Upload
mert0723
View
215
Download
0
Embed Size (px)
Citation preview
8/14/2019 jkotter-jmp.pdf
1/67
Technological Change, Job Tasks, and CEO Pay
JASON KOTTER
October 8, 2013
Job Market Paper
ABSTRACT
This paper argues that managing human capital changes the role of the CEO. Managers andnonroutine task workers exhibit synergies; by focusing on this synergy managers increase thevalue of the firm. Using text analysis of 10-K statements, I provide evidence that the focus ofmanagers has shifted away from the operations of the firm and toward the people of the firm.When accompanied by an high human capital labor force, this shift in focus leads to large (6-14%) increases in firm value. To induce managers to shift their focus and realize this synergy,shareholders optimally increase CEO pay. Using a difference in difference approach, I estimatethat changes in the human capital of the workforce induced by the computer revolution caused
CEO pay to double. This explains roughly half of the aggregate increase in CEO pay over the lastthirty years, suggesting that a substantial portion of the increase in CEO pay over the past threedecades represents an optimal response to skill-biased technological change.
JEL Classification Numbers:G32, G34, J33
Keywords:Executive compensation; task-biased technological change; human capital.
Ross School of Business, University of Michigan. Please send correspondence to [email protected].
8/14/2019 jkotter-jmp.pdf
2/67
Allocating human resources in a strategic manner is a key aspect of the CEOs role ...Little if
anything else that I do as CEO will have as enduring an impact . . . A.G. Lafley, P&G CEO
In 1910, about 14% of the U.S. adult population had graduated from high school, while only 3%
had earned a bachelors degree. By 2012, 88% of the workforce had completed high school and
nearly 31% had graduated from college (See Figure 1). How does this massive increase in human
capital affect firms? While human capital is a key component of economic growth, the increased
importance of skilled labor poses new challenges for firms (Zingales (2000)).1 Unlike physical
capital, human capital can voluntarily walk away from the firm. In addition, the productivity of
human capital is particularly sensitive to the work environment (Drucker (1999)). This suggests
that changes in the human capital of the workforce alter the role of executive managers. In this
paper, I examine how managing human capital affects the role of managers. I provide evidence
that the increase in human capital over the last thirty years caused average CEO pay to increase
by about $1.8 million, which explains about half of the actual increase in pay over this time. This
increase is consistent with an optimal compensation contract that induces the CEO to focus on
labor. I show that managers of high human capital workforces focus more on their employees
and that this focus increases the value of the firm.
[Figure 1 about here.]
The primary challenge in linking changes in human capital of the labor force to CEO out-
comes is distinguishing between treatment effects (caused by increased human capital) and se-
lection effects (associated with the type of CEO that manages skilled workforces). Since CEOs are
ultimately responsible for the composition of their workforce, highly skilled managers might dis-
proportionately hire skilled workers. This could be due to complementarities between managers
and skilled labor, but could also arise due to manager preferences to work with a similar back-
ground.2 If CEO labor markets are efficient such that more skilled CEOs command higher wages,
1A large empirical literature demonstrates that human capital leads to economic growth. Ciccone and Papaioan-nou (2009), Hanushek and Kimko (2000), Glaeser, Porta, Lopez-de Silanes, and Shleifer (2004) are a few among manyexamples of this work.
2For example, managers might give preference to hiring alumni of the University they attended.
1
8/14/2019 jkotter-jmp.pdf
3/67
this selection leads us to over-estimate the effect of human capital on CEO pay. Alternatively, if
high human capital workforces are difficult for shareholders to monitor, managers might be able
to extract more rents. This selection also leads to over-estimates of the effect of human capital.
To overcome these problems, the ideal experiment would either randomly assign CEOs to
workforces of different human capital levels or randomly assign workers of different skill levels to
CEOs. While neither of these experiments exists, there is a natural experiment that approximates
the second situation for workers that perform routine tasks. In the early 1970s, the invention of
the microprocessor and personal hard drive ushered in what is sometimes called the computer
revolution. This technology shock had a huge, but heterogenous effect on the labor composition
of firms. Technology allowed firms to computerize routine tasks (i.e., tasks that can be completed
by following explicit directions), but has not (yet) allowed firms to computerize nonroutine tasks
that involve creativity, critical thinking, and complex communication. As a result, firms that
were highly exposed to routine tasks before the computer revolution experienced large increases
in the human capital of their workforce as they replaced routine workers with computers and
hired additional nonroutine workers, while the composition of the workforce at firms with low
exposure to routine tasks stayed relatively constant. I utilize the variation in industries exposure
to routine task workers as a plausibly exogenous source of variation in the changes in a firms
human capital. This variation allows me to identify the causal effects of changes in human capital
on executive compensation.
To proceed, it is necessary to develop a measure of the human capital of a firms workforce. I
use public person-level Census data to create a measure of routine and nonroutine task intensity.
The Bureau of Labor Statistics scores each occupation in the Census (e.g., accountant) on dozens
of dimensions. I follow Autor, Levy, and Murnane (2003) and pick out several dimensions that arecorrelated with performing routine and nonroutine tasks; these dimensions are used to create an
index of nonroutine and routine task intensity for each occupation. I then aggregate each person
to the industry level (e.g., auto manufacturing), weighting by the persons full-time equivalent
hours. I do this each year from 1973 to 2009 so that I have a time-varying industry measure of
2
8/14/2019 jkotter-jmp.pdf
4/67
nonroutine and routine task intensity.3
Following the estimation strategy of Stevenson (2010), I use the pre-technology shock level
of routine task intensity in an industry to instrument for exogenous changes in the labor force.
