Comments are welcome: ptambe@stern.nyu.edu. I am grateful for valuable feedback from Ashwini Agrawal, Roy Radner, and seminar participants at New York University, the Georgia Institute of Technology, and the INFORMS Conference on Information Systems and Technology.

THE DOT-COM CRASH, THE MIGRATION OF TECHNICAL SKILLS, AND

PRODUCTIVITY GROWTH AFTER THE BUST

DRAFT – PLEASE DO NOT QUOTE

Prasanna Tambe New York University, Stern School of Business

ptambe@stern.nyu.edu

November 2011

Abstract

This paper uses data on IT labor and skills to test the hypothesis that IT-sector investment during the dot-com bubble produced a supply of new, e-commerce related technical skills that diffused to firms in IT-using industries after the bust, contributing to a) e-commerce adoption in IT-using firms and b) shifting patterns of productivity growth towards a new set of industries (the “second surge”) after 2000. The effects were strongest in cities where dot-com layoffs were concentrated—especially cities in California—and among technology “leaders” in IT-using sectors. By contrast, the older generations of technical skills demanded by technology “laggards” were readily available from other IT-using firms so the dot-com bust had little effect on these firms. These results are robust to estimators that use the geographic distribution of IT employment as a quasi-experimental source of variation in firms’ exposure to IT labor market changes after the bust. The findings have implications for research and policy related to IT productivity, IT labor markets, and high-tech clusters. Implications for the more recent wave of social media investment and data analysis skills are also discussed. Keywords: IT productivity, IT labor, IT human capital, general-purpose technologies, technology diffusion, financial bubbles, productivity growth, IT spillovers

1.0 Introduction

Although significant academic attention has been paid to the contributions of investments in IT

capital and software to productivity growth, comparatively little emphasis has been placed on how

differences in the supply and quality of technical skills and know-how contribute to productivity growth

differences, despite IT labor accounting for 30-50% of IT spending in modern firms (Gartner, 2010). This

is important because economists have argued that much of the know-how required for installing general-

purpose technologies, such as the electric dynamo or computers, is embodied in skilled technical workers,

and this know-how is likely to be an important factor of production for IT-enabled growth. Furthermore,

due to labor market frictions, differences in technical human capital have the potential to persist across

regions or industry. Therefore, factors affecting the supply of technical skills are of central interest to

understanding variation in IT adoption and productivity.

This paper examines the role of dot-com investment in producing a supply of e-commerce skills

that contributed to productivity growth in other sectors after the dot-com crash. Due to the tacit nature of

the skills required to debug a web site or implement online payment processing, development of new

technical skills requires hands-on experimentation and on-the-job learning. Therefore, while some IT

investment made during the dot-com boom produced intangible assets such as new business processes

and information practices, some of this investment produced technical human capital and new e-

commerce related skills that, unlike investments in many other forms of intangible capital, may have been

transferred to firms in other industries. Therefore, the large waves of investment into IT industries in the

late 1990’s may have been important for developing e-commerce skills that were not lost, but instead

were transferred to firms in other sectors after the dot-com bust. This transfer, in turn, may partially

explain productivity growth patterns after 2000—productivity researchers have observed that productivity

growth shifted from IT-producing sectors towards IT-using sectors after the bust, but the sources of this

shift remain puzzling (Gordon, 2004; Stiroh 2006; Oliner, Sichel, and Stiroh 2007). This paper tests the

hypothesis that the flow of e-commerce skills from IT-producing to IT-using firms after the dot-com

crash contributed to these shifts in productivity growth.1

To address this hypothesis, the paper combines several data sources. A key contribution of the

paper is the analysis of archival data describing the flow of IT skills among firms, data that is normally

not collected by administrative sources or employers. These data are generated from self-reported

                                                                                                                         

1This paper classifies “IT-producing” industries as those defined in Jorgenson, Ho, and Stiroh (2005). The same classification has been used in some prior work (Forman, Goldfarb, and Greenstein 2010). “IT-using” industries are industries that are not IT-producing. The latter classification is slightly different than some prior work that defines IT-using industries to be an IT-intensive subset of non IT-producing industries.

   

employment histories from hundreds of thousands of IT workers and illuminate previously invisible IT

labor mobility patterns. These data on IT labor flow patterns have been used in prior work connecting IT

labor flows to aggregate productivity growth patterns2, and in this paper are used to trace the flow of

technical skills from IT-producing to IT-using firms after the bust, as well as to analyze how these

movements were affected by the dot-com crash. The paper also uses new data on e-commerce adoption

and fine-grained measures of e-commerce skills to provide detailed evidence that the observed

productivity effects are due to the spread of e-commerce and not other technologies.

The analysis primarily uses fixed-effects and differences estimators to test how changes in IT

labor markets after the crash affected productivity growth in IT-using firms. Omitted variable concerns

are addressed by treating the geographic concentration of dot-com layoffs as a source of quasi-

experimental variation in the magnitude of the changes in IT labor prices across different labor markets

immediately before and after the bust. Due to labor market frictions, Firms with IT employees in cities

where dot-com layoffs were concentrated—such as Starbucks or Boeing in Seattle—disproportionately

benefited from the rapid growth in supply and sudden fall in demand from IT-producing firms associated

with the dot-com boom and bust. These differences are consistent with anecdotal evidence relating dot-

com layoffs to changes in the availability of IT skills that were limited to specific cities—for instance,

although firms in Seattle were flooded with job applications after the dot-com crash, Philadelphia-based

firms continued to face an IT skills shortage (Vaas 2001; Cook 2002). This variation is the basis of a) an

instrumental variables analysis of the main productivity regressions as well as b) a difference-in-

differences matching estimator that compares the productivity growth of otherwise similar firms that

differ in the geographic distribution of their IT employment.

To establish that observed productivity effects are principally related to the diffusion of e-

commerce technologies, supplementary analyses are presented that use fine-grained data to improve the

precision of both the dependent and key independent variables. First, the effects of these IT labor market

changes on e-commerce adoption in IT-using firms are tested using a new panel data source on the

adoption of individual technological innovations, created by text-mining firms’ financial reports (see

Saunders and Tambe 2011). The direct use of e-commerce adoption as a dependent variable has the added

benefit that it addresses a number of omitted variable concerns associated with the use of productivity

growth as a dependent variable. Second, cross-sectional data on the e-commerce skills of individual IT

workers in the sample are used to support the hypothesis that the observed productivity effects are a result

of the diffusion of new e-commerce skills, rather than a broader class of technical skills.

                                                                                                                         

2 Tambe and Hitt (2011b) use these IT labor flow data to measure the contribution of IT-related productivity spillovers to productivity growth from 1987-2006.

The analysis produces a number of key findings. First, the dot-com crash resulted in a flow of IT

labor from IT-producing to IT-using firms, primarily for IT-using firms with establishments in cities with

a high density of IT-producing firms. Workers acquired by IT-using firms from IT-producing firms after

the dot-com crash tended to be those with new and rare Internet and e-commerce skills, and a probit

analysis suggests that the acquisition of these skills by IT-using firms after the bust was associated with

the adoption of e-commerce technologies. Furthermore, the transfer of these skills across industries was

associated with faster productivity growth in acquiring firms—these effects appear in fixed-effects and

differences and are robust to the use of location-based instrumental variables and the application of the

matching estimator. The effects are strongest in cities where dot-com layoffs were concentrated (and

especially in California), and for technology “leaders” in IT-using sectors for whom frontier technical

skills were in high demand. By comparison, the older generations of technology skills required by

technology “laggards” were available from other IT-using firms. Finally, the effects are strongest for

retail, financial, and some manufacturing firms, which is consistent with the role that these industries

played in driving post-2000 productivity growth (Stiroh, 2006), suggesting that some of the growth in

these industries after 2000 can be explained by spillovers of e-commerce-related managerial and

technological innovations into firms in these sectors.

This paper contributes to two established academic literatures. First, it addresses an open question

in the productivity literature on the causes of the change in sources of productivity growth after 2000

(Gordon 2004). Notably, the “second surge” in productivity growth that occurred between 2000 and 2005

appears to have been driven by firms in a broader set of industries than the relatively narrower set of IT-

producing industries that drove productivity growth from 1995-2000. Some leading explanations for the

causes of this shift include spillovers of technical and managerial innovations to other sectors or lag

effects associated with the installation of organizational complements in large firms (Stiroh, 2008; Tambe

and Hitt, forthcoming), but the lack of firm-level data on IT investments after 2000 has made it difficult

to resolve this puzzle. This paper tests the spillovers hypothesis. This paper is also therefore closely

connected to an emerging literature on the contribution of various types of IT spillovers to productivity

(Cheng and Nault 2007; Chang and Gurbaxani, forthcoming; Cheng and Nault, forthcoming; Tambe and

Hitt, 2011b).

Second, this is one of the first studies to connect the literature on IT labor markets (Ang and

Slaughter 2001; Ang, Slaughter, and Ng 2002; Mithas and Krishnan, 2008; Levina and Xin 2008; Tambe

and Hitt, forthcoming) to the larger literature on IT productivity (e.g. see Dewan and Min 1997 or

Brynjolfsson & Hitt 2003). Prior work has examined the productivity of IT employment using headcount

or labor expenses (Brynjolfsson & Hitt 1993; Lichtenberg 1995; Tambe & Hitt, 2011a), but how IT

human capital affects productivity has received almost no attention (Bapna et al 2010 is an exception).

   

This is an important omission because a) technical human capital is a poorly understood input into

production, b) prior work suggests significant variation in technical human capital across geographic

regions (Saxenian 1996; Bresnahan and Gambardella 2000) and c) explaining regional differences in IT-

enabled productivity growth has become an active research area (Dewan and Kraemer 2000; Bloom,

Sadun, and Van Reenen 2008; Forman, Goldbarb, and Greenstein 2010; Tambe and Hitt 2011b).

