†Comments are appreciated. Address for correspondence: Stern School of Business, New York University, KMC 8-82, 44 West Fourth Street, New York NY 10012-1126. E-mail: ptambe@stern.nyu.edu. We have greatly benefited from feedback from Rocio Bonet, Benjamin Campbell, April Franco, Brad Greenwood, Damien Joseph, Paul Osterman, Michael Roach, and Aaron Sojourner as well as seminar participants at the Academy of Management, the American Sociological Association, the Wharton People and Organizations Conference, the University of Minnesota, Temple University, the INFORMS Conference on Information Systems and Technology, Purdue University, the Workshop for Information Systems and Economics, and New York University. Prasanna Tambe is very grateful to the Alfred P. Sloan foundation for financial assistance.

POACHING AND RETENTION IN HIGH-TECH LABOR MARKETS

PRASANNA TAMBE† NEW YORK UNIVERSITY

XUAN YE NEW YORK UNIVERSITY

PETER CAPPELLI U. OF PENNSYLVANIA

DRAFT – PLEASE DO NOT POST OR QUOTE

Last revised, April 11th, 2015

Abstract Using new, fine-grained data generated through job search activity, we show that greater wage premiums are required to poach IT workers from employers investing in emerging technologies. These results are specific to IT workers, and specific to their employers’ investments in emerging information technologies. We do not observe similar results for workers in other occupations, or for workers from employers distinguished along other characteristics, such as performance, aggregate IT investment, growth, R&D investment, or size. We then provide evidence that these premiums arise because the long-run value of the skills acquired by working with new technologies enters into the worker’s underlying utility function. In thick labor markets, employers’ investments in emerging technologies and in workplaces that emphasize learning and skill acquisition appear to be an important source of non-wage compensation for IT workers. These findings indicate how labor market structure can impact the firm’s ability to finance new IT innovation. Moreover, they suggest that workers with long, predictable career horizons over which to earn returns to technical skills, such as young males, have greater financial incentives to seek employment in markets characterized by early technology adoption.

I. INTRODUCTION

As software and business process digitization become increasingly important for explaining

corporate performance, access to IT workers has become a critical differentiation point for firms. Labor

markets for workers are receiving growing attention from managers, the press, and policy makers because

of topics such as whether there are IT labor “shortages” that should be addressed through immigration, as

well as a noted lack of gender and race diversity in the IT workforce. The debate surrounding these topics

has become both public and divisive (Glantz 2012). Competition for high-tech workers has also received

scrutiny as employers have used aggressive tactics to manage employee hiring and retention—including

allegedly collusive measures—drawing attention from the US Department of Justice (Streitfeld 2014).

In this paper, we study poaching and retention in high-tech labor markets, topics of significant

interest to both policy makers and managers. For managers, the acquisition and retention of IT workers

has become an increasingly important issue because access to IT labor has implications for IT strategy

and performance, and because losing IT workers to competitors is costly.

To provide empirical insight into poaching in high-tech labor markets, we analyze a new and

unique data source on the “target wages” that IT workers seek from future employers. These figures are

similar – although not identical to -- reservation wages, which represent the wage level at which an

unemployed individual will accept a job or, as in this case, the wage at which an employed worker will

change jobs. Reservation wages have been a key parameter in job search models where they indicate the

wage that a worker expects in their next job (Lancaster and Chesher 1983; Feldstein and Poterba, 1984;

Shimer and Werning 2007). Employed workers choose a reservation wage that maximizes their future

income, striking a balance between higher wages and the ability to secure job offers. For employed

workers, the target wage is likely to be closely related to the reservation wage, although it may

systematically differ in ways that we discuss later in the paper.

These data are valuable because information on the wages required to leave a firm are not

otherwise easily available. Despite its importance for labor policy questions, empirical studies using data

on reservation wages are rare because these values are difficult to measure. They are perceptual

measures, and are therefore stored only inside the heads of workers seeking new jobs. Similar data are

occasionally collected by survey from small samples of workers, but even survey-based studies of

reservation wages are subject to a series of biases, such as overconfidence, self-esteem biases (e.g.,

pressure to report higher values), and so forth. To our knowledge, data on the wages required to leave

employers are not routinely collected by any administrative agencies.

The target wage data used in this study are from a leading online jobs board on which U.S IT

workers post wages desired from future employers. They are, therefore, generated by actual job search

activity. These data are noteworthy for two reasons. First, in contrast to previous studies based on small

samples of workers, we have access to the target wages of over 200,000 information technology workers.

Second, because the data are reported in a job search context where the stated target wages have real

consequences – potential employers see them – we argue that workers are likely to report these values

accurately. In particular, reporting target wages that are too high can have financial consequences for

workers because employers may screen on this number when considering applicants. The data also

include detailed information about workers’ locations, human capital, and current and past employment

information. Individuals have an incentive to report this information accurately as well because the failure

to do so typically constitutes grounds for dismissal from any employer who would hire an applicant based

even in part on that information.

We show that conditional on the worker’s current wage, IT workers post higher target wages

when their current employers are investing in cutting-edge technologies. In other words, these workers

require higher wage increases to be pulled away from these employers. These results are unique to IT

workers, and are unique to the early technological investments of their employers. We do not observe

similar results for workers in other occupations, or for workers employed at firms distinguished along any

other characteristics, such as performance, aggregate IT investment, firm size, or R&D investment. These

findings are consistent with the perceived difficulties in hiring IT workers in high-tech labor markets.

We then explore the mechanism that drives workers to post higher target wages when employed

at these firms. To do this, we complement our wage data with a second novel data source, online reviews

posted at Glassdoor, which provide a rare and fine-grained window into employees’ perceptions of the

various compensation and benefits used by their employers. We text-mine hundreds of thousands of

reviews to develop firm-level measures of non-wage benefits and compensation, as perceived by the

firm’s past and current employees. We connect this data source with our wage data to demonstrate that

the target wage premium associated with the firms’ IT investment arises because the skills workers

acquire by working with these technologies enters into their underlying utility function. In other words,

outside offers must account for the value that IT employees assign to working with various information

technologies, which is systematically higher in high-tech labor markets.

These findings have many important implications. First, they indicate how technological

investment and skill acquisition affect the poaching and retention of key employees. Specifically, they

provide confirmatory evidence that poaching premiums are higher in high-tech labor markets, and they

shed light on why workers set higher target wages in these markets. Our findings suggest that

technological investment, in workplaces that facilitate skills and learning, provide an important non-wage

benefit to workers. There are also two policy implications. First, our findings suggest important links

between technological investment and labor market structure. Because the workers’ ability to recoup the

value of these skills depends on labor market structure, our findings suggest that workers will place

greater value on these non-wage benefits in thick markets where labor mobility is relatively unrestricted

by non-compete policies or other provisions. This, in turn, implies that firms are best able to finance

these new IT innovations in these types of labor markets. Second, the tradeoffs described above are a

source of inter-temporal earnings risk that can reduce incentives for older workers with fewer years

remaining in the workforce to sort into firms because they may miss out on the ability to recoup benefits

in the future.

Our paper principally contributes to two academic literatures. First, it connects labor market

structure to the financing of new IT innovations. Our understanding of how technological investment is

distributed across firms and regions and how it affects the strategic choices of employers is in its infancy.

The importance of research in this area is highlighted by an influential literature on how the economic

properties of science affects the career choices of scientists (Moen 2002; Stern 2004), and how these

career choices affect patenting, entrepreneurship, and other forms of innovation that matter for economic

growth. Like scientist mobility, IT labor market structure is of strategic interest because of the importance

of technical human capital stock for economic growth, and because the mobility of IT-related human

capital can produce spillovers that have implications for agglomeration and that can discourage corporate

investment in skill development. Separating the incentives for firms and workers to invest in these skills

is important, therefore, for understanding the strategic dynamics of IT investment.

Second, determinants of IT compensation and IT career paths have been the subject of a number

of prior studies (Ang et al 2002; Levina and Xin 2007; Mithas and Krishnan 2008; Mithas and Lucas

2010), and the nature of IT careers and skill acquisition has been examined in a number of recent papers

(Joseph et al 2012; Bapna et al 2013). The literature in these areas focuses on how supply and demand

factors including human capital, labor market, training, and employer characteristics (primarily size and

industry) affect IT compensation and career paths. Our paper contributes to this literature by examining

how technological investment—among the most salient features of IT labor markets—affects the mobility

of workers across firms. To the best of our knowledge, our paper is among the first to consider poaching

in these labor markets, which is currently a topic of significant interest for both policy maker and

managers, and it is the first to consider the role of non-wage benefits, which appear to play an important

role in high-tech labor markets.

As with all empirical analyses, ours has several caveats. Data collected from jobs boards are

subject to sampling limitations because web site participants are actively seeking new jobs. While posting

one’s resume to a job board does not necessarily represent aggressive job search – roughly 18 million

people in the US did so on this job board alone – our results can offer little insight about workers who do

not post their resumes on a job board like this one or those who are interested in other jobs but prefer to

search in other ways. However, given that our study is about poaching, much of the underlying

population that we are interested in is probably well represented by workers who post on job boards.

Moreover, most data used in prior work on IT wages suffers from similar or even more severe sample

limitations. We report detailed comparisons with administratively sampled data sets (such as the Current

Population Survey from Census and Occupational Employment Statistics from the Bureau of Labor

Statistics) to document how the sample composition and the reported wages compare to that of the

broader IT labor force.

The principal identification concern that arises in our analysis is that other firm-level factors,

correlated with technology investment and skill acquisition, can explain why workers at these firms raise

their target wages. Some of the primary candidates are greater levels of non-wage benefits that are not

directly measured by our Glassdoor measures. However, our analyses using the Glassdoor data as well as

other supplementary tests provide fairly robust evidence of our key hypothesis. Our key coefficient

estimates are significant when we interact learning and technology investment, and are robust to including

measures of other wage and non-wage benefits, including measures of equity compensation, office perks,

healthcare and retirement benefits, work-life balance, as well as when including measures of the overall

rating of the firm from the perspective of its employees. The class of omitted variables that could bias our

results must be correlated with investment in both new technologies and skills and learning, but not with

any other benefits or with employees’ overall perception of the quality of the firm or its compensation or

benefits.

