms (SIC 20-29) located in some parts of Northern California is about 5% higher than that of firms in these industries located in other parts of the country. Keywords: IT externalities, labor mobility, general-purpose technologies, knowledge spillovers, organizational practices, high-tech clusters, productivity Contact: ptambe@stern.nyu.edu. We are grateful for valuable feedback from David Autor, Erik Brynjolfsson, Matthew Bidwell, Peter Cappelli, Eric Clemons, Avi Goldfarb, Lori Rosenkopf, and AnnaLee Saxenian, as well as anonymous reviewers from the International Conference on Information Systems and seminar participants at the Wharton People and Organizations Conference, NBER IT and Economics Summer Institute, MIT, Purdue University, Industry Studies Association Conference, Cornell University, UC Irvine, New York University, University of Pennsylvania, University of Maryland, and the University of Minnesota.

2 Introduction Researchers have argued that like earlier general-purpose technologies such as electricity or the steam engine, information technology (IT) investments may generate substantial knowledge spillovers as new, improved methods of IT-enabled production diffuse among firms (David, 1990). Superior access to spillovers of these new production methods and work practices should lead to higher IT returns and higher productivity levels for firms. Therefore, identification and measurement of the economic impact of these spillovers has implications for understanding productivity differences and heterogeneity in IT returns, as well as in understanding how benefits from IT investments are allocated among firms. To the extent that information technology has been responsible for a significant portion of multi-factor productivity growth in the last two decades (Jorgenson & Stiroh, 2000), these spillovers are also a potentially important source of long-run economic growth in advanced economies and variations in labor mobility may partly explain the divergence of IT-related returns in different countries. The transfer of new IT-related production innovations, however, is a complex process that requires adapting new practices to fit a specific organizational context. As with spillovers of some types of scientific knowledge (Almeida & Kogut, 1999; Franco & Filson, 2006), the diffusion of new production methods is governed by access to knowledge transfer mechanisms that promote rich face-to-face interaction, such as the movements of contractors, consultants, and employees among firms (Dedrick et al, 2003). This is consistent with patterns of diffusion observed in earlier generations of new production technologies for example, economic historians have suggested that skilled factory architects and electrical engineers were instrumental in spreading work practices related to the installation of the electric dynamo in US 2

3 factories (David, 1990). 1 Because labor mobility tends to be local, location may play a particularly important role in determining access to these new production methods, and may partly explain why firms choose to locate in high-tech clusters despite facing higher factor costs for other inputs, such as land and labor. Prior research has suggested that geographic clustering in some industries is related to superior access to knowledge complementary to new IT-related production methods (Saxenian, 1996; Forman, Goldfarb, & Greenstein, 2005; Desmet & Rossi- Hanbserg, 2008). The primary goal of this study is to test the hypothesis that firms benefit from the IT investments of other firms because the mobility of technical workers generates knowledge spillovers related to new IT-related production methods. We first show that 1) location affects returns to IT investments within the US and that 2) that differences in these regional returns can be at least partially explained by knowledge spillovers generated by the movements of IT labor among firms. The primary contribution of the study is collection of a new data set describing IT labor flow patterns. Our data are derived from the employment histories of several hundred thousand US-based IT workers and include information on both employers and skills. The data represent the career movements of as much as one-sixth of the IT workforce over the last twenty years. Access to this unique sample allows us to track the mobility patterns of IT workers among US firms and the likely diffusion paths of IT-related practices, and to therefore test an important mechanism through which firms may benefit from the IT investments of other firms. These data also include detailed experience and education information about these workers, so when testing 1 The economic importance of this type of employee mobility in the IT sector is suggested by recent press covering the Department of Justice investigation of collusive non-poaching agreements in several prominent technology firms (Catan & Kendall, 2010), as well as in coverage of the difficulties faced by large tech firms in slowing the defection of employees to competitors (Efrati & Morrison, 2010). 3

4 our spillover measures, we can control for the flow of other types of IT human capital among firms. Our paper, therefore, contributes to a large literature on IT productivity, where there is currently interest in understanding the heterogeneity in IT returns that has been observed across firms and countries (Dewan & Kraemer, 2000; Van Reenen, Sadun, & Bloom, 2008). Our study is among the first to attempt to identify how IT spillovers might drive regional IT productivity differentials, and to the best of our knowledge, is the first study to investigate this issue using micro-data on labor mobility. We also contribute to a literature on externalities produced by IT investments. Although most of the attention in the literature has been focused on private returns to IT investment, economically significant IT spillovers would have important implications for understanding how firms benefit from the IT investments of other firms, which has implications for optimal investment strategy and growth policy. Our findings suggest that location explains a substantial amount of variation in IT returns. Geographic differences in turn are partially explained by spillovers produced by IT labor flows. Our spillover estimates indicate a rate of return to external IT investment that is about 20% the rate of return to internal investment. In other words, firms thatsince firms tend to hire most of their technical workers locally, being located in a region where average IT investment levels are one standard deviation higher than the mean is associated with a rate of return that is about 20% beyond that of the rate of return to private investment levels. Furthermore, higher returns are attributable specifically to the flow of labor, not just proximity, and some level of private IT investment is necessary to capture these benefitsprimarily captured by those firms that are making substantial private IT investments, suggesting that firms cannot simply free-ride on the IT investments of other firms. In the next section, we discuss the theoretical foundations of this 4