Identifying how the dot-com bust affected local labor markets and productivity growth should also be of

considerable interest to managers and policy makers because unlike IT hardware or software, local

institutions and policies create significant and persistent differences in IT labor quality, which decision-

makers should consider when organizing production or evaluating factors that affect urban growth.

2.0 Background

Computerization has been a key source of economic growth for over two decades (Jorgenson &

Stiroh 2000; Brynjolfsson & Hitt 2003). Yet, there is still little understanding of how the skills and

technical know-how required for computerizing business processes are developed and how they diffuse

through the economy. Unlike R&D know-how, which can typically be codified and transferred through

patent documents or other channels, the skills and expertise required for installing IT innovations tend to

be embodied in the IT labor force.3 Experienced and skilled technical workers can help implement new

technologies, adapt technologies to an organization’s needs, and adapt organizational practices to fit a

new technology. Therefore, the stock of technical know-how embodied in the IT workforce is likely to be

an important factor of production for IT-related productivity growth (Dedrick et al. 2003; Tambe and Hitt

2011a).

From the acquiring firm’s perspective, the supply of technical know-how embodied in the IT

labor pool is affected by the IT investments of other firms in the same labor market for two reasons: a)

learning-by-doing and b) economies of specialization. Learning-by-doing has been theorized as an

important mechanism for increasing labor productivity with new tools and technologies (Arrow, 1962),

and robust empirical evidence linking labor productivity to learning-by-doing has been provided in

contexts as varied as building aircraft (Benkard 2000), constructing naval ships (Thornton & Thompson

2001), and chemical processes (Lieberman 1984). For earlier waves of general-purpose technologies, the

development of a cohort of factory engineers with experience redesigning workflow in electrified

factories has been identified as a key mechanism for the diffusion of the new production methods (David

1990). For IT implementation, trial-and-error in an organizational context is important for the formation

                                                                                                                         

3Breschi and Lissoni (2009) provide recent evidence that most R&D knowledge localization can be explained by inventor mobility. Labor mobility, therefore, may play a disproportionate role even in the context of R&D knowledge diffusion.

of technical human capital and recent studies have provided evidence of learning-by-doing specific in

software development or IT systems implementation (Boh, Slaughter, and Espinosa 2007; Avgar, Tambe,

and Hitt 2011).

Higher IT investment levels within a labor market also lead to specialization in emerging

categories of skills. Economies of specialization have been recognized as an important source of

economic growth since Adam Smith. Rising investment levels within a labor market enable employees

to specialize in cutting-edge and rare combinations of skills, leading to higher labor productivity (Rosen,

1983; Romer, 1987; Kim 1989; Duranton and Puglia 2003). These increasing levels of specialization are

associated with higher productivity levels because workers who concentrate on a narrower range of skills

tend to have a higher marginal product (Becker and Murphy, 1992).

This paper focuses specifically on the technical human capital formed by investments in IT-

producing firms between 1995 and 2000. Although two decades of IT-enabled productivity growth have

spanned multiple waves of IT innovation beginning in the 1980’s, the period associated with the “dot-

com bubble” is viewed by many as being qualitatively different from prior waves of computerization

because of the scale of investment and because of its focus on e-commerce and Internet technologies (see

Forman, Goldfarb, and Greenstein 2010 for a similar claim). IT workers employed in dot-com companies

during these years were exposed to the development and standardization of a new family of technical

skills required for web programming, networking, e-selling and e-buying, and for communicating with

customers online, all of which were required to build infrastructure for the Internet communication driven

culture that emerged during the mid to late 1990’s.

After the collapse of the Internet bubble and the subsequent fall in demand for these skills from

the IT-producing sector, some e-commerce skills diffused to “old economy” companies, such as banks

and retailers developing e-commerce infrastructures, as well as to consulting firms that developed web

sites and e-commerce infrastructures for other old economy firms. One perspective representing the view

of many HR managers towards dot-com workers after the crash was that “employees coming from

startups bring with them a dot-com know-how that includes a comfort zone using interactive tools. They

have learned how to communicate electronically: how to speak to customers online and through e-mail.”

(John Challenger). Yet, anecdotal evidence suggests that HR managers were selective about their hiring,

and that the dot-com boom and bust principally raised demand for the IT skills require for building the

new e-commerce infrastructures: While the economic slowdown and dot-com demise have increased the overall supply of IT workers somewhat, shortages of key skills- particularly those most sought by enterprises- remain as pronounced as ever. There are several reasons for that: One, experts say, is that many dot-com refugees simply don’t have the skills for which enterprise IT hungers; another is that many enterprises have become more picky about the people- and skill sets- they’ll bring on board; and a third is that dot-com layoffs have been concentrated in a few geographic areas. … More and more, those hiring managers are after specific skills- not just anything that fits underneath the umbrella

   

term "IT." According to hiring managers, the skills they still hunger for include networking, customer relationship management, ERP (enterprise resource planning), security and e-commerce/Internet skills such as database management and server administration.” (Vaas, 2001)

Therefore, of the technical human capital developed during the dot-com bubble, it was primarily the e-

commerce related know-how that was demanded by other firms after the bust.

The transfer of productive capital produced during an economic bubble is not new. Prior research

has suggested that technology bubbles are important for the development of capital that drives

productivity growth in later periods. For instance, although companies that invested in broadband

infrastructure disappeared during the dot-com bust, much of this infrastructure was later acquired and

productively repurposed by firms such as Google (Gross 2007). Economists have noted that productivity

growth from 2000 onwards was driven by a different set of industries than the IT-producing industries

that drove much US productivity growth from 1995 to 2000. The scatter plot in Figure 1 illustrates this

shift (from Stiroh, 2006).4 This paper argues that the e-commerce related skills incubated during the dot-

com bubble and subsequently diffused to IT-using firms facilitated this shift. The next two sections

describe the data and methods used to test these hypotheses.

3.0 Data

IT Labor Flow Data

IT labor flow patterns were computed from fielded employment history information for about ten million

users who posted career histories on a leading online jobs board in 2007. In addition to employment

histories, users provided information about occupations (e.g. IT, sales, management, accounting, etc.),

education, and other demographic and human capital variables including IT skills and certifications. This

study focuses on IT workers, self-identified by choosing the IT occupation from a drop-down box.

Because there are no other sources of IT labor flow data, these data cannot be directly validated against

other measures. However, the IT labor data and their sampling characteristics have been benchmarked in

other work (see Tambe and Hitt 2011b for a thorough description of these data). When compared to IT

workers sampled in the 2006 Current Population Survey (CPS), a nationally representative Census

administered survey5, these data are not particularly skewed towards higher or lower education levels.

The average job tenure of IT workers in our sample is approximately two years lower than the average job

tenure of IT workers in the CPS survey. A key reason for this difference is that the sample contains a

                                                                                                                         

4  Note: the “IT-using” firms in this analysis correspond to both the “IT-using” and “Other” categories in Figure 1. 5The 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 http://www.census.gov/cps/.

disproportionate number of young workers and job-hoppers. However, this may not be a significant

limitation because the key population of interest in this study is the “job-hopping” IT worker, so the

sample may be a better representation of this population than the CPS sample, which includes workers

who never or infrequently switch jobs. The 2006 wages of the workers in our sample are similar to the

2006 IT wages computed in the Occupational Employment Survey. Nevertheless, IT workers that post on

this jobs board may still be different than other job-switchers along unobservable dimensions. This

selection concern is addressed further below.

IT Skills and Location Data

This study introduces new data on technical skills and certifications, self-reported by IT workers.

From these skills and certifications, IT workers are identified who report having e-commerce and Internet

related skills such as HTML, Web programming, or SQL programming.6 To the best of my knowledge,

these data represent the first large-scale measures of skills in the IT labor market. However, one

significant limitation of these data is that they are reported only in 2006. Therefore, assigning these skills

to a particular employer-employee pair in years prior to 2006 produces measurement error if workers

acquired the skill after the year of assignment. This measurement error will be larger in earlier years than

later years, and will be one-sided— it will misclassify IT workers in some years as having an e-commerce

skill who have not (yet) acquired that skill but will not omit any IT workers who do have the skill.

However, as long as these misclassification errors are exogenous to the hiring firms’ productivity levels,

measurement error of this type will only tend to make it more difficult to find a meaningful result when

using these skills data.

Metropolitan areas are also reported by workers in 2006 and in combination with the employer

data are used to create measures of the extent to which IT-using firms participate in the same labor

markets as IT-producing firms. For each firm i, a labor market index ITLM is computed as:

!"#$! = !!"(%!"#$%&!)!

!!!

where w is the percentage of a firm’s IT workers in metro-area j, and %ITPRODj is the percentage of IT

workers in metro-area j employed at IT-producing firms and j indexes all of the J metro areas in which a

firm employs IT labor. It is worth noting that there are, to the best of my knowledge, no alternative data

sources describing the distribution of a firm’s IT workers across US cities7, so the ability to create this

                                                                                                                         

6 The complete list of skills is chosen using Foote Consulting’s list of “hot” IT e-commerce skills. 7 The closest alternative are the CITBD data published by Harte Hanks which only have information on the number of programmers in establishments with at least 100 people.

   

measure using the firm-city distribution is somewhat unique and improves the accuracy of the analysis

over one in which IT employee locations would be approximated using the location of firms’ corporate

headquarters.