We also provide evidence for the robustness of our interpretation by leveraging the detailed

career histories for the workers in our sample, along with the observation that whether seeking a higher

target wage is driven by skills or greater non-wage benefits has opposing predictions for how workers of

different types set their target wage. A skills-based explanation predicts higher estimates on

technological investments for workers who require training (and are less productive on entry), whereas an

explanation based on non-wage benefits predicts either neutral effects or stronger effects for workers with

the most valuable skills if these workers have greater bargaining power. We observe higher target wages

associated with technical workers who earn lower wages, are less educated, and require more reskilling as

indicated by prior employers and job titles. We observe a similar pattern with workers with skills in data

technologies, which have been the most disrupted by recent technological change. This is consistent with

a skills based explanation. Moreover, the alternative, that workers who require the most reskilling are

rewarded with the greatest non-wage benefits, is less plausible.

II. BACKGROUND

Many of the concerns pertaining to IT labor market activity stem from how rapid shifts in the

demand for technical skills affect the IT labor force. The canonical model of human capital formation

suggests that employees invest in human capital in exchange for future income, but this conclusion does

not hold in contexts like IT where skills are general – valuable to employers elsewhere – and employees

are mobile. It may be further complicated for IT workers because technical skills depreciate quickly and

because many skills are acquired on-the-job (see Ang, Slaughter, and Ng, 2002 for evidence). How such

skills are financed is the subject of a long literature in labor economics starting with Becker (1964) that

recently has focused on the more limited circumstances under which employers can provide general

skills that are useful elsewhere (e.g., Acemoglu and Pischke 1998), typically by having employees share

the costs of such training.

IT skills are perhaps especially transferrable across employers, and when combined with the fact that they

can depreciate quickly, the ability of employers to pay for workers to learn IT skills is constrained.

Consequently, rapid technical change can affect both IT careers and wage structure in ways that have not

been studied in prior work.

Our findings also have implications for an extensive literature on IT turnover (e.g. see Igbaria and

Siegel 1992; Moore 2006; Joseph et al 2007). Although IT turnover has been recognized as a significant

problem for managers (Agarwal and Ferratt 2000), the literature on factors affecting IT turnover has

principally focused on behavioral explanations. The economics literature has investigated some of the

consequences of IT mobility (Fallick et al 2006, Tambe and Hitt 2014), but has not investigated the

economic micro-foundations of these rapid labor movements. Our study offers an economic explanation

for higher IT turnover rates. By doing so, it connects IT turnover to a larger economic literature on

technical change and generates new empirical predictions about when and where managers can expect

higher rates of IT poaching and turnover.

Wages and careers in high-tech labor markets have been the subject of an active literature (e.g.

Ang et al 2002; Levina and Xin 2007; Mithas and Krishnan 2008; Mithas and Lucas 2010; Joseph et al

2012). Existing work analyzes data from sources such as the Current Population Survey (CPS) or the

National Longitudinal Survey of Youth (NLSY) administered by the US government (e.g. Levina and Xin

2007 or Joseph et al 2012), and private survey data such as that collected by InformationWeek, a media

publication targeted towards the IT industry (e.g. Mithas and Krishnan 2008). These studies have

examined IT career paths (e.g. transitions into and out of IT occupations) as well as human capital and

labor market factors that affect IT compensation. More recent work in this area has emphasized the

changing importance of specific skills—for example, a recent literature notes a shift in demand within the

IT workforce towards softer, practical skills such as teamwork or project management (Joseph et al 2010;

Tambe and Hitt 2012).

One of the largest gaps in the empirical literature on IT labor markets relates to how IT

innovation and the ensuing depreciation of technical skills impact IT careers. Existing work on IT labor

markets has principally focused on the returns to firm-specific and general human capital endowments for

IT workers, which are relatively unaffected by technological change (Ang et al 2002; Slaughter et al

2007). Researchers have provided evidence that much of the human capital complementary to specific

technologies is acquired on-the-job (Boh et al 2007), and recent papers study the economic consequences

of the dissemination of new technical human capital for firm performance (Tambe and Hitt 2014; Tambe

2014). How this IT human capital replacement process is paid for has not yet been studied.

More generally, the management literature argues that inter-organizational careers are of growing

importance, and that the ability to acquire new skills is among the most salient determinants for inter-

organizational career success (Barley and Kunda 2006; O’Mahony and Bechky 2006; Bidwell and

Briscoe 2010). Inter-organizational careers and related skill acquisition may be particularly important for

IT workers because much technical human capital is transferrable across firms (Kim et al 2014). Prior

empirical work has shown that IT workers’ can engage in non-compensated production for signaling

purposes (Roberts et al 2006; Hann et al 2014), but the literature is only now addressing how IT workers

acquire new skills, and how this impacts careers or mobility (Bapna et al 2013 study training programs).

This gap in the literature is especially notable because anecdotal evidence suggests that due to the rapid

pace of IT innovation, continuous learning for IT workers—which is not well-captured by general or

firm-specific human capital endowments—is particularly important for explaining labor outcomes.

The absence of a literature in this area is underscored by a robust and influential literature in

management and economics on how technological change affects human capital investment and careers.

Researchers have demonstrated that preferences for science and learning can explain aspects of labor

market sorting, careers, and wages for scientists. Moen uses panel data on R&D wages to illustrate wage

profiles for R&D workers that are consistent with the argument that scientists accept lower wages early in

their careers to work with new R&D knowledge that can benefit them later in their careers (2002). Stern

uses survey data on job offers to provide evidence that scientists who prefer working with science and

new knowledge accept compensating differentials (2004).

Scholars have also demonstrated that these preferences have important implications for the

sorting of R&D workers into industry and academics (Roach and Sauermann 2010; Agarwal and Ohyama

2013; Roach and Sauermann 2014). These papers are embedded in a larger literature about on-the-job

human capital accumulation and productivity slowdowns in the presence of technological change (e.g.

Chari and Hopenhayn 1991; Jovanavic and Nyarko 1996). A related empirical literature has shown that

workers often trade off lower wages for greater external opportunity in later periods, either by working

with a newer generation of technologies or by transferring know-how to a new commercial context.

Franco and Filson analyze mobility data from the rigid disk drive industry and show that when employees

copy the know-how of prior employers, they pay for the possibility of this human capital accumulation

(2006). In the mutual fund context, researchers have demonstrated that workers divert effort from

productive activities on-the-job to learn how well suited they are for entrepreneurship (Chatterji et al

2013).

In summary, this literature suggests that the opportunities workers face for working with and

learning from new science at prospective employers has implications labor market sorting and wages. We

make a related argument here, focusing on new technical skills, rather than new scientific knowledge,

and specifically on how the acquisition of those skills is paid for. We know of , no prior work related to

scientific or IT-enabled know-how examining this question. The paper closest to ours in the scientific

world may be Stern (2004), who collects data on multiple simultaneous job offers for graduating R&D

scientists and uses differences in wages to estimate the compensating differential associated with working

with new science.

III. DATA SOURCES AND KEY MEASURES

III.A. Target Wages from Online Job Boards

To test our arguments, we use data on the target wages of IT workers. These are similar to reservation

wages, which are the wage level at which workers are indifferent between accepting and rejecting job

offers. Reservation wages have important implications for understanding labor market outcomes related

to job search, such as labor mobility. The empirical literature on reservation wages has focused on the

effects of reservation wage setting on labor outcomes such as unemployment duration, or on factors that

affect how workers set their reservation wages, such as unemployment insurance levels. The literature in

the latter area typically uses human capital variables such as experience, education, demographic data,

and especially employment status to explain reservation wage levels. Although reservation wages are of

interest for both employed and unemployed workers, the population of interest in most existing studies

has been unemployed workers due to the policy interests around unemployment (e.g. Prasad, 2003). Most

empirical work in this area has used surveys asking workers to report their reservation wages (Lancaster

and Chesher 1983; Feldstein and Poterba 1984; Jones 1988; Falk et al 2006; Hall and Krueger 2012;

Krueger 2014). These studies have varied in the definitions used. For example, the surveys used by

these authors have asked survey respondents for their “desired wage” or for their “asking wage”.

The online job board from which we collect data asks workers to report their “target wage”, but

the context in which the question is asked makes it clear that prospective employers will see that wage

and interpret it as a statement about what the individual is expecting in a job elsewhere. The posted data

on target wages are also notable because most existing data on reservation wages have been collected

from small worker samples, so data on target wages for such a large number of workers presents several

opportunities. Employed workers also post the wages they are earning at their current employer. The

difference between their stated target wage and their current wage is important because the size of the gap

tells us how much would be required for them to leave the firm.

It is certainly true that the measures of target wages here and of current wages are far from perfect.

Self-reported data on wages has many potential biases, although virtually all labor market studies are

based on self-reported data. Individuals may have much stronger incentives to inflate their current wages

here as doing so may affect wage offers from employers. But there are also powerful sanctions against

doing so in this context. The first is that lying on job applications is typically grounds for dismissal from

a job, and the second is that inflating the report of their current wages may also reduce the interest in them

from other employers by making them appear too expensive. Individuals may report higher target wages

than they actually believe are reasonable in hopes that it will lead to higher wage offers. The downside,

of course, is that it will reduce interest from many employers. Others may report lower target wages in

hopes of attracting more offers, perhaps negotiating them up along the way. The obvious downside here

is that it is much more difficult to raise wage offers when starting out with a statement that you are

interested in a lower wage.

These target wages are available for all workers who have participated on the web site. However,

we focus only on employed IT workers, which limits the sample to slightly more than 250,000 workers.

However, the sample size we use in our analysis is much smaller, due to restrictions with the availability

of supplementary data from Compustat and other sources. An analysis of how the workers used in our

regression compares to the broader underlying sample of IT workers is presented in Appendix A.

We identify IT workers as those who choose information technology as one of their three

occupational affiliations from a drop-down menu. We also know when the data were submitted, although

we do not have panel data because we do not observe the same workers over multiple years. Participants

also submit demographic, human capital, and past employment information (i.e. prior jobs and

employers). We also have information about what types of the job the worker is seeking (annual or paid

hourly, i.e., exempt from the Fair Labor Standards Act requirements or covered by them). In this study,

most of our analyses are on annual workers, although we use hourly workers for some robustness tests.