5 paper. Section 3 describes the empirical framework. In Sections 4 and 5, we describe our data and measures. Finally, in Sections 6 and 7, we discuss some our findings and their implications for managers and policy makers. Framework The literature on the productivity of IT investment has heavily emphasized the role of organizational complements in explaining heterogeneity in IT returns (Bresnahan, Brynjolfsson, & Hitt, 2002; Caroli & Van Reenen, 2002). This stream of research has produced substantial evidence that higher levels of productivity are generated not simply from investing in IT, but also by reorganizing the firm s information structures. Empirical research has primarily focused on decentralization of decision-making, but scholars have also looked at the role of other information practices, such as external information practices and workflow digitization, which require extensive know-how related to technologies, protocols, standards, and information architectures. Differences in the adoption levels of these new innovations across firms have been connected to substantial IT-related performance differentials, raising the question of why the most effective combinations of IT and these new production innovations are slow to be recognized and adopted by some firms. One explanation for these differences is that organizational transformation is a difficult and costly process. In prior research, scholars have argued that organizational adjustment costs (i.e. the costs of making changes within the workforce and organization to best use the new technology) can be substantial (Levy & Murnane, 1996; Milgrom & Roberts, 1990). Therefore, rather than being re-invented within each firm, firms are more likely to learn these practices by imitating other firms. This suggests that spillovers are likely to be substantially stronger over small distances because the knowledge transmission mechanisms rich enough to facilitate this 5

6 knowledge transfer among firms, such as employee mobility, informal social interactions, or shared membership in technical user groups, are often much more densely concentrated over small distances. Therefore, the diffusion of IT-related organizational complements, like complex scientific knowledge, should be geographically localized. 2 The central thesis of this paper is that geographic variation in IT returns is in part due to spillovers of new IT-enabled production methods that are generated through labor mobility. A common method for testing for the economic effects of spillovers is to create aggregate measures of the external pool of knowledge available to firms under assumptions about how firms access this pool of knowledge. For example, if knowledge spillovers are hypothesized to occur only among firms within the same industry, as might be expected for specialized knowledge pertaining to specialized scientific advances, measures of available extramural knowledge pools could be constructed from the R&D knowledge of other firms in the same industry. Alternatively, if spillovers are regional in scope, then spillover pools can be constructed from the investment levels of geographically proximate firms. Although the use of aggregate industry or regional investment levels has been commonplace in the literature on knowledge spillovers, the use of broad measures based on industry or regional investment levels is subject to criticism (Breschi & Lissoni, 2001; Van Reenen, Schankerman, & Bloom, 2007). In addition to confounding the effects of individual spillover mechanisms that act within industries or regions, the use of broad measures may produce spurious results in models that confound industrial or technological proximity with regional heterogeneity in cross-section and over time due to region-specific shifts in taxes, labor migration, or government policy. Moreover, to the extent that some regions have concentrations 2 A rich literature documents the localized diffusion of scientific knowledge. For example, see Jaffe, Trajtenberg, and Henderson (1993). Additional references available upon request. 6

7 in specific industries, they may experience common productivity or IT-demand shocks due to changing patterns of supply and demand for industries overrepresented in their area. Emphasis has therefore shifted towards explicit modeling of the individual mechanisms of knowledge transfer though the collection of fine-grained data on social or economic activity. As argued above, for IT spillovers an important mechanism for knowledge transfer may be labor mobility because it allows for the repeated face-to-face interaction required for adapting technical knowledge to a new organizational context. Therefore, rather than using broad and potentially problematic regional or industry measures, we explicitly model this transmission pathway using detailed data on the employment paths of IT workers, and compare these results with those from broader industry or geography based measures. In the next section, we describe the data, methods and measures we use to test these hypotheses. Empirical Framework Our empirical approach is closely related to a literature on the impact of R&D knowledge spillovers on productivity, as well as to a literature on the productivity of information technology investments. An approach common in both literatures is to use methods from production economics that enable researchers to estimate the contributions of various inputs, such as capital, labor, R&D, and IT, to firm productivity. To estimate the contribution of R&D knowledge spillovers to productivity, productivity models can be augmented with measures of the external knowledge pool available to a firm (Griliches, 1992). A similar approach has been proposed for studying IT externalities where IT is substituted for R&D as the knowledge producing asset (Draca, Sadun, & Van Reenen, 2006). To implement this approach, a standard production function relating capital (K), labor (L) to output is augmented with measures of the stock of computers (C) and the stock of computers 7