Table 2 lists the cities with the highest values for %ITPROD. The top ten regions ranked

according to this measure are consistent with prior expectations, and include San Jose, Seattle, and San

Francisco. The firm-level ITLM index value, therefore, will be highest for firms that employ a large

percentage of their IT workforce in cities such as these where dot-com layoffs were concentrated. IT-

using firms with a high ITLM index would have been the first firms to acquire the skills and human

capital displaced from IT-producing firms. Like the skills data, these IT worker location data are also

reported only in 2006 cross-section, but the magnitude of the measurement error produced by this data

limitation is bounded by the rate of entry and exit for establishments in different cities. The size of the

measurement error should be the fraction of establishments that in 2006 were in different locations than

they were in 2001. For much of this analysis, firms are grouped according to ITLM quintiles for ease of

interpretation.

Selection Concerns and the Direction of the Bias Term

A concern regarding sample selection is that although the measure of interest is the network of

firms defined by all IT labor movements, IT workers who post on job boards may be systematically

different than other IT worker who switch employers. If the historical movements of jobs board

candidates are an accurate representation of the other employers in a firm’s IT labor pool (weights

between firms), then there is no measurement error. However, systematic differences between these

workers and the larger population of job-hoppers introduces error into the measurement of this network of

firms. However, this error term should only impose an upward bias if for a focal firm, the network of

firms defined by the movements of IT jobs boards candidates is more likely than other types of candidates

to transfer skills and human capital, which is somewhat unlikely if the concern about jobs board

candidates is one of adverse selection. In other words, selection problems will tend to lower the estimate

on our parameter of interest by overweighting the investments of the ex-employers of workers who post

on jobs boards. Furthermore, because most of the analysis is in differences, this bias is most severe when

the measure of the IT-producing pool disproportionately rises as a result of the dot-com bust, which will

only be the case if the jobs board over samples IT labor in IT-producing sectors, which is also

inconsistent with adverse selection.

A similar analysis can be applied to the use of the most recently reported IT employment location

to create measures of the geographic distribution of a firm’s IT workers. The measure of interest is the

geographic distribution of a firm’s IT hires. If the workers in a firm’s IT labor force that are posting on

the jobs board are geographically distributed in a different way than the true distribution of firm’s new IT

hires, there will be error in the geographic measure. This will tend to overweight the geographic presence

of establishments in which employees are leaving at disproportionately high rates (e.g. layoffs,

offshoring). This imposes an upward bias if employees entering these establishments are

disproportionately likely to transfer productivity spillovers but this is also unlikely to be the case because

events that would raise the rate of jobs board postings in specific establishments within a firm will tend

to reduce hiring in a specific location from the underlying base rate.

IT Investment Data, E-commerce Adoption Data, and Compustat Economic Data

The IT investment data are from the IT labor series constructed and described in prior published

work (Tambe and Hitt, forthcoming). These data are based on IT labor intensity within the firm and have

been shown to perform comparably with other IT datasets such as the CITDB capital stock data in

productivity regressions. For the purposes of this analysis, they have the advantage that they are collected

in a consistent manner in the years leading up to and after the dot-com bust. The e-commerce and data-

mining adoption data used in this study are created by text-mining firm’s financial reports between 1995

and 2010 and these data and their error characteristics are also described in other work (Saunders and

Tambe 2011). Other economic data used in productivity regressions, such as value added, capital, non-IT

employment, industry, and sales, were collected from the Compustat database. Variables were created

using methods common in the micro-productivity literature.

4.0 Methods and Measures

Main Productivity Regressions

To test the main hypothesis, the productivity regressions treat the IT investments of IT-producing

and IT-using firms as separate factors of production and estimate how productivity growth is associated

with the growth of each of these factors. The Cobb-Douglas functional form is used because it has been

the most commonly used specification in the literature on IT productivity and forms the basis for US

productivity growth measurement (Brynjolfsson and Hitt 1996; Dewan and Min 1997). The starting point

for the analysis is the large literature that estimates the social returns to R&D investments (see Griliches

1992 for a review of this literature). In logs, the most basic productivity model that includes a measure

of external IT investment is:

(2) !" =∝! ! +∝! ! +∝!" !" +∝! ! + !"#$%"&' + !!"#

where all lowercase variables are in logs, k is capital stock, e is employment, it is a firm’s own IT

investment levels, and s is the firm’s external pool of IT investment. Similar specifications have been

   

used in much prior work on R&D productivity and have recently been adapted for the study of returns to

external IT investments (e.g. see Chang and Gurbaxani 2010).

IT investment and IT intensity have been used in prior research as a measure of the firm’s

technical capabilities (Bharadwaj and Bharadwaj 1999; Brynjolfsson, Hitt, and Yang 2002; Saunders

2010)—these studies argue that firms with greater IT intensity tend to have superior IT capabilities and

superior performance. Furthermore, higher demand for frontier technologies in IT-intensive firms

suggests that these firms are more likely to employ technical employees with newer and specialized

technical skills—therefore, firms investing in frontier technologies will tend to obtain these skills from

other IT-intensive firms that have made similar investments. The use of IT intensity as a measure of the

skills embodied in individual workers is also supported by a literature demonstrating that firm-level

investment in R&D is a superior measure of human capital accumulation than calendar time or other

measures (Lieberman 1984).

Variation in the measure of the external IT pool across firms requires measuring differences in

firms’ abilities to capture returns from this pool of investment (Griliches 1992). In keeping with the

skills-based mechanism of productivity transfer proposed in this paper, measures of each firm’s pool of

external IT investment are computed by weighting the IT-intensity of other firms by the share of IT

employment flows from that firm in each year. Because the hypothesized channel for this productivity

transfer is specialization and the accumulation of technical human capital, this tests the hypothesis that

firms derive productivity benefits from the IT investments of firms from which they acquire IT labor. To

test the hypothesis that IT-sector investment generated a new stock of technical skills, the analysis further

separates firms into two sets8, treating the IT investments made by IT-using and IT-producing firms as

separate factors of production. The aggregate IT investment measures from the two different sectors are

computed as in (3) and (4).

(3) !!"! = !!"#!"!"

!"!!"#$!!!

(4) !!"! = !!"#!"!"!"!!"#$%!!!

Sit is the IT pool of firm i in year t, the p and u superscripts signify investments in either IT producing or

IT using sectors, IT is the IT intensity of firm j in year t, the weighting term (Φijt) is the fraction of firm i’s

incoming IT workers (summed across both IT producing and IT using firms) that come from firm j in

year t and the superscript signifies that the summation is restricted to IT producing or using firms. The

employee weights across both measures add to 1. A shift in the acquisition of IT labor towards IT-

                                                                                                                         

8  Orlando (2005) is another example of an analysis that separates external investment pools into sets to separate the contribution of each.  

producing firms corresponds to a shift in the fractional weighting terms that increases the value of the

measure in (3) relative to that in (4).

The final productivity model that incorporates both IT investment pools and accommodates

differences in how these pools contributed to productivity before and after the bust is:

(5) !" =∝! ! +∝! ! +∝!" !" +∝!!! !! +∝!!! !! +∝! !! ∗ ρ +∝! !! ∗ ρ +∝! ρ + !"#$%"&' + !!"#

Lowercase variables are in logs, k is capital stock, e is employment, it is a firm’s own IT investment

levels, su is the IT know-how of IT-using firms transmitted through IT labor flows and sp is the IT know-

how of IT-producing firms transmitted through IT labor flows, ρ is a dummy variable that assumes the

value 1 in the years after the dot-com crash and 0 for all other years. The model in (5) is tested in fixed-

effects and differences. Fixed-effects models remove most sources of firm-specific unobserved

heterogeneity. However, differences estimates provide more suitable controls for idiosyncratic firm

characteristics that evolve over time. The hypothesis that the flow of skills from IT-producing industries

contributed to productivity growth in other industries after the bust is supported if the coefficient estimate

∝!!!!  is significantly greater than zero. This specification is used to estimate effects across the entire

sample of IT-using firms, as well as within specific regions and industries. Where appropriate, standard

errors are clustered on firm. The next sections describe approaches used to address endogeneity concerns

associated with estimating (5).

Endogeneity

The regression approach described above filters out many omitted variable concerns. It uses both fixed

effects and differences, and any remaining omitted variables that bias the value of ∝!!!!  must share the

idiosyncratic timing of the dot-com bust—they must be correlated with the pool of IT investment

captured through IT labor from the IT-producing sector, but only after 2000. However, a serious

remaining concern is that high-growth IT-using firms were most successful in acquiring IT skills from the

IT sector after the bust. To address these, the analysis takes two different approaches both of which are

described below.

Instrumental Variables

To test the causal direction of the relationship in (5), a quasi-experimental source of variation in

IT labor price changes is differences in dot-com layoffs across IT labor markets. Because layoffs were

concentrated in a relatively small number of cities, changes in IT labor prices after the crash also varied

significantly across cities. If IT-using firms are otherwise evenly distributed across labor markets, these

differences identify the effects of these changes on productivity growth for IT-using firms. This analysis

uses two instrumental variables based on geographic location: a) the labor market index described above

   

(ITLM) and b) the total layoffs from IT-producing firms in the county of the firm’s corporate

headquarters. These variables are meant to capture the extent to which the dot-com bust affected a firm’s

local IT labor market. The descriptive evidence in the next section presents evidence that skills from IT-

producing firms flowed into IT-using firms within the same metropolitan areas. However, the validity of

this research design requires productivity growth to be randomly distributed over the different labor

market categories in the absence of the dot-com bust.