III.B. Selection Concerns and Data Benchmarking

The principal concern with data collected from online job boards is that participants are actively engaged

in job search, and therefore, are representative of a particular sub-population of the IT workforce.

Broadly, when evaluating the sample, it is important to consider that the jobs board sample are more

likely to be job seekers—and possibly younger, higher mobility, and less likely to be embedded in firms

with deep internal labor markets. On the other hand, job seekers within the IT workforce, who are

particularly well represented on the jobs board, are the principal group of interest for many of the issues

considered in this study. To be more precise about how our sample compares to the IT workforce-at-large,

we can conduct direct comparisons between the workers in this sample and the workers in other datasets

with known sampling properties.

To assess how this sample compares to the IT workforce, we conduct two sets of comparisons,

the first with workers in the Current Population Survey (CPS) and the second with the Occupational

Employment Survey (OES). For both comparisons, some flexibility was required to create comparison

samples because the occupational classifications done by government agencies do not fit neatly with the

IT classification in the jobs board sample. To bridge this gap, we compared IT workers on the job boards

sample with workers in the government statistics who fell into “Computer and Mathematical Occupations”

(O-Net code 15). Workers in these occupations comprise the majority of the workers commonly

considered to be within the IT workforce, and especially those in technical occupations, such as

programmers, network administrators, and database administrators.

In Table 1, we compare demographic and human capital statistics from annually paid workers in

the job boards sample with IT workers in the Current Population Survey (CPS), which is administered by

the US Census Bureau, and is designed to be statistically representative of the US workforce. The

educational and industry distributions in the two samples are similar. Education levels in the CPS sample

are slightly higher than in the job boards sample, but the differences are not large. At the industry level,

there are more IT workers from finance and real estate industries in the job board sample but again the

differences do not appear to be very large.

There are some differences along demographic lines. Relative to the CPS, the job board sample

has proportionately fewer white males among IT workers. However, this is somewhat difficult to pin

down because race and gender are not required categories for job board workers, and the similarity in

distributions is dependent on the composition of those who choose not to provide this information, which

we do not observe. Table 2 compares the wages of workers in the job board to the wages of workers

collected through the Bureau of Labor Statistics Occupational Employment Statistics (OES) program.

Like the CPS, the data from the OES are sampled in a way to be statistically representative of the US

workforce. We compared wages along three dimensions that we use for analysis later in the paper:

industry, geography, and occupation.

The comparison reveals some interesting differences. When comparing key labor markets, the

wages of IT workers in the jobs board sample are systematically higher than the wages of IT workers

collected by the OES. This may, in part, be due to the occupational classification problems discussed

earlier. On the job boards sample, workers employed in IT industries who are not technical workers (e.g.

senior sales manager at a high-tech firm) may identify with the “Information Technology” profession and

we do not attempt to separate these workers out of our sample. If these workers have higher wages than

most technical workers in these regions, this could produce differences in the effects that we observe.

Because IT industries tend to be geographically concentrated in high-tech regions, this bias is likely to be

highest in the high-tech metropolitan areas for which the comparisons are presented.

Across industry, wages from the jobs board sample are lower in some industries, such as software

publishing, but they are higher in others, such as Internet Service Providers. The occupation-based wage

differences are more consistent—the wages of workers on the jobs board sample are lower than workers

in the OES. Because this distribution controls for occupational heterogeneity, the differences between

these and the regional distributions are not surprising.

Overall, these comparisons suggest that the wages of the IT workers in our sample differ in some

predictable ways from the broader IT workforce, but the variation in IT wages across industries and

regions is broadly similar to that reported in administrative data. The analyses below should be taken as

evidence for IT workers who are seeking alternative employment, specifically via job boards. That

population appears to be considerable. There are many job boards in addition to the one we access here,

and while we have no way of knowing how many candidates use multiple job boards, there are millions

of individuals whose information is listed on job boards.

III.C. Firm-Level Data from Glassdoor

From Glassdoor.com, we use online reviews to measure a number of firm-level attributes such as work

environment, perks and benefits. Glassdoor is an online labor market intermediary launched in 2008,

where employees anonymously review companies and their management. There are over 1.4 million user

generated ratings and reviews (700,000 for US firms) that comment on the “pros” and “cons” of

employment at these companies. The mechanism that Glassdoor uses to incentivize providing review

information is that employees who want to read past website comments must contribute at least one piece

of information to the sites. "Reviews" are one of the featured sections on Glassdoor.com, meant to reveal

information about how the work practices of each company are perceived by employees.

There are six rating fields in the "Reviews" section that employees are required to fill in,

including overall ratings, work life balance, compensation and benefits, culture, senior management, and

career opportunities. In addition to these fields, website participants are encouraged to enter text

information, including “pros”, i.e. company features that the employee likes, the “cons”, i.e. the aspects

that the employee dislikes. For our analysis, we focus on the text in the “pros”. The raw data that we use

include the text in the “pros” section of all 700,000 U.S. reviews from Glassdoor. To convert these

reviews to firm attributes, we 1) define the constructs 2) convert the review data into text features and 3)

compute firm measures for each construct. We discuss each step in turn in the following paragraphs.

There are no widely accepted taxonomies characterizing non-wage benefits. We therefore use

data centric methods to develop our measures of non-wage benefits (see Archak et al 2011 for a similar

argument in the product space). We pre-process all of the reviews by removing stop words, and then

transform each review into bigrams. We use bigrams as the minimum level of analysis because unigram

combinations may have different semantic meanings.1 We compute the term frequency of the bigram in

the whole text corpus. In total there are 2.8 million unique bigram terms, of which 2 million of them only

appear once. The top 1% of bigrams by frequency have a term frequency of 40 across the entire corpus.

We then manually review the top 2000 most frequent bigram terms, and place them into one of

the following non-wage benefit categories: 1) Work environment and quality of co workers (CULTURE);

2) Opportunities for career advancement and development (ADVANCE); 3) Benefits, perks and

compensation (BENEFITS); 4) Work-life balance (BALANCE) and 5) Opportunities for skills, learning

and training (SKILLS). We choose the first four categories of non-wage benefits because they correspond

to the top level categories used by Glassdoor. The final measures is the central theoretical construct of our

paper. Therefore, the final representation of each firm is a five dimensional vector.

We then convert each review into measures. Because we use the “pros” section as the raw data

source, we count the reviews that mention any of the keywords related to one of the five dimensions

mentioned above. We do not differentiate adjectives (e.g. “good”, “great”) because of the heterogeneity in

linguistic expressions. Since it is difficult to exhaustively label all bigram counts, and to minimize human

input, we associated bigrams with dimensions using the following procedure: 1) we defined a set of

unigrams that represent each of the dimensions, and 2) we then selected bigram terms containing those

unigrams and appearing at least thirty2 times in the corpora. If a review contained a bigram

corresponding a certain dimension, it received a score of one for that dimension. We selected keywords

with minimum ambiguity to represent each of the five dimensions. The final result from this procedure is

a five dimensional binary vector for each employee review.

The final task is computing firm level measures. For each firm, we compute the share of positive

counts. The assumption embedded in our approach is that workers’ likelihoods of mentioning bigrams

related to these constructs in the pros section of the Glassdoor reviews is an estimate of the underlying

measure of workers’ perceptions of each of these non-wage benefits of the firm. Whether this is an

accurate reflection of the actual value of these benefits is not relevant to our estimation task because we

are ultimately estimate the value it provides to employees.

III.D. Description of Key Measures 1 For example, “flexible”, “flexible schedule” and “flexible payment” suggest different management practices. 2 The results are robust to using a threshold of 10, 20, or 100.

Current Wages and Target Wages: Our principal variables are the target and current wages of employed

workers. Our sample includes 197,490 employed individuals in IT professions who post current wages

and target wages and who identified both wages as earned on annual basis. Among these IT workers,

125,553 report themselves as currently employed.3 To remove potentially noisy outliers from the data,

we drop workers who report earning less than $10,000 per year as their current or target wage and those

who report more than $1,000,000 per year as their current or target wage as those figures are unrealistic

for annual wages in this field and may represent typographical errors.

Education: In our regressions, we control for one of seven education levels (e.g. high school, college).

Race and Gender: Workers’ race and gender were reported by filling out drop-down boxes. Reporting

this information is voluntary, which contributed to missing information.

Experience and Job Tenure: Job tenure was computed from the job history of the employee. Job tenure is

the time starting from when they began their most recent job up until the time at which the wage

information was posted. Therefore this measure is current tenure not completed job tenure.

Year: The database includes data on the year in which the data were submitted. Year dummies control for

systematic changes in labor market conditions across year.

Location: Workers report their city, zip code, and state. Using these data, we assigned each worker to a

metropolitan statistical area (MSA), which we used to define their labor market. We then established

whether workers were located in one of ten high-tech metropolitan areas. In some regressions, we include

a separate dummy variable for each of these regions, and in others, we include a single dummy variable if

the employee is in any one of the high-tech regions.

Occupation: We constructed occupational measures using workers’ reported job titles in their current

employment spells. Using third-party software, we matched job titles to codes in the O*NET-SOC

database which is a comprehensive database developed by the Bureau of Labor Statistics.4

Industry: We matched employers with industries by matching employer names to their Compustat ID

using software provided by Capital IQ (recent acquirers of the Compustat database). We used the

Compustat database to create measures of industry based on North American Industry Classification

System (NAICS) codes. An industry was classified as high-tech if employment in IT-oriented

occupations accounted for a proportion of total employment that is at least twice the percentage average

for all industries (4.9%).

Technological Investment: We use several different measures of whether firms are likely to be working

with emerging information technologies (corresponding to the years between 2006 and 2011). We use a) 3 Examples of IT workers are: “Assistant Technical Director”, “Digital Printer Operator”, “Data Processing Technician”, “ Computer /Network Technician” etc. according to workers’ self-reported job titles. 4 See http://www.onetonline.org.

external datasets on investment into Hadoop-based systems as a proxy, which industry observers have

identified as a key marker of investments related to big data technologies5, b) aggregate IT intensity

measured by IT employment, c) geographic location, and d) industry. More details on these measures are

presented later in the paper.