8 held by other firms in close proximity (S). Taking logarithms of both sides of an augmented Cobb-Douglas production function produces: (1) vaijt = α kkijt + α Llijt + α ccijt + α ssijt + ε ijt A positive and economically significant coefficient on the spillover term ( α ) is interpreted as evidence that spillovers generate substantial productivity benefits. Because data on investments in capital, labor, and output are widely available, the most significant challenge in using this approach is the development of reliable measures of firm s internal and external knowledge sets. A great deal of attention has been paid in the literature on knowledge spillovers to how best to measure these external knowledge pools (Griliches, 1992). Historically, most research on R&D spillovers has relied on the assumption that firms benefit from the knowledge of other firms when they are close, in a technological or geographical sense (Griliches, 1992; Jaffe, 1986). Under these assumptions, the knowledge available to the focal firm i (K i ), is modeled as the weighted sum of the knowledge of other firms in the sample (K j ), where the weights (Φ ij ) between firms i and j reflect the amount of knowledge leakage between the two firms, proxied by a measure of proximity. (2) Sit = j φ i ijt K jt A similar approach is possible for the measurement of IT externalities, but the use of broad proximity measures based on technological position or geographic location to measure spillovers has received criticism for the reasons described above. These criticisms can be avoided by direct measurement of the spillover path. In this study, we focus on the flow of IT employees between firms as the measure of proximity, based on the argument that knowledge of IT-related process innovations is disproportionately transmitted by the flow of IT workers across firms. For comparison, we also model the potential IT spillover pool available to firms based on s 8

9 industry and geographic proximity. The spillover augmented production function in (1) can be estimated using standard regression techniques such as ordinary least squares (with suitable standard error corrections for panel data) or panel methods such as fixed effects as well as other approaches from the micro-productivity literature for addressing endogeneity in the production function framework. Data Our employee mobility data were obtained from a leading web site through which participants post employment histories (resumes) online. We have access to employment information from about ten million users who posted or modified their career histories through this service in Users provide information about occupation (e.g. information technology, sales, management, etc.), education, and other demographic and human capital variables. In this study, we primarily focus on IT workers, although we use data from other occupations to test the robustness of our results. Our data set includes career information about full-time employees as well as hourly workers. We include both types in our analysis because our primary interest is in knowledge transfer, and contractors and part-time workers also play an important role in knowledge transfer. We use the inter-firm mobility patterns of workers in our data to model the flow of IT workers among firms in the US. The educational and occupational characteristics of the IT workers in this data set are shown in Table 1. We compare their educational attainment to that of the IT workers sampled in the 2006 Current Population Survey (CPS), a nationally representative Census administered survey. 3 The educational distribution of the workers in our sample is 3 The survey is administered monthly to 50,000 households and is conducted by the Bureau of Labor Statistics (BLS). The sample is selected to represent the civilian non-institutional population. More information is available at 9

10 similar to that of the IT workers in the CPS. Our data do not appear to be particularly skewed towards either higher or lower education levels, although workers with vocational training appear to be slightly over-represented. We also compared job tenure statistics against the job tenure of information technology workers who appeared in the CPS Job Tenure supplement, last published in Unsurprisingly, the average job tenure of workers in our sample is about two years lower than the average job tenure of workers in the CPS survey (p<.01). One reason for this difference is that our sample is likely to contain a higher fraction of younger workers and job-hoppers. However, because the population of interest in this study is the job-hopping IT worker, our sample may actually be a better representation of this population than the CPS sample, which includes workers who never or infrequently switch jobs. To develop inter-firm measures of IT worker flows from these employee micro-data, we associate employer names with unique identifiers by matching them against external databases of publicly listed companies and subsidiaries, including Compustat, the Compact Disclosure Database, and the NBER Patents database, and then aggregate the data by firm-year. We restrict our analysis to the movements of employees between public firms because of the availability of supplementary economic data on these firms through Compustat. Aggregating these data to the firm-level provides information on 1) how many IT workers were employed at a particular firm in a particular year 4, and 2) how many workers a firm hired from each other firm in our dataset. In the next section, we explain how we use these data to develop measures of IT investment and the IT spillover pool. Measures 4 In Tambe & Hitt (2008), we show that these firm-level IT worker measures compare favorably to external data sets on firm-level IT employment. 10