Matching Estimator

A second “experimental” comparison uses a difference-in-differences matching estimator. Matching

estimators have the advantage of being non-parametric, and have been widely used in the empirical

economics literature on program evaluation (e.g. see Heckman, Ichimura, and Todd, 1997; Abadie et al

1994). Firms are placed in the treatment group when they are in the top ITLM quintile, and the treatment

effect is estimated by comparing differences in productivity growth immediately after the bust for a

treated firm with that of another, otherwise similar firm that differs only in its exposure to thick IT labor

markets. The analysis matches firms on total employment, the location of its corporate headquarters, and

industry. Therefore, the productivity growth effect in this analysis is estimated, for example, by

comparing two similar-size pharmaceutical companies located in New Jersey, where one firm operates an

IT development center located in a high-tech city. For this analysis, using matching estimators to

complement the prior analyses has the advantage that the results do not rely on the specification of the

external IT pool, and therefore estimate the average productivity growth effect of being located in these

regions, irrespective of their success in hiring IT labor.

Evidence for E-commerce

The approaches described above link IT investment in the IT sector to subsequent productivity

growth in IT-using firms. The analysis takes two approaches to further establish that these effects reflect

the diffusion of new e-commerce technologies. First, results are reported from an analysis that associates

the flow of IT labor to e-commerce adoption, rather than productivity growth. The use of e-commerce

adoption as a dependent variable is somewhat less subject to the endogeneity concerns that accompany

productivity growth. To test the hypothesis that IT labor from IT-producing industries specifically

facilitates e-commerce adoption, a probit regression of the following form is tested:

(1) !"#$$ =∝!!! !! +∝!!! !! +∝(!!!)! !! ∗ ρ +∝(!!!)! !! ∗ ρ +∝! ρ + !"#$%"&' + !!"#

The dependent variable ecomm is a binary variable indicating whether a firm has invested in e-commerce

technologies in any prior year. This dependent variable is measured using a text-mining dataset described

in other work that captures whether a firm mentions making e-commerce related investments in their SEC

filings (Saunders and Tambe 2011). On the right-hand side of the equation, su and sp are the IT

investments of the relevant group of IT-using firms and IT-producing firms respectively, ρ is a dummy

variable that assumes a value of 1 in the years after the dot-com crash (2000 and onwards) and 0 for prior

years (these variables are identical to the aggregate investment variables used in the productivity

regression). An estimate of ∝(!!!)!  that is significantly greater than zero indicates that IT-producing

investment prior to the dot-com bust had a significant association with IT-using firms’ adoption of e-

commerce technologies through the flow of technical skills. Standard errors for all probit regressions are

clustered on firm. The robustness of this result is further tested by using the adoption of other types of

technologies as the dependent variable. For example, if skills developed during the dot-com bubble are

primarily e-commerce related, the pool of investment associated with the ∝(!!!)! coefficient should not

have a significant economic impact on the adoption of non e-commerce related technologies.

The second approach to establishing that any observed productivity growth is an “e-commerce

effect” addresses the limitations of using the flow of undifferentiated IT labor in creating the weighting

matrix. The use of aggregate IT labor flow data does not distinguish between the flow of old and new

technical skills although a key argument of the paper is that IT-sector investment produced a new

category of e-commerce skills. To test the hypothesis that the productivity growth results reflect the

returns to e-commerce skills, alternative measures of the pools of external IT investment are computed

using only the flow of e-commerce labor (i.e. any IT worker reporting having at least one e-commerce

skills) to create the weighting terms for the IT-intensity of prior employers. When embedded in a

production function, this measure connects productivity growth with IT-producing sector investment

specifically through e-commerce skills, which suggests that productivity benefits were primarily

transferred through the flow of new e-commerce skills rather than aggregate IT skills. The statistical

power of this measure is limited relative to the first measure due to the cross-sectional nature of the e-

commerce skills data. However, it can provide corroborating evidence that the observed productivity

effects are in fact returns to a shift in prices for e-commerce skills rather than technical skills overall.

5.0 Descriptive Statistics

IT Labor Flows

Figure 2 uses Bureau of Labor Statistics (BLS) data to illustrate the employment collapse that

occurred in IT-producing industries after the dot-com crash. The job separation rate rose from 40,000-

50,000 annual separations in IT industries during the late 1990’s to a rate of about 200,000 in 2000, and

remained at higher than normal rates through 2003. These changes indicate a qualitative shift in the levels

of IT labor flowing out of IT sectors during the dot-com bust and for a few years afterwards.

   

The micro-data used in this study enable detailed analysis of the migration paths of IT labor

displaced from the IT sector. Figure 3 shows industry destinations for IT employees leaving IT sector

firms during the years surrounding the dot-com crash. A drop in the percentage of displaced IT workers

hired back into IT-sector firms was offset by a rise in the percentage of IT workers entering firms in the

retail, finance, and manufacturing sectors. Figure 4 shows source industries for IT labor hired into the

finance, real estate, and insurance (FIRE) industries in the years surrounding the dot-com crash; the trends

indicate a sharp rise in the percentage of IT labor acquired from IT-sector firms during and after the dot-

com crash, offset by smaller percentages of IT labor being hired from within the FIRE sector.

These figures reflect aggregate trends across regions and skills, but anecdotal evidence suggests

significant differences in the geographic concentration of IT layoffs during the dot-com bust. For IT-

using firms, Figure 4 shows changes in the percentage of IT labor acquired from the IT sector, separated

by labor market quintile (bottom, middle, and top 20% by ITLM). The flow of IT labor from IT-

producing to IT-using firms was principally concentrated in firms with establishments in the highest

ITLM quintile. There is comparatively little effect for firms in the middle and lower labor market

quintiles. These differences suggest that IT-using firms located close to IT-producers may have

disproportionately benefited from the flow of IT skills out of the IT sector after the dot-com bubble bust.

It is also somewhat interesting to note that prior to the dot-com bust, IT-using firms in the top labor

market quintile absorbed a smaller percentage of workers from IT-producing firms than IT-using firms in

the middle quintile. This is consistent with claims of skills shortages in these cities—due to high levels of

IT labor demand in these regions during this period, IT-using firms may have had a particularly difficult

time hiring workers when competing with IT-producing firms.

This effect becomes significantly more pronounced after narrowing IT labor to those employees

with e-commerce skills in Figure 5. For IT-using firms in the top labor market quintile, the fraction of IT

workers with e-commerce skills from IT-producing firms jumps significantly, from about 20% before the

crash to 80% after the crash. By contrast, there is little change in the composition of acquired IT labor for

non e-commerce related skills, indicating that the patterns observed in the prior figures were primarily

driven by the migration of e-commerce skills from IT sectors to other industries. Collectively, these

figures suggest that the dot-com bust was associated with a flow of e-commerce and Internet skills from

the IT sector to IT-using firms, but that IT-using firms that were located in the same geographic regions

as IT-producing firms disproportionately benefited from this shift.

External IT Investment

Figures 7 through 10 consider growth in external IT investment as a factor of production and

examines how this input changed during the dot-com bubble and bust in different sectors and regions.

Figure 7 examines how external IT investment changed over time in select industries. The vertical axis in

Figure 7 is the mean-removed growth in the measure of external IT investment—this measure rises when

the fraction of IT labor acquired from IT-intensive firms rises faster than at other firms. For IT-sector

firms, this growth rate peaks in the late 1990’s, which is consistent with IT investment levels during that

period and the flow of IT labor within the IT sector during this time period. The timing of the transfer of

the technical skills developed in this industry is suggested by the lines associated with the financial and

retail industries. This rises for finance firms as it begins to fall in the IT sector but before the dot-com

crash, suggesting that financial firms are not sensitive to changes in the price of new IT skills. However,

rising levels of this measure for retail firms correspond with falling demand for this input from the IT

sector.

Figure 8 illustrates these trends for financial firms, deconstructed by sector to illustrate how much

of the shift in this measure is driven by IT-producing firms. The measure is rising in both sectors in

Figure 8, but there an apparently sharp upward shift in the measure coincides with the dot-com crash.

One of the main arguments of this paper is that this shift was qualitatively different in its skill content and

its geographic distribution. Therefore, to sharpen the comparison, Figure 9 narrows IT labor to workers

with e-commerce skills and to IT using firms located in the highest IT labor market quintile. These

adjustments produces a sharp rise in external IT measure for IT-using firms, driven almost exclusively by

IT-producing firms and suggesting the emergence of a new class of skills from the IT sector. By contrast,

Figure 10 displays the same e-commerce based measures for firms in the lowest IT labor market

quintile—for these firms, the dot-com bust did not produce a comparable increase. The next section

reports results from regressions that analyze the associations between the variation in this production

input and a) e-commerce adoption and b) productivity growth.

Aggregate Productivity Growth

Finally, figures 9 through 11 provide some preliminary evidence associating these changes with

faster productivity growth. Figures 9 and 10 compare the productivity growth of IT-using and IT-

producing firms from 1999 to 2003 where the vertical axis corresponds to productivity growth and the

categories on the horizontal axis correspond to the five IT labor market quintiles. There is a distinct

difference in the relationship between productivity growth and labor markets for IT-producing and IT-

using firms, with IT-using firms in thicker IT labor markets achieving higher levels of IT productivity

growth during these years, and IT-producing firms in these markets doing relatively poorly, consistent

with a potential transfer of assets across organizations for firms in these areas. For comparison, Figure 11

indicates that prior to the dot-com bust, IT-using firms in thicker IT labor markets did not experience

particularly high levels of productivity growth from 1994 to 1998—in fact, firms in the upper labor

   

market quintiles appear to have been somewhat less productive in the years leading up to the dot-com

bust. The next sections attempt to more rigorously test how the IT labor flow patterns observed above are

associated with these productivity growth patterns.

6.0 Regression Results

Main Productivity Regressions

Table 3 presents productivity estimates from equation (5). The estimates from columns (1) and

(2) use fixed-effects estimators which remove the effects of time-invariant omitted variables, including

most persistent differences associated with firms’ positions in the IT labor flow network. The estimates in

(1) are from IT-using firms only and indicate that faster productivity growth is associated with growth in

investment from both IT-using (t=2.56) and IT-producing industries (t=3.30) after the dot-com bust. The

results in (2) are restricted to firms in IT-producing industries and do not reject the hypothesis of no

statistical association between productivity growth and growth in the investment of either sector.