III.E. Descriptive Statistics and Figures

Table 3 presents the statistics for our key variables. The average annual wage earned by IT workers in

our sample is about $66,000 and the average target wage reported by workers is slightly over $69,000,

which is 1-2% higher than their current wages. The fact that the target wage is not far removed from the

current wage suggests that the former are not unrealistic expectations of the respondent’s worth in the

market. It also suggests that it would not take much to get a typical IT worker to move, which is

consistent with findings on IT mobility (Saxenian 1996; Fallick et al 2006). The average IT worker in our

sample has about eleven years of experience. Slightly over 20% of our sample is located in a high-tech

MSA. The comparable statistics for non-IT workers are shown in the right half of the table.

Figure 1 plots the gap between these two figures against the technological orientation of the firm.

Although both curves for IT workers have a similar general shape, rising with greater investments in one

specific emerging technology, the gap between employees’ wages and their target wages becomes larger

at employers with heavier investments in emerging technologies. The amount required to poach IT

workers, therefore, appears to be higher at firms using emerging information technologies.

We also plot the distribution of these effects across US labor markets. Prior work suggests that

due to the potential for spillovers, both technological investment and complementary human capital

formation cluster in high-tech regions. Employers in thick labor markets are more likely to invest in new

technologies because they can find specialized employees; employees in thick markets invest in

specialized human capital because they have a greater likelihood of finding subsequent employers that

value their specialized human capital (Acemoglu 1997; Rotemberg and Saloner 2000; Moretti 2010).

Figure 2 overlays wage information onto a map of the United States. Darker areas correspond to a

larger gap between IT workers’ target wages and their pay. Gray areas on the map correspond to areas

with less than ten workers in the zip code, so we do not compute statistics for those regions. The patterns

in Figure 2 indicate greater poaching premiums—those closer to the right side of the graph in Figure 1—

are concentrated in regions where IT industries cluster, such as Silicon Valley, New York, Boston, and

the DC-Virginia corridor. This is consistent with the focus on poaching and retention in high-tech labor

markets, and it is also consistent with the argument that the production of IT human capital tends to

5 A more detailed description of Hadoop is presented later in the analysis.

concentrate in high-tech clusters due to the potential for spillovers (Tambe 2014), as well as with prior

work that documents steeper wage profiles in high-tech labor markets (Freedman 2008). Within IT

workers, the highest deviation metropolitan areas are Santa Barbara-Santa Maria-Goleta (10.23%), San

Jose-Sunnyvale-Santa Clara (8.19%), Washington-Arlington-Alexandria (7.12%), San Francisco-

Oakland-Fremont (6.82%), New York-Northern New Jersey-Long Island (6.23%), Bridgeport-Stamford-

Norwalk (6.18%), Trenton-Ewing (5.09%), Provo-Orem (4.39%), Los Angeles-Long Beach-Santa Ana

(4.33%), and Seattle-Tacoma-Bellevue (3.39%), where the comparison group is all other metro areas. All

of these differences are significant at the 5% level.

IV. EMPIRICAL ANALYSIS

IV.A. Empirical Framework

The principal goal of this paper is to analyze poaching wages in high-tech labor markets. To

understand the advantages of our data for analyzing this question, it is useful to consider biases that arise

when running wage regressions using a cross-section of IT workers employed at different firms. Because

firms that adopt early-stage technologies may employ higher quality workers, estimates using a cross-

sectional regression of employers’ technological investment on workers’ wages will be biased upwards by

unobserved heterogeneity in worker quality. It is important to control for that difference, as we do here.

Moreover, variation in panel data on IT wages may be confounded if workers exhibit a ‘preference’ for

learning new technical skills and do not see it as an investment, and therefore accept lower wages at

firms that offer such opportunities. That preference is especially likely for younger workers.

Short of identifying workers before they enter the field of IT, it is difficult to address all possible

sources of endogeneity in the factors that make IT workers choose target wages that are systematically

different from that of other workers. It might be possible with other data to identify situations where the

type of technology used by an employer was exogenous to the attributes of the firm and employee – e.g.,

the technology changed but the workers did not. Short of contexts such as that one, one way to address

the concern about possible factors that could confound the relationship between the use of cutting edge

technology and target wages is with data on how much money workers require to leave the firm or by

using the earnings of workers at different employers who fall along different points on the wage-

technology spectrum. Because the target wage represents the value workers require to leave the firm, it

allows us to back out the value of the non-wage benefits attributed to their current work context.

To test the hypothesis that higher wage offers are required to induce workers from early

technology adopters to leave their employers, we estimate the following multivariate framework:

(1) Log! TW! = !β! + β!Log CW! + β!Technology! + β!Geographic! + β!Occupation! + β!Controls! + ε!!

TW is the worker’s target wage, CW is the worker’s current wage, and the technology measures indicate

the technological orientation of the employer. There are alternative dependent variables we could

consider, such as measuring the gap as a percentage of current income, or testing how much employees

earn, conditional on sharing a common target wage, but we use the model in (1) because it is most

consistent with the existing literature on determinants of workers’ reservation wages.

Because the dependent variable is in logs, the coefficient estimates are interpreted as the

percentage change in the target wage associated with a change in the independent variable. Aside from

the technology measures, controls include education, experience, job tenure, year, race, gender, and

ethnicity. The error term εi in equation (1) captures the effects of factors that are unobserved but that

could impact how high the worker sets her target wage.

Note that because we include current wages on the right hand side, this error term principally

captures the effects of omitted variables that are correlated with the most recent employment spell.

Current wages absorb much of the unobserved heterogeneity among workers that may affect how they set

their target wage, except for heterogeneity relating to differences in the most recent employment

experience that may have caused the target wage to deviate from the current wage. The difference

between the target and current wage, therefore, captures either switching costs or non-wage benefits at

their current employers. Since we do not explicitly model switching costs, an assumption that is

embedded in our estimating framework is that conditional on the control variables, which include industry,

region, occupation, and worker demographic variables, the omitted switching costs are not systematically

correlated with our technology or non-wage benefit measures.

When compared with existing work on IT wages, therefore, this specification is fairly robust to

unobserved differences in worker quality. Our regression results, however, are still subject to concerns

about omitted variables that are related to the technology investments and non-wage benefits that we are

trying to study. We address in a dedicated section in our analysis below.

IV.B. Technology Investment and Target Wages

In Table 4, we report results from baseline regressions. The dependent variable in all regressions

is the logged target wage of the employee. Independent variables include technology measures, logged

current wages, and demographic and human capital variables.

First, we test the hypothesis that conditional on the current wage, technical workers post higher

target wages at firms making investments in emerging information technologies. The paucity of data on

firms’ technological investments is a well-known issue in the empirical IT literature, even at an aggregate

expenditure level, let alone at a level that allows classification of the types of information technologies

firms choose (see Tambe and Hitt 2012 for a detailed discussion). Therefore, we use a menu of measures

that capture different elements of the firm’s technological position.

First, we use a measure of investment in one specific, but important, information technology.

Hadoop-based systems, according to many industry observers, have been a key marker of an emerging

wave of investment in data analytics coinciding with the years covered by our panel (ranging from 2006

through 2011). These technologies have been associated with a new generation of data management skills,

such as Hadoop, Apache Pig, and Map/Reduce, the demand for which industry observers expect to grow

in the coming years. Measurement of employers’ investments in Hadoop is not a comprehensive measure

of early IT adoption; IT applications in mobile, cloud computing, and other areas are also growing rapidly.

However, this class of technologies has been identified by the business press as particularly important for

defining the next wave of innovation and productivity growth (Mckinsey 2011), as well as for defining

future IT labor market trends (Royster 2013). Therefore, in the years covered in our panel, new data

analytic technologies may be among the most important set of disruptive and emerging IT. Moreover,

data on Hadoop investments are a useful proxy for investments in other emerging IT, if on average during

the years in our sample, firms that employ workers with Hadoop skills are more likely than other firms to

be investing in emerging information technologies.

The technology investment data we use are from a recent paper (Tambe 2014). The data measure

year-on-year levels of human capital investment in Hadoop, aggregate IT, and for a smaller sample of

firms, Structured Query Language (SQL), a widely used relational database technology that is relatively

mature by comparison to Hadoop and other emerging database technologies. In the period covered by our

sample, on-the-job learning was probably less valuable for SQL due to the multiplicity of channels

through which these skills can be acquired (e.g. universities, training programs). Therefore, we expect

Hadoop investments to be a superior marker of firms where on-the-job learning may be valuable.

In Table 4, we report estimates from regressions using our Hadoop-based measures. The

coefficients on the full list of demographic variables are omitted for brevity, but demographic variables

and MSA effects are both included in the regressions. Column (1) in Table 4 indicates that Hadoop

investment is associated with workers posting a 1.5% higher target wage, conditional on current pay

(t=4.09). Column (2) adds SQL and aggregate IT investment measures into the regression. The estimates

indicate that workers post a higher target wage when employers are investing in Hadoop. At firms

distinguished by investments in mature data technologies or aggregate IT investments, we do not observe

technical workers posting similarly higher target wages.

To test whether the correlation between technological investment and workers’ pay reflects

heterogeneity in the occupational or industry composition of firms making these investments, we include

industry and occupational fixed effects in subsequent columns. These results are reported in columns (3)

and (4). None of these has a substantive effect on the estimates, which indicates that the relationship

between employers’ technological investments and higher target wage levels is not confounded by

heterogeneity at the occupation or industry level.

The remaining columns in Table 4 probe the robustness of this result in different ways. Columns

(5) through (8) incorporate measures of the firm’s R&D investment, performance (sales/employee), and

size (in assets), and they also report the results for non-IT workers. Although we present a formal

discussion of endogeneity further below, these results suggest that the among firm characteristics, higher

poaching wages are idiosyncratic to firms with emerging IT investments, and they are concentrated within

the IT workforce. The results from this table suggest that the gap between the worker’s target wage and

current wage is 1 to 2% larger at firms that made Hadoop investments during this time period, and that

these effects are unique to Hadoop investments—they are not correlated with aggregate investments in IT

or other firm characteristics like R&D intensity or other performance measures, and we do not observe

such effects for workers in other occupations.