11 IT Expenditures Our IT investment measures, intended to capture a firm s investments in hardware, software, human capital, and reengineering, are based on IT employment intensity. Construction of this IT data set, its sampling properties, comparison with other external IT data sets, and its behavior in productivity regressions is described in greater detail elsewhere (Tambe & Hitt, 2010). The data set describes firm-level IT employment over the last two decades, which we use to measure firms IT expenditures. Although aggregate firm-level IT expenditure data has been used in some recent studies, it is based on limited surveys, available only for a small sample of firms. By contrast, broader archival datasets such as our IT employment data are available for much longer timer periods, and for a much larger sample of firms, and are therefore often used to measure IT expenditures in larger samples of firms (Lichtenberg, 1995; Brynjolfsson & Hitt, 1996). Correlations reported in survey-based studies have demonstrated that the IT intensity of firms, measured by IT employment or capital data, is closely associated with complementary investments in organizational and information practices (Bresnahan, Brynjolfsson, & Hitt, 2002; Tambe, Hitt, & Brynjolfsson, 2010). The use of IT investment to measure IT know-how also closely follows an approach from the literature on R&D spillovers, where R&D expenses are used to represent a firm s R&D capital stock. 5 Among archival data sets, our employment based data have some advantages over alternative IT data sets. Although recent research on IT productivity has relied heavily on the Computer Intelligence Technology Database (CITDB), the main panel of these data is restricted to Fortune 1000 firms, the definitions of variables changed significantly after 1994 and most 5 Our IT measures in most regressions are IT labor investment flows, but we show in our analysis that our results are not substantively different if we convert these measures into IT labor stock measures or if we use IT capital stock rather than IT labor stock. 11

12 importantly, the CITDB no longer includes direct measures of IT capital stock. 6 Our employment data are highly correlated with other available IT capital, employment, and labor expense data, suggesting that it is a reasonable proxy for aggregate IT expenditure. 7 In this study, therefore, we primarily use IT employment as our IT measure. However, we also report results from regressions using the CITDB IT capital stock data and IT stock measures inferred from our IT employment data to demonstrate the robustness of our results to our choice of IT measure. Measuring the IT Spillover Pool Our spillover measure is computed as the IT intensity of other firms from which a firm hires at least 5% of its new IT workers. This approach is similar to that used in other knowledge spillover studies, where the R&D spillover pool is constructed as the aggregate R&D capital stock of other firms within a certain radius or within a particular industry (e.g., see Orlando, 2004). The assumption is that if employee mobility is an important mechanism for the transfer of IT practices, a firm receives greater knowledge spill-ins from organizations from which they hire large numbers of IT workers. 8 In this sample, the average firm has slightly over 6 other firms from which it hires at least 5% of its IT workers. Towards the end of our analysis, we also compare results using this threshold with those using a threshold of 1%, for which the average firm has over 28 firms in its network, as well as an employee-flow weighted measure with no threshold. 6 Chwelos, Ramirez, Kraemer and Melville (2007) provide a method for extending CITDB 1994 valuation data through 1998 by imputing the values of equipment in the earlier part of the dataset and adjusting for aggregate price changes. However, this differs from the method employed by Computer Intelligence which determined equipment market values by looking at actual prices in the new, rental and resale computer markets. 7 Correlations between our employment data set and other well known data sets, such as the CITDB, ComputerWorld, and InformationWeek data sets are generally above.5 (Tambe & Hitt, 2008). 8 Although we use a 5% threshold for most of our regressions, our primary results are not significantly affected when using 1%, 2%, 5%, 10%, or 15% thresholds. Robustness results using these thresholds are reported. 12

13 Our primary measure is increasing in the IT investment levels of other firms. From a network perspective, if IT mobility is a key mechanism for the transfer of IT-related knowledge, the spillover pool is larger for firms that hire employees from other IT-intensive firms. Employers embedded in high-tech regions, for example, are more likely to hire employees from high-tech firms, and should benefit from the flow of skills and knowledge from the employees coming in from these firms. However, a key distinction of our study is the focus on a specific mechanism. Therefore, positive returns to our spillover measure imply that firms that hire employees from other IT-intensive firms, regardless of their location, should capture spillovers. To behave under the log transformation, spillover pools with zero values are seeded with a minimum non-zero value--in Table 8, we test the sensitivity of our results to this seeding by altering the seeding assumptions, excluding firms with a zero spillover pool and estimating a Heckman-style selection model. None of these adjustments substantially affect our results. We also create spillover measures based on industrial and geographic proximity. Industry spillover measures are constructed as the IT intensity of other firms within the same four-digit SIC industry. Adding industry based IT investment measures to our regressions accounts for the competitive effects of IT investment which are needed to identify technological spillovers because the positive effects of technological spillovers in productivity regressions are confounded with the negative effects of competitive investment (Bloom, Schankerman, & Van Reenen 2007). Geographic proximity is measured in two different ways. First, we construct broad measures by grouping firms that have headquarters within the same state and county, where county and state of a firm s corporate headquarters are from the Compustat data set. These data are available for most firms and years in our sample, making it possible to match these data to 13