As discussed earlier, fixed effects estimates may not provide suitable controls for idiosyncratic

firm characteristics that evolve over time. Therefore, in columns (3)-(6), differences results for IT-using

firms with varying difference lengths are presented. In addition to providing the same control for firm-

specific effects as a fixed-effects analysis, the coefficient estimates at varying difference lengths can be

interpreted as a comparison of the short run (1st differences) versus long run (3 year or more differences)

(Bartelsman, Caballero and Lyons 1994). Longer differences may also be less subject to biased estimates

due to measurement error (Greene 1993). Estimates from the differences regressions also suggest a

statistically significant association between the acquisition of IT skills from IT-producing firms after the

bust and productivity growth but the point estimates do not become significantly different from zero until

the differences window is extended to four years (t=2.22).9 For comparison, the results from a four-year

differences regression on the sample of IT-producing firms are reported in (7). These indicate that the

acquisition of skills from other IT-producing firms has a significant association with productivity but it

does not change after the dot-com bust. Together, the results from the differences estimates are consistent

with the interpretation of IT-producing firms as incubators for skills that drove productivity spillovers

within the IT producing sector during the dot-com bubble and in IT-using sectors after the bust.

Productivity Regressions by Region, IT investment, and Industry

Dot-com layoffs were geographically concentrated, and due to housing and family constraints,

displaced IT labor is more likely to seek employment within the same metropolitan area. Therefore,                                                                                                                          

9 They also remain significant for five and six year differences, but are not shown for compactness.

short-run productivity differences may have been stronger in regions where dot-com layoffs were

concentrated. Table 4 shows productivity regressions separated by IT-producing/IT-using industry and by

whether the firm is in the top or bottom half according to the labor market index (ITLM). The results in

(1) and (2) suggest that the transfer of know-how from IT-producing firms was only important for firms

in cities in the top half scored by ITLM index, and that these effects appeared only after the dot-com crash

(t=2.33). This is consistent with the geographic concentration in IT layoffs restricting the transfer of

know-how to specific regions. For IT-using firms that were not located in these cities, the dot-com bust

did not have a discernable productivity impact through the transfer of technical know-how. Comparable

results from firms in IT-producing industries are in (3) and (4). The results indicate that IT-producing

firms in high ITLM locations appear to benefit from the flow of skills from other IT-producing firms

(t=1.72), but as expected, these effects are not significantly different after the dot-com bust. IT-producing

firms that are not located in IT-intensive cities do not derive comparable productivity benefits from the

flow of skills.

Columns (5) and (6) estimate the impact of these skills on firms that are technology leaders and

technology laggards, respectively, measured by whether or not the firms’ total IT intensity levels are

above or below mean levels relative to other firms. For technology laggards, the productivity effects are

weakly significant for the acquisition of technical know-how from IT-using firms (t=1.67) but there is no

effect from IT-producing firms, and the dot-com bust had no impact on productivity growth for these

firms—the effects of acquiring IT labor from IT-using or IT-producing firms are not significantly

different before and after the bust. On the other hand, for technology leaders, there is a positive and

significant association between changes in the acquisition of technical know-how from both sectors and

productivity growth after the bust. These results suggest that the flow of IT skills incubated in IT-

producing sectors affected productivity for IT leaders, for whom access to the skills needed to install

frontier technologies would have been in demand. For IT laggards, who may have been installing older

generations of technologies, the necessary skills were already freely available from a broader set of IT-

using firms. Therefore, the changes in the supply of cutting-edge skills caused by the dot-com bust would

have had little impact on productivity growth.

Finally, columns (7) and (8) focus the analysis on California, the region most closely associated

with dot-com investment and dot-com layoffs. Although the productivity effects of IT sector labor

acquired after the dot-com bust are significant in firms both inside and outside of California, the effect

sizes are much larger for California firms (t=2.79). Overall, the results in this table are consistent with the

interpretation that technology leaders would have found it beneficial to locate in high-rent locations (such

as Silicon Valley) for access to cutting-edge technical skills from IT-producers. When technology

   

bubbles burst, firms in these regions will derive the strongest productivity benefits. IT laggards, on the

other hand, may be served equally well by locating away from these locations.

In Table 5, estimates are reported using the same fixed-effects specification as in Table 4, but the

sample of firms in each column is limited to a single one digit SIC industry. The results indicate that

productivity spillovers from IT-producing firms after the dot-com bust primarily benefitted firms in

specific industries: heavy and light manufacturing and retail industries. The productivity effects are not

significant for firms in other industries. It is interesting to compare this pattern of results with the

productivity growth results after 2000 from prior work. The estimates from these regressions are

consistent with the relatively fast pace of productivity growth observed in retail and durable and non-

durable manufacturing industries between 2000 and 2004. Although it is likely that this shift had a

number of causes beyond computerization, these results indicate that productivity spillovers of

managerial and technological innovations related to e-commerce were at least one important reason.

Instrumental Variables and Matching Estimator

The results from IV regressions using different sets of instrumental variables are reported in

Table 6. The model is a four-year differences regression, restricted to the years after the dot-com bust and

instruments are used to explain variation in the measure of the IT pool from IT-producing sectors.

Column (1) uses both the IT labor market variable and the IT layoffs variable as instruments. The

estimates from this model are significant (t=1.82), but the first-stage regression has fairly little

explanatory power, resulting in inflated estimates of the effect of IT-producing investment on IT-using

productivity in the second stage. The over-identification statistic (Hansen’s J) is not very close to zero, so

we cannot reject the hypothesis that the instruments are uncorrelated with the disturbance term. Columns

(2) and (3) add interactions between IT labor market and industry to the instrument list. Adding industry

at the 2-digit level significantly increases the explanatory power of the first-stage and lowers the IT pool

estimate closer to the values from our main productivity regressions, but the over-identification statistic is

close to being significant. In aggregate, these various IV regressions suggest that using location as a

quasi-experimental source of variation in IT labor market changes supports the hypothesis that IT-

producing investment contributed to IT-using productivity growth through the development of a body of

technical skills. The IV statistics reported at the bottom of the table are consistent with the idea that ITLM

classification was exogenous with respective to post-bust productivity growth for IT-using firms. In (4),

the instruments in (2) are applied to a sample of IT-producing firms and the estimates for this regression

are insignificant.

The results from a matching estimator are reported in Table 7. The sample used for the matching

estimator includes only firms in the highest and lowest IT labor market categories and matches firms that

are similar in employment size, industry, and the location of corporate headquarters, but differ in their IT

labor market classification. The results in the top row are from 4 year productivity growth differences

from 2001 to 2005 matched on different sets of variables where “exact” matches are enforced where

possible. The first column matches firms on size and the state of corporate headquarters with no

restriction placed on industry, and where exact matches are enforced on state. The percentage of exact

matches is high—over 90%--and the results indicate faster productivity growth levels for firms in the top

IT labor market quintile. Columns (2) through (4) use different levels of industry classification for

matching, from 1 digit industry through 4 digit industry, where increasing the precision of the industry

classification reduces the percentage of exact matches. The estimates across all combinations of

matching variables indicate that firms in the highest ITLM category experienced significantly faster

productivity growth than firms in the lowest ITLM category, conditional on size, industry, and the

location of corporate headquarters. Column (5) further narrows the location match for firms’ corporate

headquarters to state and county. This significantly reduces the percentage of exact matches but the

results are not substantially different.

The second set of results at the bottom of the table differ only in that the four year period

considered is 1995 through 1999, which predates the dot-com bust. Notably, the productivity effects of

being in the highest IT labor market quintile disappear entirely, and the estimates weakly suggest that

firms in the highest IT labor market category during this time period were actually less productive than

their competitors in other locations. Therefore, it does not appear that the faster productivity growth

reported in the top row were due to persistent differences in firms that locate establishments in these

regions.

Evidence for an E-commerce Based Explanation

A remaining question is whether the observed productivity effects are due to the diffusion of e-

commerce technologies, or can be explained by the spread of other types of technical human capital out of

the IT sector. Table 8 presents results from probit adoption regressions that span the years 1995 to 2006

and use e-commerce adoption as a dependent variable. The binary dependent variable in (1) takes the

value 1 if the firm indicates making e-commerce related investments in its 10-K financial reports in any

prior year and 0 otherwise. The results from (1) suggest that the acquisition of IT labor after the bust is

positively associated with the adoption of e-commerce technologies, but these effects are only weakly

significant (t=1.93). In (2), the sample is limited only to firms in the two highest IT labor market

quintiles. After restricting the sample to these regions, The effects are larger and more precisely

estimated (t=2.81), indicating that the e-commerce adoption benefits that IT-using firms derived from IT

firms after the bust were primarily within high-tech cities.

   

As a falsification test, the adoption of two other technologies, data mining and ERP, are used

instead of e-commerce as the dependent variable. If dot-com investment primarily produced e-commerce

related skills, know-how from the IT sector made available after the bust should not have statistical

associations with the adoption of other technologies. The regression in (3) uses the adoption of data

analysis technologies—including storage, analytics, machine learning—rather than e-commerce

technologies, as the dependent variable, and the sample is limited to firms in the top labor market

quintiles for direct comparability with (2). The estimates from this regression indicate that the adoption of

data technologies is associated with the flow of skills from other IT-using firms (t=3.73) with no

significant differences before or after the bust—in other words, the skills required to install these

technologies were not produced in the IT sector during this time period, but instead, were broadly

circulated among firms in all sectors. Using ERP technologies in (4), rather than e-commerce

technologies, produces no statistical associations between ERP adoption and any of the measures of IT

investment during this time period.