Finally, column (9) includes measures of equity compensation into the regression. To compute

equity compensation measures, we follow existing work in the accounting and finance literatures that

studies the effects of employee stock option plans (Oyer and Schaefer 2005; Hallock and Olson 2006;

Hallock and Olson 2010). The most common way to measure employee stock options is by using the

Compustat ExecuComp database, which contains the number of stock options granted to the top five

executives in the firm, along with data on the total number of stock options that have been granted by the

firm from the Compustat Fundamental Annual database. These numbers can be used to infer the stock

options granted to non-executive employees at the firm. The result in column (9) suggests that the

relationship between technological investment and target wages is not reflecting unmeasured equity

compensation at the worker level. In a subsequent section, we further analyze the potential effects of

equity compensation on our key results.

Alternative Measures of Technology Investment. In Table 5, we substitute Hadoop investment

measures with three alternative measures of the firm’s technological position: 1) the fraction of IT

workers at the firm (IT employment intensity), 2) whether the employer is located in a high-tech

geographic cluster, and 3) whether the employer is in an IT industry. All of these measures capture

different elements of firms’ technology investment choices.

Column (1) replicates our results using the Hadoop measures from Table 4. Column (2)

substitutes a binary measure of whether the employer is in an IT industry. Firms in IT industries tend to

be early adopters of new information technologies. The coefficient estimate on the IT industry variable is

slightly smaller than the estimate in (1), but it is positive and significant. The smaller coefficient on the IT

industry measure reflects that this measure captures broad industry classifications based on outputs, rather

than technological activities based on inputs—there is likely to be considerably heterogeneity both inside

and outside of the IT industry with regards to technological investment.

The largest estimates are generated by using our IT employment intensity measure in column (3).

This measure is the fraction of total workers at the firm who are IT workers, and is the closest measure to

a firm-level aggregate IT investment measure to which we have access (see Tambe and Hitt 2012 for

construction and justification of this measure). The rationale behind the use of this measure is that IT-

intensive firms are more likely to be investing in new technologies. Compared to other measures used in

this analysis, it has the benefit of being a continuous variable, so there is variation beyond the binary

classification of our Hadoop investment or IT industry variables. However, because the measure can be

broadly interpreted, it is also likely to be correlated with a number of other firm-level characteristics,

which is perhaps reflected in the larger coefficient estimate produced by using this measure.

A more precise measure may be labor market location choice. We would expect these effects to

be largest in high-tech clusters for the reasons described above; workers are most likely to specialize in

technical skills when they know there is a thick market for these skills. Moreover, firms should only be

willing to bear the higher costs of land and labor in these areas if they are investing in newer information

technologies, and can therefore directly benefit from these investments. To that end, we can directly use

measures of location as a rough measure of the firm’s technological position. The coefficient estimate on

the high-tech MSA variable that we use in (4) is similar in magnitude to the IT intensity variable.

The most important takeaway from Table 5 is that workers post higher target wages when

employed at firms that are characterized by various measures of technological leadership—not just the

Hadoop investment measure used in Table 4. Although none of these measures perfectly captures a

firm’s technological orientation, they each capture different elements of firms’ investment choices.

IV.C. Technology Investment, Other Non-Wage Benefits, and Target Wages

The coefficient estimates on the technology measures that we observe above can arise for a

number of different reasons. For example, firms that make such technology investments may offer better

internal opportunities or workers at these firms may simply enjoy working with newer technologies, for

which they are willing to accept lower wages at their current employers. This section uses the Glassdoor

online review data to generate direct measures of different categories of non-wage benefits. The methods

used to construct non-wage benefit measures along a number of dimensions, including compensation,

work-life balance, culture, and promotion opportunities was described in an earlier section.

Columns (1) through (5) in Table 6 present estimates from regressions incorporating different

categories of non-wage benefits—skills, overall benefits and compensation, culture, and work-life

balance. All regressions include controls for MSA and four-digit industry.

Panel A presents the results for IT workers. The principal takeaway from Panel A is that among

categories of non-wage benefits, IT workers only consistently post a higher target wage when at

workplaces encouraging skill acquisition and learning. It is important to note that this means that for

omitted variables to bias our results, they must be more correlated with the learning measure than with

any other category of non-wage benefit. Therefore, we expect that omitted variables like employee stock

options, if they were driving workers to raise their target wages, would be more correlated with the

benefits and compensation measure than with the learning measure.

For sake of comparison, Panel B performs a similar analysis for non-technical workers.6 More

so than for technical workers, non-IT workers do appear to raise their target wages in at firms that provide

greater non-wage benefits, including work-life balance, internal career opportunities, and perks and

benefits, but they place less value on learning opportunities at the firm. In sum, the results from Panels A

and B indicate that relative to non-technical workers, technical workers place greater value on workplaces

that facilitate learning and skills as a form of non-wage benefit.

If employees value technological investment because it affords them opportunities to build new

technical skills, then we should only observe a correlation between technological investment and higher

target wages at workplaces that facilitate learning and skills. Table 7 examines how the non-wage benefits

analyzed above interact with technological investment. The estimates in column (1) indicate that for IT

workers, skills are an important form of non-wage compensation, and this relationship persists when

including technology measures into the regression in column (2). Column (3) includes an interaction term

between the two measures and finds that the relationship between technological investment and a higher

target wage is only statistically significant in workplaces that encourage skills and learning.

Columns (4) through (6) provide additional evidence using differences in annual and hourly

workers. The non-wage benefits that we discuss above fall into two categories, those that are primarily

rewarded to annual workers, such as healthcare benefits, and those such as skills and workplace capture,

that provide value to all employees regardless of their employment standing. This distinction is a useful

one, because some of omitted variables of concern, such as equity compensation, are principally rewarded

to annual workers. However, the results in (4) through (6) suggest that the benefits workers derive from

the firm’s investments in technologies are valuable to IT workers but not non-technical workers, although

6 In other regressions, we separately estimate these values for workers in different non-IT occupations, and these behave in predictable ways. For example, sales workers particularly value bonuses, and managers value promotion opportunities. These regression results are not shown for brevity, but are available upon request.

it does not matter if these IT workers are full-time or contract workers. The results in (5) and (6) indicate

that, by comparison, benefits such as healthcare are valuable to workers, regardless of whether they are

technical or non-technical, but only if they are full-time workers. Better workplace culture is a non-wage

benefit captured by all employees, regardless of their occupation or employment status.

The interaction term between emerging technologies and learning provides evidence for our main

argument, which is that technological investment raises the outside wage offer required to lure workers

away from their employers, because of the value that workers attribute to the opportunity to acquire skills

related to these new technologies. Beyond this, the falsification tests in Table 7 are also important

because they provide evidence against several sources of endogeneity. The results are consistent with a

non-wage benefit only for technical workers at firms that have invested in emerging technologies and that

emphasize learning and skill acquisition. Unlike equity compensation or healthcare, this source of non-

wage benefits is specific to technical workers, but available to both hourly and full-time workers. This

means that the class of omitted variables that could bias our estimates must act at the confluence of all of

these variables, and not be correlated with any of the main effects or other interaction effects. In other

words, any omitted variables that bias our estimates must also be specific to technical workers at firms

investing in emerging technologies at workplaces that encourage skill acquisition, and they cannot be

limited to full-time workers. In the next section, we report additional results that minimize the likelihood

that our key results are biased by omitted variables.

IV.D. Additional Robustness Tests

The results presented above suggest that the gap between workers’ earnings and their target wage

is positively correlated with technological position, but only for emerging technologies, and only for

workplaces that encourage skill acquisition. This is consistent with the argument that IT workers place

value on the opportunity to acquire skills, but there may be other factors correlated with skills that can

cause workers’ internal pay to deviate from their poaching wages.

In this section we use the detailed employment histories for workers in our sample, and report

results from a number of falsification tests that take advantage of the fact that these non-wage benefits

should be valued most highly by skill-intensive workers, and particularly for workers whose technical

skills are depreciating due to technological change. Moreover, these workers are not commonly thought

of as the ones that receive the highest pay or the most benefits or have the greatest advancement

opportunities. In other words, the different non-wage benefits have opposing predictions about for which

types of workers we may expect to see higher target wages.

Table 8 reports the results from these tests. In Panel A, we exploit heterogeneity in workers’ prior

career histories. Columns (1) and (2) show that the value assigned to the employer’s technological

investment is larger for technical workers that earn wages below the median level. Columns (3) through

(5) illustrate that the largest effects are for technical workers in the lowest educational group. Together,

these results indicate that lower wage and lower education technical workers are the ones most likely to

express a larger gap between their current pay and their target wage when employed at firms investing in

advanced or emerging information technologies. The human capital offered by these workers to their

employers is more often specific to a tool or technology, so these findings are consistent with the

argument that at least some portion of the poaching wage premium associated with employers’

technological investment is due to skill acquisition. By contrast, these workers are no more likely to

receive high levels of equity compensation or have access to other perks and benefits to which higher paid

workers with greater levels of general human capital do not have access.

Panel B of Table 8 uses elements of prior career histories. These measures are based upon

whether workers are likely to require significant on-the-job learning, or whether they are transitioning

from other firms at which they are likely to have already acquired the skills and human capital required to

support new technologies. The first column in Panel B is based upon measures of the reskilling that

workers require when entering their current jobs based on their prior jobs. The results indicate that for IT

workers, target wages are higher for workers with backgrounds that require more on-the-job learning.

The magnitudes of these effects are not large, but they are statistically significant. The smaller magnitude

of the estimate is due to the fact that we include in the sample workers switching from other backgrounds

at all technology firms, not just the early technology adopters on which our study focuses. The

subsequent two sets of tests narrow the sample to early technology adopters and the magnitude of the

estimates are commensurately larger.

Columns (1) and (2) compare workers who transition from other early technology adopters from

those who transition from other firms. The point estimates are larger for workers who transition from

other early technology adopters and are of a similar magnitude to the estimates reported in our key

regression tables, although the estimates are not statistically significant due to the smaller size. The

comparison is more informative in columns (3) and (4) where we compare workers switching to early

technology adopting firms from other, non-IT backgrounds. The gap between internal pay and the posted

target wage is larger and the difference is statistically significant, suggesting that workers requiring

greater reskilling based on prior career histories have the largest gaps between their wages and their target

wage. These results are also consistent with the argument described above, that workers who derive

greater benefits from skill acquisition are likely to value these non-wage benefits the highest.