14 almost our entire panel. The use of corporate headquarters to identify regional influences on a firm, however, is an imperfect measure because IT workers are generally geographically scattered across multi-establishment firms. To test how our results are affected by this type of measurement error, we use workers reported metro-area locations in 2006 to construct finergrained establishment-level measures for each firm at the metro-area level. Using these metroarea data, we consider two firms to be geographically proximate if they have establishments within the same metropolitan area. Across establishments, the aggregate pool for each firm is computed as the pool for each of its establishments, weighted by the share of IT employment at that establishment. This measure is more precise than our earlier measure based on the location of corporate headquarters, but the drawback to using these measures is that these employment dispersion data are only available in the final year of our sample, so only cross-sectional comparisons are possible. Finally, we test if spillovers generated through IT labor flows explain localized spillovers. To separate the contributions of job-hoppers from other potential localized spillover channels, we jointly estimate the contributions of four separate spillover pools that vary along two dimensions, labor flows and geographic distance (for a similar example, see Orlando, 2005). Therefore, the IT investments of other firms are broken into one of four pools: firms that are 1) proximate in terms of employee flows and geography, 2) proximate in terms of employee flows but geographically distant, 3) distant in terms of employee flows but geographically proximate, and 4) distant in terms of both employee flows and geography. The division of firms into these four groups separates the estimates of the effects of employee flows from other spillover channels. 14

15 In Tables 2 and 3, we show the correlations between the spillover pools created using the 2006 establishment-metro area detail. The only notable correlation in Table 4 is between the labor mobility spillover pool and the spillover pool constructed from average IT investment levels within a county, which is.38. This correlation is expected because firms tend to hire locally. The correlations between the spillover measures created using the 2006 metro data, shown in Table 5, are higher. In particular, correlations between a firm s within region spillover pools and within hiring network spillover pools are also quite high. However, there is still substantial variation in these data to separate the impacts of these different spillover pools on firm productivity. Other Inputs, Human Capital, Wages, and Sales Compustat data were used to compute employment, capital, sales, value added, and materials in constant 2000 dollars, using standard methods from the micro-productivity literature. Non-IT employment was computed by subtracting IT employment from total employment. Means, standard deviations, and correlations for these variables are shown in Table 4, and the industry composition of the firms in our sample is shown in Table 5. In some regressions, we also include measures of the human capital and wages of incoming IT workers and of the firms IT workforce. We construct these measures using data on the education and experience of the workers employed at firms and the workers moving between firms. Education, experience levels, and 2006 wages are self-reported by individual IT workers. Education is selected from one of a number of different education levels, including doctorate, graduate school, 4 year degree, vocational degree, some college experience, high school degree, and some high school experience. Experience in the relevant industry is self-reported in To compute experience in a particular year, we adjust 2006 experience levels backwards, 15

16 assuming full employment in the interim periods. In addition to experience, we separately estimate worker age based on the years they report attaining various education levels, as reported in the full text portion of their resume. 9 Measures of firm-level IT wages are constructed by averaging the wages of workers within a particular firm and year. We also include several control variables in our model. We used Compustat data to assign firms to a SIC industry. Depending on the specification, we included industry dummies at either the one or two digit SIC level. We also include year dummies to account for time trends and unobserved sources of productivity that vary over time. Results Location, Regional IT Investment, and IT Productivity We first use our IT expenditure data to benchmark IT returns over the last two decades. In column (1) of Table 6, we report estimates from a model that includes capital, labor, and IT employment measures in a Cobb-Douglas framework. Our estimates indicate an output elasticity of slightly less than.09 (t=12.3) for IT employment, which is consistent with IT productivity estimates from other studies using IT employment data in cross-sectional regressions (Lichtenberg, 1995; Brynjolfsson & Hitt, 1996). In Columns (2)-(4), we look for evidence of regional differences by testing whether IT employment is associated with higher productivity in California. We chose California because it has been the focus of much of the literature on hightech clusters, job hopping, and IT innovation (Saxenian, 1994; Fallick et al. 2006; Freedman, 9 For every worker we have both fielded data which forms the bulk of our measures as well as the full text of their resume. Using a text mining application we are able to identify the years that a particular employee completed different levels of education (high school, vocational school, community college, or four year college). We assume that high school is completed at age eighteen, vocational or community college is completed at age twenty and other degrees are completed at age twenty-three. We compute a separate age estimate for each school attended and then take the minimum of these estimates to accommodate the possibility of a delay between entry into higher education and the completion of high school. 16