These analyses indicate a specific relationship between the acquisition of IT labor from the IT

sector and e-commerce adoption. To provide further support for the causal direction of this relationship,

the probit analysis in (5) is restricted to years after the dot-com bust, and uses the same location-based

instrumental variables that were used for the productivity regressions above as instruments for the change

in the structure of the IT labor flow network. The association between access to skills produced in the IT

sector and subsequent e-commerce adoption persists in the IV regression (t=3.05) suggesting that the

change in demand for e-commerce skills benefitted IT-using firms in dot-com cities, facilitating e-

commerce investments in these firms.

Table 9 reports estimates using measures of the external IT investment pool created using the e-

commerce skills data. OLS estimates are reported first due to cross-sectional limitations with the skills

data—they are most accurate in recent years, so the distribution of these skills across firms is more

representative than within-firm changes in this distribution over time—furthermore, the number of

samples of movements for workers with these skills is relatively small compared to the number of

samples of workers in firms, so the signal to noise ratio will be higher across firms. The OLS estimates in

columns (1) and (2) suggest that the IT investment of firms from which IT skills were acquired had a

positive association with productivity after the bust (t=1.87) and that the productivity effects of the

broader measure of labor-weighted IT investment disappears when jointly estimated with a measure based

on e-commerce weights. There are also significant association with the IT-using pool (t=1.75), although

these do not appear to be significantly different before or after the dot-com bust. These effects disappear

in the fixed-effects estimates in (3) and (4) which is likely due to the poor signal to noise ratio in the

longitudinal dimension of the skills data. The fixed-effects estimates when including the e-commerce pool

measure are similar to those reported earlier with no e-commerce measures included in the model.

Nevertheless, the fact that the productivity effects of the technical know-how measure after the bust

disappear when including direct e-commerce measures provides some support that the earlier results

specifically reflect the transfer of e-commerce know-how from IT-producing to IT-using firms.

7.0 Conclusion

The main finding of this study is that the technical human capital developed in IT sectors during

the dot-com bubble contributed to productivity growth in IT-using sectors after the bust. These results are

robust to multiple specifications and dependent variables. They are also robust to IV regressions that use

the geographic concentration of IT layoffs after the dot-com bust as a source of quasi-experimental

variation in changes in the local IT labor pool. Productivity benefits were largest for IT-using firms

located near IT-producing firms, suggesting that firms in IT-intensive cities had the best access to e-

commerce know-how immediately after the dot-com crash. The effects are particularly strong in

California, and they are stronger for technology-intensive IT-using firms, for which access to cutting-edge

e-commerce skills would have been a bottleneck to technological progress. For IT-using firms investing

in older generations of information technologies, the technical skills required for productivity growth

were readily available from other IT-using firms.

For managers, these results provide insight into how a supply of technical skills is created and

diffused across region, time, and industry. After the dot-com bust, hiring managers commented on the

skills and broad e-commerce exposure that dot-com employees brought to new employers. These findings

provide empirical support for these statements, especially regarding the importance of location for access

to new and rare skills developed through on-the-job learning at other firms. However, the estimates

suggest that access to these skills is important only for IT leaders. For IT laggards, regional location may

not be particularly important. These findings also have implications for understanding how technology

bubbles facilitate the development of new skills which can diffuse to other industries. For instance, a

concern with the current “social media” bubble was voiced by an early Facebook employee, who was

quoted as saying “The best minds of my generation are thinking about how to make people click ads. That

sucks.” (Vance, 2011) This study suggests that the “big data” skills being developed in social media

companies may be productively transferred to companies in the financial, retail, and healthcare sectors as

firms in these industries begin to exploit valuable databases using techniques that are currently being

pioneered by social media companies. Consequently, new skills related to big data (e.g. Hadoop and

MapReduce) that are currently concentrated in a few cities like Sunnyvale, CA may diffuse to firms in

other industries if and when there is a shakeout in the social media sector—with firms in close geographic

proximity to social media companies being among the first to benefit.

   

For policy makers, these results provide some insight into the importance of technical skills for

productivity growth, and the important role that dot-com bubble investment played in incubating a supply

of technical human capital. Although evaluating the efficiency of technology bubble investment through

this lens is well beyond the scope of this paper, these results provide evidence that at least some of the

“intangible assets” produced through dot-com investment were transferrable to other firms. The

importance of on-the-job learning for the development of technical skills suggests that regulatory

environments that encourage experimentation with new technologies may be an effective way to develop

a supply of skills around new technologies. The regional results in this paper also provide support for

policies that encourage investment to facilitate the development of a skilled workforce within a region.

They also suggest the importance of encouraging labor mobility for economic growth, which is relevant

to the assessment of laws such as non-compete agreements or health insurance portability that affect labor

mobility rates.

Finally, these findings suggest directions for future research. This paper suggests the importance

of technical skills for productivity but does not discuss how to attract and retain high-quality IT workers.

Research on the role of compensation and other firm-level factors for locating, attracting, retaining, and

motivating skilled IT workers to complement a firm’s technical initiatives would provide important

guidance about how firms can build a productive IT workforce. Furthermore, this study discusses the

productivity benefits of the dot-com bubble captured by firms through the transfer of skills and human

capital, but importantly, but it does not discuss how the dot-com crash affected labor outcomes for IT

workers. An interesting question for future research is how the collapse of this bubble affected the value

of skills for IT workers, and how this affected long-term labor outcomes for IT workers, especially in

conjunction with offshoring and other changes to IT labor supply that became increasingly important after

the bust.

References (Preliminary and Incomplete)

Ang, S., Slaughter, S., Ng, K. 2002. “Human Capital and Institutional Determinants of Information Technology Compensation: Modeling Multilevel and Cross-Level Interactions,” Management Science (48:11), pp. 1427-1445.

Attewell, P. 1992. “Technology Diffusion and Organizational Learning: The Case of Business Computing,”

Organization Science (3:1), pp. 1-19. Bapna, R., Langer, N., Mehra, A, Gopal, R. 2010. “Examining Return on Human Capital Investments in the Context

of Offshore IT Workers,” Working Paper. Bartel, A., Ichniowski, C., and Shaw, K. 2007. “How Does Information Technology Affect Productivity? Plant-

Level Comparisons of Product Innovation, Process Improvement, and Worker Skills,” Quarterly Journal of Economics (122:4), pp. 1721-1758.

Benkard, C. Learning and Forgetting: The Dynamics of Aircraft Production. American Economic Review, (90:4), 2000, 1034-1054.

Boh, W., Slaughter, S., Espinosa, J. 2007. “Learning from Experience in Software Development: A Multilevel

Analysis,” Management Science (53:8), pp. 1315-1331. Breschi, S. and Lissoni, F. 2001. “Knowledge spillovers and local innovation systems: a critical survey,” Industrial

and Corporate Change (10:4), pp. 975-1005. Bresnahan, T., Brynjolfsson, E. and Hitt, L. 2002. “Information Technology, Workplace Organization, and the

Demand for Skilled Labor : Firm-Level Evidence,” Quarterly Journal of Economics (117:1), pp. 339-376. Bloom, N., R. Sadun, and J. Van Reenen. 2008. “Americans Do I.T. Better: US Multinationals and the Productivity

Miracle,” Working Paper. Brynjolfsson, E. and Hitt, L. 1996. “Paradox Lost? Firm-Level Evidence on the Return to Information Systems

Spending,” Management Science (42:4), pp. 541-558. Brynjolfsson, E. and Hitt, L. 2003. “Computing Productivity: Firm Level Evidence,” The Review of Economics and

Statistics (85:4), pp.793-808. Brynjolfsson, E. and Hitt, L. 2000. “Beyond Computation: Information Technology, Organizational Transformation

and Business Performance,” Journal of Economic Perspectives (14:4), pp. 23-48. Catan, T. and Kendall, B. 2010. “US Tech Probe Nears End,” Wall Street Journal, Sept 17. Cook, J. 2002. Workers go from dot-com to down to earth. Seattle Pi. July 25th. Chang, Y. and V. Gurbaxani. The Impact of IT-Related Spillovers on Long-Run Productivity. Information Systems

Research, forthcoming. Cheng, Z. and B. Nault. 2007. Industry Level Supplier-Driven IT Spillovers. Management Science. (53:8) 1199-

1216. Cheng, Z. and B. Nault. Relative Industry Concentration and Customer-Driven IT Spillovers. Information Systems

Research. Forthcoming. David, P. “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox”,

American Economic Review 80(2), 1990, pp. 355-361. Dedrick, J., K. Kraemer, and V. Gurbaxani. 2003. “Information Technology and Economic Performance: A Critical

Review of the Evidence.” ACM Computing Surveys (35:1), pp. 1-28. Dewan, S. and Kraemer, K. 2000. “Information Technology and Productivity: Evidence from Country-Level Data,”

Management Science (46:4), pp. 548-562. Desmet, K. and Rossi-Hansberg, E. “Spatial Growth and Industry Age,” Journal of Economic Theory, Forthcoming. Fallick, B., Fleischman, C., Rebitzer,J. 2006. “Job-Hopping in Silicon Valley: Some Evidence Concerning the

Microfoundations of a High-Technology Cluster,” Review of Economics and Statistics (88:3), pp. 472-481. Farrell, D., M. Baily, J. Rames. 2005. US Productivity After the Dot Com Bust. Mckinsey Global Institute Report. Fichman, R., Kemerer, C. 1997. “The Assimilation of Software Process Innovations: An Organizational Learning

Perspective,” Management Science (43:10), pp. 1345-1363.