Finally, one alternative explanation left unaddressed is that the gap between worker’s pay and

their perceived market value can be attributed to workers in the sample who have mobility restrictions,

because high-tech labor markets are characterized by relatively larger numbers of skilled visa holders

with restricted cross-firm mobility. We can conduct a direct test of this alternative explanation. Along

with other demographic variables, employees in our database report information about their work status—

specifically, whether they are US citizens, if they are able to work for any employer, or whether they have

work restrictions. The first two groups of workers face no mobility restrictions, so if our estimates are not

substantially changed when we limit our sample to these workers, then it is unlikely that our results are

biased by mobility restrictions. The results from this test are reported in column (5). The sample size is

smaller, as expected, but workers who report free mobility comprise the majority of the sample (the

difference in observations between this and the full sample also includes workers who choose not to

specify their work status). The most important takeaway from this test is that the magnitude of the gap for

IT workers between their current pay and perceived value does not change much after we limit the sample

to workers with no mobility restrictions. A Wald test does not reject the hypothesis that the IT worker

coefficients are the same in the two columns (Chi Sq(1)=0.112).

Table 9 reports results based on technical skills and occupational training levels. The results

above present evidence that the employees most affected by recent technological change are those in

skill-based occupations with less general education. For our Hadoop based measures, the specific IT

workers who face the largest disruption to the value of their existing human capital are those who work

with data technologies. The estimates in column (1) of Table 7 are consistent with the argument that the

largest effects are for workers specializing in data management and administration technologies. We

observe these results when using both the data on occupational titles as well as the certificates and skills

data. The second two columns indicate that within IT, there are larger effects for workers who require

more on-the-job training (as judged by the Bureau of Labor Statistics job attributes), but on-the-job

training scores appear to matter little for non-IT workers. This implies that these effects are greatest for

workers for whom on-the-job training matters most, which is again more consistent with a human capital

based explanation than alternative explanations.

V. CONCLUSIONS

This paper provides empirical evidence that higher wage offers are required to hire IT workers

away from early technology adopters than from other firms. We then argue that an underlying reason for

this finding is that the opportunity to acquire new technical skills is an important source of non-wage

compensation in high-tech labor markets.

We make two key contributions to the literature on IT labor markets. First, we provide evidence

that technological investment is an important source of non-wage compensation in high-tech labor

markets. This is important because it has implications for how labor market structure influences IT

investment. Second, if workers attribute compensatory value to these investments, it has implications for

IT career choices, including which workers choose to enter IT professions, because this value can be more

easily captured by some workers than others.

Our findings have implications for employers and for policy makers. Our results connect

employers’ technological orientations with IT compensation. Competition among employers for IT labor

has been a subject of much discussion in the press for at least the last decade. For managers, our findings

provide insight into why poaching in high-tech labor markets requires higher offers than in other markets.

Moreover, our findings suggest that the firm’s technological choices can be an important determinant of

retention outcomes. For a given wage point, in work environments where workers feel that their technical

skills are atrophying, workers may be harder to retain than in environments where workers are learning

new technical skills that will be valuable in their future careers.

Furthermore, our findings help connect labor market structure to the financing of new IT

innovations. One stylized fact from the literature on urban economics is that investments in information

technologies are geographically concentrated in high-tech clusters. Our findings show that in thick labor

markets with unrestricted mobility, workers should be more likely to value human capital as a form of

non-wage compensation. This suggests why it may be advantageous for firms investing in newer

technologies to locate in these markets.

For policy makers, an important implication of our paper is that IT labor supply is sensitive to

technological innovation. This suggests that workers’ income streams are vulnerable to external

technological change, and especially workers who are mid-career during periods of rapid technological

change. This is a source of labor market risk that, to the best of our knowledge, has not been studied in

the existing academic literature. This source of income risk has implications for what types of workers

choose to enter professions, and may help to shed light on why workers in some demographic groups are

less likely to choose IT professions, as well as why workers in IT professions burn out at a rapid rate.

There are a number of caveats to our study. Our study documents wage differentials across firms

with different characteristics, but the sample is limited to job seekers. It would be important to evaluate

whether the sample considered in this study differs in important ways from IT workers in careers

characterized by internal labor market movement. Our data are also from a relatively short (five-year)

panel. It is impossible, therefore, to take a strong position on whether the effects we document are

idiosyncratic to this wave of technological investment, or are more broadly generalizable to IT-enabled

change. Should the data become available, a more robust test of the relationship between technological

orientation and wages might analyze data from a longer time period, encompassing more than one

technology diffusion cycle.

There is significant scope for future work in this area. Because there are many policy questions

related to IT labor supply, the dynamics of IT labor markets remain a promising area of research. How IT

workers acquire skills, how these skills are rewarded on the labor market, and the depreciation of these

skills, all have implications for the IT labor force, the demographic composition of workers who choose

these professions, and the wages they earn. These issues have received relatively little attention in the

academic literature, but the answers to these questions are increasingly important for information policy

decisions as well as for managing technical skills in a tight labor market.

REFERENCES

Acemoglu, D. (1997). Training and innovation in an imperfect labour market. The Review of Economic Studies, 64(3), 445-464.

Agarwal, R., & Ohyama, A. (2013). Industry or academia, basic or applied? Career choices and earnings

trajectories of scientists. Management Science, 59(4), 950-970. Andersson, F., Brown, C., Campbell, B., Chiang, H., & Park, Y. (2008). The Effect of HRM practices and

R&D investment on Worker Productivity. In The Analysis of Firms and Employees: Quantitative and Qualitative Approaches (pp. 19-43). University of Chicago Press.

Ahuja, M. K., Chudoba, K. M., Kacmar, C. J., McKnight, D. H., & George, J. F. (2007). IT road warriors:

Balancing work-family conflict, job autonomy, and work overload to mitigate turnover intentions. Mis Quarterly, 1-17.

Ang, S., Slaughter, S., & Yee Ng, K. (2002). Human capital and institutional determinants of information

technology compensation: Modeling multilevel and cross-level interactions. Management Science, 48(11), 1427-1445.

Bapna, R., Langer, N., Mehra, A., Gopal, R., & Gupta, A. (2013). Human Capital Investments and

Employee Performance: An Analysis of IT Services Industry. Management Science, 59(3), 641-658.

Barley, S. R., & Kunda, G. (2006). Gurus, hired guns, and warm bodies: Itinerant experts in a knowledge

economy. Princeton University Press. Ben-Porath, Y. (1967). The production of human capital and the life cycle of earnings. The Journal of

Political Economy, 352-365. Bidwell, M., & Briscoe, F. (2010). The dynamics of interorganizational careers. Organization Science,

21(5), 1034-1053. Boh, W., Slaughter, S. A., & Espinosa, J. A. (2007). Learning from experience in software development:

A multilevel analysis. Management Science, 53(8), 1315-1331. Brown, C., Reich, M., & Stern, D. (1993). Becoming a high-performance work organization: the role of

security, employee involvement and training. International Journal of Human Resource Management, 4(2), 247-275.

Brynjolfsson, E., & Hitt, L. (1996). Paradox lost? Firm-level evidence on the returns to information

systems spending. Management Science, 42(4), 541-558.

Chari, V. V., & Hopenhayn, H. (1991). Vintage human capital, growth, and the diffusion of new

technology. Journal of Political Economy, 1142-1165. Chatterji, A. K., de Figueiredo Jr, R. J., & Rawley, E. (2013). Learning on the Job? Entrepreneurial

Spawning in the Asset Management Industry. Working Paper. Fallick, B., Fleischman, C. A., & Rebitzer, J. B. (2006). Job-hopping in Silicon Valley: some evidence

concerning the micro-foundations of a high-technology cluster. The Review of Economics and Statistics, 88(3), 472-481.

Feldstein, M., & Poterba, J. (1984). Unemployment insurance and reservation wages. Journal of Public

Economics, 23(1), 141-167. Forman, C. (2005). The corporate digital divide: Determinants of Internet adoption. Management Science,

51(4), 641-654. 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), 389-420.

Fox, J. T. "Estimating the Employer Switching Costs and Wage Responses of Forward-Looking Engineers." Journal of Labor Economics 28.2 (2010).

Franco, A. M., & Filson, D. (2006). Spinouts: knowledge diffusion through employee mobility. The

RAND Journal of Economics, 37(4), 841-860. Freedman, M. L. (2008). Job hopping, earnings dynamics, and industrial agglomeration in the software

publishing industry. Journal of Urban Economics, 64(3), 590-600. Gordon, R. H., & Blinder, A. S. (1980). Market wages, reservation wages, and retirement decisions.

Journal of Public Economics, 14(2), 277-308. Griliches, Z. (1992). The Search for R&D Spillovers. The Scandinavian Journal of Economics, S29-S47. Hall, R. E., & Krueger, A. B. (2012). Evidence on the incidence of wage posting, wage bargaining, and

on-the-job search. American Economic Journal: Macroeconomics, 4(4), 56-67. Hallock, K. F., & Olson, C. (2006). The value of stock options to non-executive employees (No. w11950).

National Bureau of Economic Research. Hallock, K. F., & Olson, C. A. (2010). New data for answering old questions regarding employee stock

options. In Labor in the new economy (pp. 149-180). University of Chicago Press. Hann, I. H., Roberts, J. A., & Slaughter, S. A. (2013). All Are Not Equal: An Examination of the

Economic Returns to Different Forms of Participation in Open Source Software Communities. Information Systems Research, 24(3), 520-538.

Igbaria, M., & Greenhaus, J. H. (1992). Determinants of MIS employees' turnover intentions: a structural

equation model. Communications of the ACM, 35(2), 34-49.

Igbaria, M., & Siegel, S. R. (1992). The reasons for turnover of information systems personnel.

Information & Management, 23(6), 321-330. Joseph, D., Fong Boh, W., Ang, S., & Slaughter, S. A. (2012). The career paths less (or more) traveled: a

sequence analysis of IT career histories, mobility patterns, and career success. MIS Quarterly, 36(2).