17 2008). We include a dummy variable into our main regression indicating whether a firm s corporate headquarters are located in California, and we include an interaction term between this variable and IT employment. The interaction term in (2) is positive and significant (t=2.30), and suggests that the measured output elasticity of IT investment for California firms is higher than for the average US firm. In Column (3), we show that this result is robust to including 2-digit industry controls that account for differences in industry composition across states (t=3.0). In Column (4), we include interactions between the California dummy variable and capital and non- IT employment, but the estimates on the other interaction terms are not significant, indicating that the higher elasticity documented in (2) and (3) is unique to IT employment. In Columns (5) through (8), we add measures of regional IT employment intensity into our regression. We construct our initial measure of regional IT intensity as the IT intensity of other firms with corporate headquarters in the same county. Column (5) shows the results using the sample for which headquarters location data are available. The results in (6), in which we include the location investment levels and interactions with own IT investment, indicate that adding these variables removes some of the IT-related productivity benefits associated with being located in California. This effect is more pronounced in the fixed-effect regressions shown in (7) and (8), which indicate that this effect largely disappears after including regional IT employment intensity. The main effect estimates on the industry and location spillover pools are negative, which is consistent with the negative competitive effects of increased IT investment. There are at least two reasons why higher IT returns might be associated with being located in IT-intensive areas within the US. The first, consistent with the argument in this paper, is that IT returns are higher in regions where complementary factors are more readily available. For instance, in regions characterized by spillovers of ideas related to new IT production 17

18 methods, firms may experience higher returns to computerization. Alternatively, fluctuation in regional demand for output in regionally-concentrated industries may simultaneously drive higher computer investment. In this case, a larger spillover pool might simply indicate that regional productivity growth led to higher IT investment levels. To separate these stories, we model spillover paths using employee mobility, which allows for variation in spillover levels within industry and region, and which our prior discussion suggests may better represent the actual paths by which a spillover could be transmitted. IT Labor Mobility and Spillovers In Table 7, we report our main results from models where IT labor flow patterns, rather than location, are used to model the potential spillover pool. For ease of comparison, in (1) we report our baseline results from Column (1) of Table 6. Column (2) shows our estimates after including the spillover pool generated from IT labor flows. The coefficient estimate on the pool is positive and significant (t=14.0), indicating that hiring IT workers from IT-intensive firms is associated with higher productivity levels. After including the spillover term, the size of the estimate on private IT investment becomes smaller, suggesting that some of the returns to private IT investments may reflect access to extramural sources of IT know-how. The coefficient estimate on the industry spillover term is also positive and significant, which indicates that the competitive effects caused by the IT investments of competitors are confounded with unobserved industry effects at the 1 digit industry level. The magnitude of the spillover estimate indicates that IT spillovers generated from IT mobility are associated with a rate of return to external IT investment that is about 20-30% of the internal rate of IT investment, potentially large enough to influence firm s strategic location or human resource decisions. 18

19 In columns (3) and (4), we report fixed effect estimates which remove the effects of many time-invariant factors that could bias the spillover pool estimates. The effect sizes implied by the fixed effect model in (3), without the spillover term, are consistent with prior research on IT productivity, and suggest an own-it elasticity of about The estimate on the spillover term, shown in Column (4), continues to be positive and significant (t=5.0). The negative and significant coefficient on the industry spillover term indicates that after controlling for unobserved firm-level heterogeneity, IT investment by industry competitors has negative effects on firm productivity. The magnitude of the spillover estimate from the fixed effects model is consistent with the interpretation that the external rate of return to IT investment for firms produced by knowledge spillovers is about 20% that of internal IT investment. In the next two columns, we report results from a variety of other estimators, including the Levinsohn-Petrin (LP) and Arellano-Bond (AB) estimators, which account for various cases in which the regressors may be correlated with the error term (Arellano and Bond, 1991; Levinsohn and Petrin, 2003). The LP estimator utilizes changes in firm materials inputs to measure the effect of short run productivity shocks and uses this information to correct the estimates of other input terms in practice it tends to return results similar to levels regressions (rather than differences or fixed effects). The Arellano and Bond estimator uses an optimal weighting of two regressions: a regression using input levels as instruments for changes in inputs as well as a regression using differences as an instrument for levels. Empirically, the Arellano and Bond estimator tends to perform closer to an instrumental variables regression in first differences. The coefficients on the terms of interest in the LP and AB estimators are very similar to those in the corresponding OLS/Fixed-effects regression. Thus, while they suggest that endogeneity may play a role in the precise estimates of the elasticities of various inputs, 19