   

Forman, C., Goldfarb, A., Greenstein, S. 2005. “How did location affect adoption of the commercial Internet? Global village vs. urban leadership,” Journal of Urban Economics (58:3), pp. 389-420.

Forman, C., Goldfarb, A., Greenstein, S. “The Internet and Local Wages: Convergence or Divergence?,” American

Economic Review, forthcoming. Gartner Group. 2010. “IT Spending: How do you Stack Up?” http://www.gartner.com/research

/attributes/attr_47450_115.pdf Lichtenberg, F. 1995. “The Output Contributions of Computer Equipment and Personnel: A Firm-Level Analysis,”

Economics of Innovation and New Technology (3:3-4), pp. 201-217. Lieberman, M. 1984. “The learning curve and pricing in the chemical processing industries,” Rand Journal of

Economics (15:2), pp. 213-228. Mann, A. and Liou, T. 2010. “Crash and Reboot: Silicon Valley High-Tech Employment and Wages, 2000-2008,”

Monthly Labor Review, January, pp. 59-73. McElheran, K. 2011. Do Market Leaders Lead in Business Process Innovation? The Case(s) of E-Business

Adoption. Harvard Business School Technology & Operations Management Unit Working Paper No. 10-104.

Mithas, S., Krishnan, M. 2008. “Human Capital and Institutional Effects in the Compensation of Information

Technology Professionals in the United States,” Management Science (54:3), pp. 415-428. Oliner, S., Sichel, D. 2000. “The Resurgence of Growth in the late 1990’s: Is Information Technology the Story?”

Journal of Economic Perspectives (14:4), pp.3-22. Saunders, A. and P. Tambe. 2011. Shining Light on Dark Matter: Text-Mining the Digital Economy. Working

Paper. Saxenian, A. Regional Advantage: Culture and Competition in Silicon Valley and Route 128, Harvard University

Press, Cambridge, MA, 1996. Stiroh, K. 2002. “Information Technology and the US Productivity Revival: What Do the Industry Data Say?”

American Economic Review (92:5), pp. 1559-1576. Tambe, P. and Hitt, L. 2011a. “The Productivity of Information Technology Investments: New Evidence from IT

Labor Data”, Information Systems Research, Forthcoming. Tambe, P. and Hitt, L. 2011b. “Job Hopping, Information Technology Spillovers, and Productivity Growth”,

Working Paper. Thornton, R. and Thompson, P. 2001. “Learning from Experience and Learning from Others: An Exploration of

Learning and Spillovers in Wartime Shipbuilding,” American Economic Review (91:5), pp. 1350-1368. Vaas, L. 2001. Testing the New IT Job Pool. eWeek. May 21st. Vance, A. 2011. The Tech Bubble is Different. Bloomberg Businessweek.

http://www.businessweek.com/magazine/content/11_17/b4225060960537.htm Varian, H., Farrell, J., Shapiro, C. 2004. Economics of Information Technology: An Introduction. Cambridge

University Press.

Table 1: Top 10 large metropolitan areas for IT-Sector employment

Rank City State 1. San Jose CA 2. San Diego CA 3. Austin TX 4. Phoenix AZ 5. Los Angeles CA 6. Raleigh NC 7. Dallas TX 8. Denver CO 9. San Francisco CA 10. Seattle WA Cities are ranked by percentage of IT workforce employed in IT industries. Only includes large cities where large cities are those with at least 1000 IT workers in the data sample.

   

Figure 1: Job separations in IT-producing industries (Source: BLS Mass Layoffs Data)

Table 2: Key variable definitions

Variable Data Source Description Value Added Compustat Sales minus materials Capital Compustat Physical capital stock Non-IT Employment Compustat Number of non-IT employees IT Employment IT Labor Data Number of IT employees IT Produce Pool IT Labor Data IT Intensity of firms in IT-producing sectors, weighted by IT labor flows IT Using Pool IT Labor Data IT Intensity of firms in IT-using sectors, weighted by IT labor flows Post-Crash Dummy variable indicating whether the year is before or after the dot-com crash. E-commerce Skills IT Labor Data Self-reported IT skills and certifications Industry Compustat 2 digit SIC classification ITLM IT Labor Data % of IT labor in city employed in IT-producing firms E-commerce (binary) Adoption Data Extracted from 10-K text, from Saunders and Tambe (2011) ERP (binary) Adoption Data Extracted from 10-K text, from Saunders and Tambe (2011) Data mining (binary) Adoption Data Extracted from 10-K text, from Saunders and Tambe (2011)

050

000

1000

0015

0000

2000

00Se

para

tions

1995 2000 2005 2010Year

Computer Hardware Software and Computer Services

Communications Equipment Communications Services

All Industries

Figure 2: Destinations for IT Labor Displaced from IT Producing Industries

Figure 3: Source Industries for IT Labor Hired into FIRE Industries

0.2

.4.6

% IT

Wor

kers

Ent

erin

g Ea

ch In

dust

ry

1998 2000 2002 2004 2006Year

Retail ITManufacturing FIRE

.1.1

5.2

.25

.3.3

5%

Sou

rce

Indu

strie

s fo

r IT

Wor

kers

1998 2000 2002 2004 2006Year

Retail ITManufacturing FIRE

   

Figure 4: Lowest, Middle, and Highest ITLM Quintiles

Figure 5: Narrowed to e-commerce skills

0.0

2.0

4.0

6.0

8.1

% IT

Wor

kers

from

IT S

ecto

r

1996 1998 2000 2002 2004 2006Year

Lowest 20% LMO Middle 20% LMOHighest 20% LMO

0.2

.4.6

.8%

IT W

orke

rs fr

om IT

Indu

strie

s

1998 2000 2002 2004 2006Year

Other IT (High 20%) E-commerce (High 20%)Other IT (Low 20%) E-commerce (Low 20%)

Figure 7: Growth in Factors by Sector (Financial Firms, All IT)

.005

.01

.015

.02

.025

1996 1998 2000 2002 2004 2006year

IT-Using IT-Producing

-.20

.2.4

.6.8

Mea

n Re

mov

ed IT

Poo

l

1985 1990 1995 2000 2005Year

retail finance IT

Figure 6: External IT Investment Across Industries

   

Figure 8: Growth in Factors (Top 20% ITLM, E-commerce)

Figure 9: Growth in Factors (Bottom 20% ITLM, E-commerce)

 

0.0

01.0

02.0

03.0

04

1996 1998 2000 2002 2004 2006year

IT-Using IT-Producing

0.0

01.0

02.0

03

1996 1998 2000 2002 2004 2006year

IT-Using IT-Producing

Figure 10: Productivity Growth of IT Producing Firms from 1999-2003 by ITLM Quintile

Figure 11: Productivity Growth of IT Using Firms from 1999-2003 by ITLM Quintile

-.2-.1

5-.1

-.05

0

4 ye

ar g

row

th in

logg

ed p

rodu

ctiv

ity

1 2 3 4 5

-.02

0.0

2.0

4.0

6

4 ye

ar g

row

th in

logg

ed p

rodu

ctiv

ity

1 2 3 4 5

   

Figure 12: Productivity Growth of IT Using Firms 1994-1998 by ITLM Quintile

 

Table 3: Main productivity growth estimates

DV: Log(VA) (1) (2) (3) (4) (5) (6) (7) (8) FE FE 1 Yr Diffs 2 Yr Diff 3 Yr Diff 4 Yr Diff 4 Yr Diff 4 Yr Diff

IT Using Firms

IT Prod Firms

IT Using Firms

IT Using Firms

IT Using Firms

IT Using Firms

IT Prod Firms

IT Using Firms

Log(IT-Use) 0.00194 -0.00717 -0.000924 0.000973 0.00138 0.00407 -0.00410 0.00371

(0.00282) (0.00854) (0.00164) (0.00204) (0.00237) (0.00268) (0.00985) (0.00268)

Log(IT-Prod) -0.00184 0.00649 -0.000813 -0.00150 -0.00137 -0.000689 0.0134* -0.000611

(0.00289) (0.00690) (0.00173) (0.00222) (0.00274) (0.00273) (0.00744) (0.00273)

Log(IT-Use)⋅Post 0.00888** 0.0157 0.00140 0.00414 0.00557 0.00382 0.0140 0.00409

(0.00347) (0.0101) (0.00266) (0.00311) (0.00340) (0.00376) (0.0131) (0.00373)

Log(IT-Prod)⋅Post 0.0115*** -0.000823 0.00122 0.00475 0.00585 0.00861** -0.0130 0.00820**

(0.00349) (0.00813) (0.00246) (0.00304) (0.00376) (0.00387) (0.00982) (0.00384)

Log(IT Emp) 0.0296*** 0.0538*** 0.0350*** 0.0354*** 0.0313*** 0.0305*** 0.0549** 0.0322***

(0.00408) (0.0142) (0.00581) (0.00632) (0.00667) (0.00685) (0.0234) (0.00690)

Log(Capital) 0.173*** -0.00576 0.0996*** 0.114*** 0.112*** 0.121*** -0.00164 0.118***

(0.00626) (0.0179) (0.0229) (0.0199) (0.0211) (0.0217) (0.0318) (0.0210)

Log(Non IT Emp) 0.727*** 1.069*** 0.693*** 0.742*** 0.769*** 0.765*** 1.033*** 0.765***

(0.00780) (0.0243) (0.0338) (0.0294) (0.0303) (0.0309) (0.0439) (0.0301)

Observations 29,324 6,972 24,951 21,653 18,945 16,649 3,081 16,649 R2 0.688 0.571 0.304 0.453 0.536 0.595 0.615 0.605 Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1; Estimates are from a Cobb-Douglas production function using value added as the dependent variable. Regressions (1)-(7) include controls for year and one-digit industry. Column (8) uses 2 digit industry, year, and state controls.