Joseph, D., Ng, K. Y., Koh, C., & Ang, S. (2007). Turnover of information technology professionals: a

narrative review, meta-analytic structural equation modeling, and model development. MIS Quarterly, 31(3), 547-577.

Jovanovic, B., & Nyarko, Y. (1996). Learning by Doing and the Choice of Technology. Econometrica:

Journal of the Econometric Society, 1299-1310. Kim, K., Mithas, S., Whitaker, J., & Roy, P. Industry-Specific Human Capital and Wages: Evidence from

the Business Process Outsourcing Industry. Information Systems Research, forthcoming. Kuhn, P., & Skuterud, M. (2000). Job search methods: Internet versus traditional. Monthly Lab. Rev., 123,

3. Lancaster, T., & Chesher, A. (1983). An econometric analysis of reservation wages. Econometrica:

Journal of the Econometric Society, 1661-1676. Levina, N., & Xin, M. (2007). Research Note-Comparing IT Workers' Compensation Across Country

Contexts: Demographic, Human Capital, and Institutional Factors. Information Systems Research, 18(2), 193-210.

Marx, M., Strumsky, D., & Fleming, L. (2009). Mobility, skills, and the Michigan non-compete

experiment. Management Science, 55(6), 875-889. McElheran, K. (2015). Do market leaders lead in business process innovation? The case (s) of e-business

adoption. Management Science. Møen, J. (2005). Is Mobility of Technical Personnel a Source of R&D Spillovers? Journal of Labor

Economics, 23(1), 81-114. Moretti, E. (2010). Local Labor Markets, in Handbook Of Labor Economics, Vol 4A (Vol. 4). Elsevier.

Edited by O. Ashenfelter and D. Card. Mithas, S., & Krishnan, M. S. (2008). Human capital and institutional effects in the compensation of

information technology professionals in the United States. Management Science, 54(3), 415-428. Mithas, S., & Lucas Jr, H. C. (2010). Are foreign IT workers cheaper? US visa policies and compensation

of information technology professionals. Management Science, 56(5), 745-765. Moore, J. E. (2000). One road to turnover: An examination of work exhaustion in technology

professionals. Mis Quarterly, 141-168.

O'Mahony, S., & Bechky, B. A. (2006). Stretchwork: Managing the career progression paradox in external labor markets. Academy of Management Journal, 49(5), 918-941.

Oyer, P., & Schaefer, S. (2005). Why do some firms give stock options to all employees?: An empirical

examination of alternative theories. Journal of financial Economics, 76(1), 99-133. Roach, M., & Sauermann, H. (2010). A taste for science? PhD scientists’ academic orientation and self-

selection into research careers in industry. Research Policy, 39(3), 422-434. Roberts, J. A., Hann, I. H., & Slaughter, S. A. (2006). Understanding the motivations, participation, and

performance of open source software developers: A longitudinal study of the Apache projects. Management science, 52(7), 984-999.

Royster, S. (2013). Working with big data. Occupational Outlook Quarterly. Accessed at

http://www.bls.gov/opub/ooq/2013/fall/art01.pdf on September 2nd, 2014. Rotemberg, J. J., & Saloner, G. (2000). Competition and human capital accumulation: a theory of

interregional specialization and trade. Regional Science and Urban Economics, 30(4), 373-404. Sauermann, H., & Roach, M. (2014). Not all scientists pay to be scientists: PhDs’ preferences for

publishing in industrial employment. Research Policy, 43(1), 32-47. Saxenian, A. Regional advantage. Harvard University Press, 1996. Shimer, R., & Werning, I. (2007). Reservation wages and unemployment insurance. The Quarterly

Journal of Economics, 122(3), 1145-1185. Slaughter, S. A., Ang, S., & Fong Boh, W. (2007). Firm-specific human capital and compensation

organizational tenure profiles: An archival analysis of salary data for it. Human Resource Management, 46(3), 373-394.

Stern, S. (2004). Do scientists pay to be scientists? Management Science, 50(6), 835-853. Tambe, P., & Hitt, L. M. (2013). Job hopping, information technology spillovers, and productivity growth.

Management Science, 60(2), 338-355. Tambe, P. (2014). Big Data Investment, Skills, and Firm Value. Management Science, forthcoming. Thatcher, J. B., Stepina, L. P., & Boyle, R. J. (2003). Turnover of information technology workers:

Examining empirically the influence of attitudes, job characteristics, and external markets. Journal of Management Information Systems, 19(3), 231-261.

Thibodeau, P. (2014). IBM cuts pay by 10% for workers picked for training. ComputerWorld. Accessed

at http://www.computerworld.com/article/2683239/ibm-cuts-pay-by-10-for-workers-picked-for-training.html on September 26, 2014.

TABLE 1: SAMPLE COMPARISONS WITH IT WORKERS IN CURRENT POPULATION SURVEY (CPS)

Job Board CPS

A. Gender Female 17.4 28.1 Male 60.9 71.9 Not Specified 21.7 0.00

B. Educational Degree

None 26.2 16.1 Vocational School 3.06 4.77 High School & Below 7.15 9.00 Two Year Degree 14.5 5.98 Four Year Degree 36.2 42.8 Graduate Degree 15.3 19.0 Doctorate Degree 0.70 1.86

C. Ethnicity

African American/Black 12.0 6.25 Asian 6.99 13.7 Pacific Islander 0.41 0.47 Hispanic 5.52 0.00 American Indian 0.55 0.20 Native American 1.25 0.00 Other 28.6 21.6 White 44.8 78.0

D. Industry

High-Tech 47.6 45.3 Finance and Insurance 16.3 11.5 Other 36.1 43.2

TABLE 2: IT WAGE COMPARISONS WITH THE BLS OCCUPATIONAL EMPLOYMENT SURVEY (OES)

Jobs Board OES

A. By Industry (High-Tech Industries) Computer and Peripheral Equipment Manufacturing 78792.76 74990 Software Publishers 86825.73 77570 Other Telecommunications 72854.64 54940 Internet Service Providers and Web Search Portals 59400 64900 Data Processing, Hosting, and Related Services 64863.35 57450 Computer Systems Design and Related Services 66659.22 72380 Scientific Research and Development Services 72890.63 71190

B. By Metropolitan Area (High Tech Regions) Los Angeles-Long Beach-Santa Ana, CA 91961.96 72400 San Francisco-Oakland-Fremont, CA 109733.8 84470 Washington-Arlington-Alexandria, DC-VA-MD-WV 97297.33 81110 New York-Northern New Jersey-Long Island, NY-NJ-PA 101890.5 79390 Seattle-Tacoma-Bellevue, WA 96643.69 79320 San Jose-Sunnyvale-Santa Clara, CA 113790.4 94590 Santa Barbara-Santa Maria, CA 71000 69120

C. By Occupations (Technical Workers)

Computer Programmers 62042.18 69500 Software Developers, Systems Software 73429.2 87250 Executive Secretaries and Administrative Assistants 54415.33 39160 Computer Systems Analysts 64848.62 72230 Network and Computer Systems Administrators 55306.53 65260 Computer Support Specialists 46953.73 44350

TABLE 3: STATISTICS FOR KEY REGRESSION VARIABLES FOR EMPLOYED WORKERS

Mean Std. Dev Observations Current Wage 66238.2 41613.1 125,553 Target Wage 69193.5 37196.1 125,553 Log(Current Wage) 10.985 0.462 125,553 Log(Target Wage) 11.051 0.418 125,553 Experience 10.588 7.428 125,455 Experience2 167.3 225.0 125,455 Tenure 3.033 3.082 125,375 Tenure2 18.7 51.3 125,375 High Tech MSA(0/1) 0.2200161 N/A 66,936 EMTECH 0.3156624 N/A 19,052 Non ESO Share 0.813 0.191 9,644

FIGURE 1: TECHNOLOGY INVESTMENT AND GAP BETWEEN CURRENT AND TARGET WAGE

FIGURE 2: GEOGRAPHIC DIFFERENTIALS FOR GAP BETWEEN TARGET AND CURRENT WAGE

T

AB

LE 4: FIR

M L

EVEL M

EASU

RES O

F EM

ERG

ING

TEC

HN

OLO

GY

INV

ESTMEN

T, 2006-2011

DV

: Log(TW)

(1) (2)

(3) (4)

(5) (6)

(7) (8)

IT

Sales O

ther Equity

EM

TECH

0.018

*** 0.00911

* 0.0192

** 0.0178

*** 0.0164

** -0.00643

0.00344 0.0138

*

(0.01)

(0.00) (0.01)

(0.00) (0.01)

(0.05) (0.01)

(0.01) Log(M

ature IT)

-0.00129

(0.00)

Log(Total IT)

0.00279

(0.00)

Log(Assets)

-0.00605*

-0.00305 -0.000523

(0.00) (0.02)

(0.00)

Log(R&

D)

0.00545*

0.00458 0.00245

(0.00) (0.02)

(0.00)

Sales/Employee

-0.00393 -0.0259

-0.00421

(0.01)

(0.03) (0.01)

Stock O

ptions

-0.002

(0.02)

Log(CW

) 0.758

*** 0.746

*** 0.752

*** 0.768

*** 0.747

*** 0.843

*** 0.792

***

(0.01)

(0.00) (0.01)

(0.00) (0.01)

(0.05) (0.01)

Obs.

5,828 20,117

5,828 5,828

4,521 238

1,836 4,461

R2

0.7512 0.728

0.751 0.728

0.757 0.807

0.770 0.789

Fixed Effects M

SA

MSA

O

ccupation Industry

MSA

M

SA

MSA

M

SA

R

obust standard errors in parentheses; *** p<0.01, ** p<0.05, * p<0.1; All regressions control for dem

ographic information and the year in

which the em

ployment history is posted. C

olumn (3)-(5) show

s the results using the same specification as colum

n (1) with subsam

ple of workers

that have all information about location, occupation and industry by adding fixed effects.