20 unobserved differences in firm productivity, do not appear to substantially bias our spillover results. To some extent this is not surprising, as our estimates should only be biased in the presence of heterogeneity that is coincidentally transmitted through a path similar to IT labor mobility patterns. Finally, in Columns (7) and (8), we test whether firms can free-ride on the IT investments of other firms, or whether investments in internal IT infrastructure are required to successfully capture IT externalities. Our estimates from Columns (7) and (8) are from OLS/Fixed-Effect regressions that include an interaction term between own IT investment and the external IT pool. The estimates from both models suggest that IT externalities transmitted through labor flows are increasing in a firm s own IT investment. Access to an external IT pool that is one standard deviation higher than the mean increases the output elasticity of a firm s own IT investment by about 7%. At mean levels of IT investment, this increase in mean productivity is consistent in magnitude with estimates from our earlier models with no interaction terms. In Table 8, we test the sensitivity of these estimates to sample selection. In columns (1)- (3) we present results from models where we include only firms with non-zero spillover pools. The estimate on IT investment rises, which is consistent with higher IT returns in larger firms, and the estimates on the spillover pools and interaction terms remain significant and positive. We also omit firms with zero spillover pools in column (4), which includes an interaction term between private IT employment levels and the IT pool. The results are consistent with the results when using the broader sample and including interaction terms in the last two columns of Table 7. In (5) and (6), we include all firms in the sample, but we include a dummy variable for firms that have a seeded spillover pool. The estimates from this model indicate that our estimates are not sensitive to the value we choose to seed the spillover pool. Finally in (7) and 20

21 (8), we show that our estimates are substantively similar when using location, industry, year, sales growth, employment growth, size, and IT employment to predict selection in a Heckman regression model. Comparing Spillovers from IT Labor Mobility with Regional Measures In this section, we report estimates from models in which the spillover pool is broken into four different pools that vary along two dimensions, geographic proximity and IT labor flows. The aggregate IT intensity of other firms falls into one of the four pools based on their location and whether they are an important source for IT labor. The average size of each IT pool and the average number of firms in each pool is shown in Table 9. In Table 10, we report our estimates from this decomposition. In our OLS and fixed effect estimates in Columns (1) and (2), the coefficient estimates on the spillover pools are only positive and significant when a firm is an important source of IT labor. Otherwise, the IT intensity of other firms, even when located in the same region, is not significant. These estimates indicate that for spillovers generated by IT investments, 1) regional spillovers appear to be driven by IT labor flows, and 2) IT labor flows appear to be an important source of knowledge spillovers even outside of a fixed region. A potentially important source of error in (1) and (2) is that we use corporate headquarters to fix firm location although in larger firms, IT workers are likely to be distributed across establishments in different states. This type of measurement error should create a downward bias on the estimated effect of the near-near spillover pool and an upward bias on the near-far pool. In Column (3), we test a model where our proximity measures are constructed at the establishment-city level we consider two firms to be geographically proximate when they have establishments in the same cities. The estimates produced from this characterization provide further validation for the hypothesis that IT labor mobility is an 21

22 important channel for knowledge spillovers as with our earlier estimates, only the coefficient estimates on the near-near and near-far pool are significant, indicating that knowledge spillovers are generated from labor mobility, and that these effects are stronger over smaller distances. Furthermore, the higher coefficient estimate on the near-near pool is consistent with the reduction of measurement error in our construction of the near-near and near-far pools, as would be expected if the use of corporate headquarters as the location of a firm s IT workers is a noisy measure. In Columns (4) and (5) we report estimates from the subsample of firms in the software publishing industry (SIC 7372), an industry which is highly agglomerated (Freedman, 2008), and in which most cross-firm moves will therefore occur within the same region. Estimates from this sample of firms indicates that in these industries, within-region job-hopping has the most significant impact on productivity. 10 After including measures based on within-region jobhopping, regional IT investment measures no longer appear to have an important effect on productivity. Beyond demonstrating that IT labor flows explain much of the regional spillover effect in technology clusters, the estimates in Columns (1)-(5) suggest that it is unlikely that the estimates in Table 10 are being caused by regional or technological heterogeneity. Any biases associated with unobserved firm status or reputation should not be constrained to firms within the same geographic region. Similarly, regional heterogeneity, such as that related to demand shocks or changes in state-wide policy, should not be limited to firms from which employees are being hired. It is unlikely, therefore, that our estimates are being driven by these types of influences. In the next section, we address some remaining endogeneity concerns. 10 The effects may also influence strategic location decisions to maximize appropriability. See, for instance, Alcacer and Zhao (2008). 22