0.0

2.0

4.0

6.0

8.1

4 ye

ar g

row

th in

logg

ed p

rodu

ctiv

ity

1 2 3 4 5

Table 4: Fixed-Effects Estimates by IT Labor Market and IT Investment

DV: Log(VA) (1) (2) (3) (4) (5) (6) (7) (8)

IT-Using High ITLM

IT-Using Low ITLM

IT-Prod High ITLM

IT-Prod Low ITLM

Below IT mean

Above IT Mean

Drop CA

Only CA

Log(IT-Use) 0.00592 0.00287 -0.00258 -0.00781 .00465* .00699 .00469* -.00227

(0.00422) (0.00340) (0.0123) (0.0161) (.00279) (.00461) (.00283) (.0125)

Log(IT-Prod) -0.000285 -0.000487 0.0155* 0.0111 .00193 .00259 .00131 -.0152

(0.00429) (0.00353) (0.00911) (0.0129) (.00312) (.00372) (.00245) (.0118)

Log(IT-Use)⋅Post 0.00474 0.00274 0.0148 0.0112 .00295 .0131** .00708* .0220

(0.00578) (0.00493) (0.0156) (0.0235) (.00406) (.00582) (.00376) (.0168)

Log(IT-Prod)⋅Post 0.0128** 0.00324 -0.0178 -0.00373 .00542 .0122** .00747** .0419***

(0.00555) (0.00541) (0.0122) (0.0167) (.00470) (.00505) (.00362) (.0150)

Log(IT Emp) 0.0444*** 0.00955 0.0468* 0.0726* .0370*** .0208*** .0247*** .0287

(0.00916) (0.0102) (0.0280) (0.0420) (.00765) (.00796) (.00622) (.0232)

Log(Capital) 0.0753** 0.173*** -0.0158 0.0282 .277*** .235*** .215*** .301***

(0.0299) (0.0289) (0.0355) (0.0637) (.0236) (.0136) (.0142) (.0658)

Log(Non IT Emp) 0.827*** 0.699*** 1.054*** 0.989*** .648*** .715*** .721*** .656***

(0.0406) (0.0429) (0.0533) (0.0771) (.0313) (.0178) (.0186) (.0846)

Observations 8,891 7,758 2,081 1,000 11,840 17,484 26,173 3,151 R-squared 0.603 0.591 0.603 0.648 2,128 3,570 3,129 441 Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1; Estimates are from a Cobb-Douglas production function using value added as the dependent variable. Each regression includes controls for year and one-digit industry.

 

Table 5: Fixed Effects Estimates by Industry

DV: Log(Value Added) (1) (2) (3) (4) (5) (6) (7)

Industry Heavy Manuf

Light Manuf Transport Wholesale Retail Finance Business

Services

Industry SIC range 24,25 32-39

20-23 26-31 41-49 50,51 52-59 60-67 70-87

Log(IT-Use) .00866 .00163 .00260 .0167 .0130** -.00164 -.0124

(.00530) (.00588) (.00581) (.0132) (.00636) (.0124) (.00851)

Log(IT-Prod) -.00628 -.00398 -.00816 -.00102 .00972 .00651 .0107

(.00527) (.00635) (.00607) (.0125) (.00684) (.0120) (.00841)

Log(IT-Use)⋅Post .00448 .0177** -.00774 -.0464*** .00182 .0253* .0303***

(.00657) (.00738) (.00716) (.0161) (.00804) (.0148) (.0103)

Log(IT-Prod)⋅Post .0164*** .0141* .00658 .00530 .0177** -.00325 -.000722

(.00633) (.00780) (.00740) (.0157) (.00827) (.0145) (.00992)

Log(IT Employment) .0484*** .0499*** .0118 .0122 .0151 -.00839 .0227*

(.00708) (.00914) (.00766) (.0203) (.00931) (.0197) (.0130)

Log(Capital) .0411*** .291*** .389*** .00277 .0112 .123*** .202***

(.0133) (.0183) (.0137) (.0253) (.0165) (.0218) (.0147)

Log(Non IT Employment) .921*** .662*** .548*** .977*** .958*** .563*** .698***

(.0156) (.0200) (.0157) (.0382) (.0212) (.0280) (.0206)

Observations 8,312 6,382 3,458 1,506 3,669 2,345 3,615 R-squared 0.692 0.635 0.808 0.685 0.817 0.640 0.732 Number of panelid 982 703 358 155 446 357 539 Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Estimates are from a Cobb-Douglas production function using value added as the dependent variable. Each regression includes controls.

   

Table 6: Instrumental Variable Tests

DV: Log(VA) (1) (2) (3) (4)

Instruments ITLM IT Layoffs

ITLM IT Layoffs

1 dig Ind*ITLM

ITLM IT Layoffs

2 dig Ind*ITLM

ITLM IT Layoffs

1 dig Ind*ITLM IT-Using IT-Using IT-Using IT-Producing 2001-2006 2001-2006 2001-2006 1995-2000 Log(IT-Use) 0.114* 0.120*** 0.0264 -0.0144 (0.0628) (0.0395) (0.0165) (0.0463) Log(IT-Prod) 0.0116 0.0138** 0.0106* 0.00180 (0.00786) (0.00690) (0.00629) (0.00498) Log(IT Emp) -0.0348 -0.0306 0.00139 0.0385 (0.0305) (0.0254) (0.0190) (0.0238) Log(Capital) 0.0986* 0.110*** 0.119*** 0.157*** (0.0504) (0.0425) (0.0404) (0.0507) Log(Non IT Emp) 0.821*** 0.813*** 0.787*** 0.752*** (0.0618) (0.0524) (0.0501) (0.0728) R2 0.247 0.450 0.0264 .718 First-stage R2 .044 .053 .136 .101 Anderson LR .085 .294 .002 .113 Hansen’s J .66 .81 .12 .28 N 2,355 2,355 2,355 1,991 All regressions are 4 year differences regressions and include controls for 1-digit industry and year. Standard errors are clustered on firm.

 

Table 7: Matching Estimator Output

(1) (2) (3) (4) (5)

DV: Productivity Growth State State 1 digit

State 2 digit

State 4 digit

State/County 1 digit

2001-2005 2001-2005 2001-2005 2001-2005 2001-2005

Top quintile ITLM 293.4** 270.5** 209.7** 228.2** 246.9**

(117.4) (118.1) (105.4) (107.0) (105.7)

Exact matches 92.6% 65.6% 30.8% 9.4% 15.9% Observations 479 479 479 479 479 1994-1999 1994-1999 1994-1999 1994-1999 1994-1999 Top quintile ITLM -34.0 -168.1* -145.0 -143.7 -44.0 (90.3) (96.9) (97.2) (97.2) (86.9) Exact matches 93.2% 62.6% 22.0% 5.4% 18.5% Observations 368 368 368 368 368 Robust standard errors in parentheses; *** p<0.01, ** p<0.05, * p<0.1. Estimator matches firms in top ITLM quintile to closest match in bottom ITLM quintile. Matching variables include total employment, location, and industry (level of location and industry variables described in column headers).

Table 8: E-commerce adoption estimates

(1) (2) (3) (4) (5)

DV: Technology Adoptiona e-commerce e-commerce data mining erp e-commerce IV

Location All Top ITLM Top ITLM Top ITLM All Log(IT-Use) 0.0536 0.0592 0.196*** 0.0306 -.088

(0.0332) (0.0603) (0.0526) (0.0529) (.054)

Log(IT-Prod) -0.0109 -0.0928* 0.0349 0.00131 .452***

(0.0283) (0.0560) (0.0497) (0.0382) (.148)

Log(IT-Use)⋅Post -0.00808 0.00922 -0.0418 -0.00173

(0.0328) (0.0469) (0.0573) (0.0533)

Log(IT-Prod)⋅Post 0.0586* 0.164*** 0.0683 0.0565

(0.0303) (0.0584) (0.0488) (0.0421)

Observations 8,346 2,597 2,597 2,597 4,170 aTechnology adoption dependent variables are based on text-mining 10-K reports and described in Saunders and Tambe (2011). Robust standard errors in parentheses; *** p<0.01, ** p<0.05, * p<0.1. IV Regression in (5) only uses observations after 2000.

Table 9: Regressions Using E-Commerce Skill Measures

DV: Log(VA) (1) (2) (3) (4)

OLS OLS FE FE

Log(IT-Prod)

0.0359***

-0.00941

(0.0113)

(0.00678)

Log(IT-Prod)⋅Post

-0.0241*

0.0139*

(0.0134)

(0.00773)

Log(IT-Use)

0.0357***

0.00965

(0.0119)

(0.00694)

Log(IT-Use)⋅Post

0.000786

-0.00858

(0.0134)

(0.00782)

Log(IT-Prod E-Commerce) -0.00934 -0.0308 -0.0297 -0.0274

(0.0260) (0.0274) (0.0292) (0.0294)

Log(IT-Prod E-Commerce) ⋅Post 0.0519* 0.0664** 0.0367 0.0330

(0.0278) (0.0297) (0.0303) (0.0305)

Log(IT-Use E-Commerce) 0.0412* 0.0150 0.0257 0.0232

(0.0235) (0.0243) (0.0295) (0.0297)

Log(IT-Use E-Commerce) ⋅Post -0.00611 0.00610 -0.0234 -0.0217

(0.0262) (0.0270) (0.0304) (0.0308)

Observations 5,577 5,577 5,577 5,577 R-squared 0.928 0.929 0.680 0.680 Number of firms 1,726 1,726 Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Estimates are from a Cobb-Douglas production function using value added as the dependent variable. Each regression includes controls for year and one-digit industry. All regressions also include IT employment, capital, and non IT-employment.