TABLE 5: OTHER FIRM LEVEL MEASURES OF TECHNOLOGICAL LEADERSHIP

DV: Log(TW) (1) (2) (3) (4) EMTECH 0.0136*** (0.01) IT Industry (0/1) 0.00920** (0.00) IT Employment Intensity 0.0569*** (0.01) High Tech Region 0.0470*** (0.00) R-squared 0.750 N/A N/A N/A Obs. 61,652 13,151 9,888 17,188 Fixed Effects MSA MSA MSA MSA Table Notes: Robust standard errors are reported in parentheses; *** p<0.01, ** p<0.05, * p<0.1; All results include controls for individual level demographics including gender, ethnicity, and wage posting year. Emerging IT is 1 if a company hired at least one Hadoop engineer in a certain year, 0 otherwise. IT Worker Share is the percentage of IT workers in a company.

TABLE 6: GLASSDOOR MEASURES OF NON-WAGE BENEFITS

(1) (2) (3) (4) (5) Non-Wage Benefits LEARNING ADVANCE CULTURE COMP BALANCE Panel A: IT Workers BENEFIT 0.150** 0.0354 0.0388 -0.0215 -0.00605

(0.08) (0.04) (0.05) (0.03) (0.05) Log(CW) 0.785*** 0.785*** 0.782*** 0.778*** 0.777***

(0.01) (0.01) (0.01) (0.01) (0.01)

Controls NAICS4 NAICS4 NAICS4 NAICS4 NAICS4 Observations 4,427 5,344 5,165 5,309 4,843 R-squared 0.807 0.815 0.807 0.814 0.804 Fixed Effects MSA MSA MSA MSA MSA Panel B: Non IT Workers BENEFIT 0.180 0.0993** 0.125** 0.0808** 0.0708

(0.14) (0.04) (0.07) (0.04) (0.09) Log(CW) 0.704*** 0.705*** 0.720*** 0.706*** 0.720***

(0.01) (0.01) (0.01) (0.01) (0.01)

Controls NAICS4 NAICS4 NAICS4 NAICS4 NAICS4 Obs. 3,763 4,631 4,516 4,534 4,030 R-squared 0.751 0.747 0.757 0.741 0.754 Fixed Effects MSA MSA MSA MSA MSA Table Notes: Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1; Each column represents a firm level attribute associated with that keyword. The intuition is the percentage of workers who are willing to comment on a certain attributes, identified by a set of bigrams that share the same keyword, in the "pros" section. Higher value means the terms are more likely to be mentioned in the "pros" section. Procedures used to construct measures: 1. Preprocess the "pros" comments with standard procedure (removing stop words etc.), transform each review into a bigram representation; 2. Compute the term frequency for bigrams; 3. Select bigrams that contain the keyword choose bigrams that appears at least 30 times in the corpora. This procedure ensures a certain level of common consensus for the semantic meaning of each bigram; The results are robust to setting threshold at 10 or 20. 4. A comment is labeled as 1 if any of the bigrams appears in the bigram representation of the reviews; A firm level benefit measure is the share of comments equal to 1. Results are estimated using firms with at least five reviews and at least one review commenting on the corresponding benefit.

TABLE 7: COMPARISON OF TECHNOLOGICAL INVESTMENT AND BENEFITS

(1) (2) (3) (4) (5) (6) HEALTH CULTURE EMTECH 0.0024 0.0456 0.0011

(0.01) (0.03) (0.01)

LEARNING 0.168** 0.173** 0.0665 (0.08) (0.08) (0.08) EMTECH x LEARNING 0.324* (0.19) EMTECH x IT 0.0122** (0.01) EMTECH x ANNUAL 0.0040

(0.01)

BENEFITS BENEFITS x IT 0.00151 0.0168 (0.01) (0.01) BENEFITS x ANNUAL 0.0246** 0.0172 (0.01) (0.01) Obs. 4,427 4,427 5,774 25,504 49,810 49,810 R-squared 0.807 0.807 0.852 0.867 0.852 0.852 Fixed Effects MSA MSA MSA MSA MSA MSA Table notes: each column represents the independent variable in the regression. The benefit variable row corresponds to the specific benefit in the column header. Perks: discount, lunch, snacks, dinner, coffee, tea, gym, commute; Health: insurance, health, dental vision; Finance: pension, 401k, stock options, retirement plan, equity; Family: maternity, work from home, children, flexible

TABLE 8: EMPLOYEE CHARACTERISTICS AND TARGET WAGES

Panel A: Worker Characteristics Below median

wage Above median

wage High school or

below 2 Year

/4 Year Degree Graduate /Doctoral

(1) (2) (3) (4) (5) EMTECH 0.019*** 0.009 0.020** 0.011 0.014 (0.01) (0.01) (0.01) (0.010) (0.010) Observations 4,405 4,835 2,827 4,737 1,676 R-squared 0.557 0.498 0.725 0.774 0.768 Fixed Effects MSA MSA MSA MSA MSA Panel B: Prior Experience EMTECH=1

in prior firm EMTECH=0 in prior firm

Previously IT Worker

First IT Job Citizen

(1) (2) (3) (4) (5) EMTECH 0.011 0.022 -0.021 0.054* 0.013**

(0.038) (0.014) (0.031) (0.030) (0.01) Observations 16,795 465 1,458 509 5,972 R-squared 0.760 0.684 0.799 0.634 0.778 Fixed Effects MSA MSA MSA MSA MSA Table Notes: Column(1) and (2) show the results based on sub sample of workers have different income level in the data set. The median income is 65,000. Column(3)-(5) split sample based on education group. *Notes: Robust standard errors are reported in parentheses; *** p<0.01, ** p<0.05, * p<0.1; All regressions control for demographics including gender, degree, ethnicity, resume edit year, experience and tenure. Non ESO Share is computed as NES = 1 − !"#$!%&!!"#$%&'!!"!!"!#$%&'!(

!"#$!%&!!"#$%& Options granted to executives are calculated by summing the options granted to the top five executives (options_award_num in Compustat ExecuComp); Options granted are directly from Compustat (OPTGR); Regressions also control for firm size (LOG (AT)) and total shares outstanding (CSHO). Each worker is matched to the year in which she was hired by tracking the start year of the most recent job. Work life balance and benefit and compensation measures are created using the average ratings (scale from 1-5) reported by workers on Glassdoor.com. Bonuses includes workers’ self reported cash bonuses, stock bonuses and profit sharing. Column (1) is restricted to a sample of workers who report themselves as “citizens“ or “can work for any employer” and therefore have no mobility restrictions. Column (2) and (3) report results with non-executive employee share of stock options;

TABLE 9: SKILL GROUPS AND TARGET WAGES DV: Log(TW) (1) (2) (3) IT Occupations IT Occupations Non-IT Occupations Network Occupations -0.003 (0.01) Database Occupations 0.039*** (0.01) Programming Occupations 0.02 (0.01) Network Skills 0.002 (0.01) Database Skills 0.047*** (0.01) Programming Skills -0.019 (0.01) Higher Education 0.059*** 0.043*** (0.00) (0.01) More Experience 0.053*** 0.033*** (0.01) (0.01) Onsite training 0.027*** 0.0111 (0.01) (0.01) On-the-job training 0.019*** -0.006 (0.01) (0.01) Observations 4,328 42,083 27,808 R-squared 0.737 0.621 0.667 Fixed Effects 2 Dig SOC Robust standard errors in parentheses; *** p<0.01, ** p<0.05, * p<0.1; *Notes: All regressions control for demographics including gender, degree, ethnicity, experience, tenure and resume edit year. Columns (1) includes dummy variables that are equal to one if the job title contains a certain keyword indicating the type of task an IT worker performs and dummy variables equal to one if the self reported skills (certifications) contains a keyword indicating the skill obtained by an IT worker. List of keywords for screening skills/certificate: General Skills: A+, MCSE, MCP, Microsoft, CompTIA, Windows, MCSA,MS; Network Skills: CCNA; Network+; CCNP, Net+, Cisco, N+, CNE, Novell, Web, Security+; Data Base Skills: Oracle, SQL, Access, Database, ITIL, Sun; Programming Skills: Java, Basic, Program, Visual; Other Skills: PMI, PMP, Sigma, Quality List of keywords for screening job titles: Network Occupations: Network, Web, LAN, Webmaster, Networking; Database Occupations: Data, Database, Oracle, SAP, Warehouse, SQL, MIS; Programming Occupations: Programmer, Programming, JAVA, Software, software, Developer

APPENDIX A: COMPARISON OF COMPUSTAT WORKER SAMPLE WITH LARGER RESUME SAMPLE

TABLE A.1: COMPARISON OF KEY REGRESSION VARIABLES

Workers in Resume + Compustat Sample Workers in Resume Sample Mean Obs. Mean Obs. Target Wage 75258.87 19052 68108.444 106501 Current Wage 72242.446 19052 65164.124 106501 Log(TW) 11.137032 19052 11.035507 106501 Log(CW) 11.078707 19052 10.968516 106501 Experience 10.738078 19040 10.560927 106415 Tenure 3.1689711 19021 3.008735 106354

TABLE A.2: COMPARISON OF DEMOGRAPHIC VARIABLES

Workers in Resume +

Compustat Sample Workers in

Resume Sample Gender Obs. % Obs. % Female 3,144 17.79 15,899 16.04 Male 10,081 57.04 59,964 60.49 Unknown 4,448 25.17 23,261 23.47 Total 17,673 100.00 99,124 100.00 Education Obs. % Obs. % None/Not Specified 4,677 24.65 26,505 24.98 Vocational Degree 411 2.17 3,055 2.88 High School Degree 669 3.53 4,014 3.78 2 Year Degree 2,371 12.50 15,293 14.42 4 Year Degree 7,390 38.95 39,454 37.19 Graduate Degree 3,358 17.70 17,005 16.03 Doctorate 99 0.52 752 0.71 Total 18,975 100.00 106,078 100.00 Ethnicity Obs. % Obs. % African American/Black 2,368 12.55 12,239 11.60 Asian 1,547 8.20 7,152 6.78 Native Hawaiian/Pacific Islander 74 0.39 445 0.42 Hispanic 878 4.65 5,772 5.47 American Indian or Alaska Native 78 0.41 562 0.53 Not Specified 6,021 31.91 31,686 30.03 Other 204 1.08 1,308 1.24 White 7,701 40.81 46,335 43.92 Total 18,871 100.00 105,499 100.00