23 Robustness Tests We conducted a number of additional robustness tests. One possible criticism of our findings is that hiring workers from IT-intensive firms may reflect greater human capital in incoming employees, or alternatively higher status levels in the hiring firm, which may be associated with higher productivity. Alternatively, proximity in the IT labor flow network may reflect heterogeneity in the underlying production function. The level of detail in our data allows us to test the robustness of our results to many of these alternative stories. First, because we have data on the human capital of individual employees moving between firms, we can control for human capital differences. We first include the education and experience levels of the firm s existing IT workforce, using human capital data for each of the workers in our data set. The point estimates on IT education and IT experience in Column (1) of Table 11 are positive and significant, but including these measures does not substantially change the estimate on the spillover term. In Columns (2) and (3), we include measures for the human capital of incoming IT workers. The estimates in (2) indicate that higher education levels in the incoming IT labor pool are associated with higher productivity, but that higher experience levels in the incoming IT worker pool are associated with lower productivity. However, after adding the average age of incoming IT workers in (3), the point estimate on experience is no longer significantly different than zero, suggesting that the estimates on experience and age reflect a selection effect in which younger workers choose higher productivity, higher growth firms. Nonetheless, the coefficient estimate on the spillover term does not change significantly in any of these models, so our spillover results are unlikely to reflect differences in IT human capital among firms, either in firm s existing stock of IT human capital, or in the pool of incoming IT workers. The spillover 23

24 estimate remains unchanged in (4), after including human capital measures for both the incoming IT labor pool and firm s IT workforce in a single regression. A related question is to what extent these spillovers are captured in the labor market through higher wages for more productive IT workers. In (5) and (6), we use our cross-sectional 2006 IT wage data to create measures of IT labor expenses and non-it labor expenses (i.e. total labor expense-it labor expense) which we use instead of employment based measures. Including these wage data in our regressions reduces the spillover estimate to about half of its earlier value (t=3.5), indicating that employees are partially compensated for the IT expertise they bring into the firm from other firms. On the other hand, the coefficient on education is not significantly different than zero after including IT wages. The regression in (6) also includes an explicit measure of the wages of incoming IT workers in Neither the estimate on the wage pool nor the spillover pool is significant, in part due the small number of samples per firm when dealing with the wages of incoming IT workers. We also test if the IT investments of other firms might reflect other attributes of a firm or its neighbors, such as status or the quality of management. We construct an alternative network measure where value-added per employee is substituted for IT-intensity, using the same employee-flow based weights as the IT spillover pool. We jointly estimate these in a productivity framework along with the original IT spillover pool. The results from this regression are shown in Column (1) of Table 12. The coefficient estimate on the IT spillover pool of other firms is.027 (t=5.4), and the coefficient estimate on the value-added measure is not significantly different from zero, suggesting that firms uniquely benefit from the IT intensity of other firms from which they hire new IT workers, not performance levels. 24

25 Finally, we test whether our estimates are uniquely generated by the mobility of IT workers rather than other types of workers, as would be expected if our estimates reflect productivity benefits from the transfer of IT know-how. We construct spillover pools identical to the one using the mobility of IT workers, but using instead the mobility patterns of workers in all other major occupations in our data set, including management, sales, production, and finance. IT workers represent a much smaller number of observations than other types of employees, so to the extent that statistical power is limited by sample size, our test is conservative. Our estimates are shown in Columns (2) and (3) of Table 12. The estimate of.021 (t=4.2) on the spillover pool generated by IT workers is similar to earlier estimates, but the estimate on the non- IT worker spillover pool is not significantly different from zero. These results rule out several alternative explanations. If our estimates reflected unobserved factors related to hiring workers from other high-performing firms, the pool constructed from value-added would have been a superior measure. Furthermore, we would have expected that the effect sizes on the potential spillover pool constructed by using the mobility of other workers would be significant because these workers are just as likely as IT workers to be sensitive to firm reputation. If our estimates reflected returns to overall human capital levels available in IT-intensive firms, its transfer would not be restricted to IT workers. If our estimates primarily reflected geographic effects, or the effects of other spillover mechanisms correlated with employee mobility, than the mobility of other workers, who are more likely to be geographically bounded in their job search than high-skill IT workers, would have provided a superior measure than IT workers. Instead, our results indicate that 1) firms benefit from the ITintensity of other firms, and 2) that these benefits are accessed through incoming IT workers, not 25