Industries with high routine task intensity form my treatment group, while industries with low
routine task intensity form my control group. I then use a difference-in-difference approach to
estimate the effect of increases in human capital on CEO pay.
Conceptually, my approach compares the growth rate of CEO pay at a firm that was highly
dependent on routine workers before the computer revolution (such as Ford) to the growth rate
of CEO pay at a firm that was not dependent on routine workers (such as Pfizer). The computer
shock allows Ford to replace many of their production line workers with robots; simultaneously,
Ford hires technicians to service the robots and analysts to sift through the data produced by
robots in search of potential efficiencies. While the technology shock increases the productivity
of research scientists at Pfizer, the composition of the workforce (which was always relatively
high skill) does not change. Comparing the difference in growth rates of CEO pay over time
at Pfizer and Ford reveals the effect of managing human capital on CEO pay as long as I have
adequately controlled for any other differences between Pfizer and Ford that also affect pay.
Importantly, this estimation strategy does not just compare high skill versus low skill indus-
tries (e.g., pharmaceutical versus manufacturing). Many highly skilled industries were also highly
exposed to routine tasks. For instance, the banking industry, though dependent on high skilled
labor, also relied extensively on many routine bookkeeping tasks. The computer shock allowed
banks to replace many bookkeepers with Excel spreadsheets. Similarly, many lower skilled ser-
vice industries, such as hair salons, do not perform routine tasks. As a result, my treatment and
control groups include firms across a wide spectrum of skilled labor. The identification of theeffect of human capital on CEO pay comes not from variation in labor force skill, but from varia-
tion in exposure to the computer revolution. This variation is plausibly exogenous to changes in
CEO pay.
3Autor et al. (2003) create broad aggregate measures of nonroutine and routine task intensity. To my knowledge,this is the first attempt to create disaggregate measures that can be used at the firm level.
3
8/14/2019 jkotter-jmp.pdf
5/67
Using this difference-in-difference framework, I estimate that the change in human capital
over the last thirty years caused CEO pay to increase by a factor of 2.4. Actual CEO pay increased
about 4 times, implying that changes in the skill-level of the workforce explain approximately 50%
of the increase in CEO pay since 1984. This is, to my knowledge, the first direct evidence of the
role of skill-biased technological change on CEO pay.
These results are robust to a variety of alternative specifications, including an instrumental
variable approach where identification relies on unanticipated changes in computer prices. Gov-
ernance failures cannot explain the relationship between human capital and executive pay, nor
can offshoring.
Why does managing human capital increase pay? Because human capital is a key driver of
growth, managers have strong incentives to pay attention to high human capital workers.4 For
example, in 2005 Jeff Immelt, GEs CEO, wrote in his annual letter to shareholders, Developing
and motivating people is the most important part of my job. This focus on people changes
the role of the CEO from command and control to coach (Hambrick, Finkelstein, and Mooney
(2005), Finkelstein and Peteraf (2007)). IBMs CEO Sam Palmisano puts it this way, You just cant
impose command-and-control mechanisms on a large, highly professional workforce. Im not
only talking about our scientists, engineers, and consultants. More than 200,000 of our employees
have college degrees. The CEO cant say to them, Get in line and follow me. Or Ive decided
what your values are. Theyre too smart for that. And as you know, smarter people tend to be,
well, a little more challenging; you might even say cynical.
As a coach or mentor, the CEO exhibits particular synergies with nonroutine task employees.
By definition, nonroutine tasks are somewhat nebulous; as a result it is costly for a worker to
figure out how to accomplish the task. A CEO in a mentor role is able to decrease the cost ofeffort to the employee (and thus increase productivity of the employee) by more clearly defining
the scope of the task.5 Managed well, nonroutine task workers are able to innovate and increase
4Management scholar Peter Drucker describes it this way: The most valuable assets of a 20th-century companywere its production equipment. The most valuable asset of a 21st-century institution, whether business or non-business, will be its knowledge workers and their productivity (Drucker (1999)).
5This doesnot necessarily happen at the one-on-one level. Creating corporate cultures through vision statements,
4
8/14/2019 jkotter-jmp.pdf
6/67
the production possibility frontier for the firm. Consistent with the contracting model in Edmans,
Goldstein, and Zhu (2011), shareholders optimally increase pay to induce managers to focus on
this synergy.
To help confirm that this is the channel driving increased pay and to better understand the
changing role of the CEO, I create a measure of CEO focus using text analysis of 10-K reports.
Using the management discussion and analysis (MD&A) section of the 10-K, I count the number of
people related words (e.g., employee, staff, labor, etc.) and the number of operations related words
(e.g., cash flow, margin, performance, etc.) and scale each count by the total number of words in
the MD&A. Using these measures, I show that management focus on people has increased and
focus on operations has decreased from 1994 to the present. These trends do not seem to driven
by aggregate changes in the economy (i.e., a switch from a manufacturing to a service economy).
While I cannot claim causality, this shift in management focus is positively correlated with
the nonroutine intensity of the firms workforce. Additionally, there is evidence of a positive
synergy between managers and nonroutine workers. Focusing on people increases both the value
and profitability of the firm, but only for firms with high human capital workforces. Using the
estimate of the CEO fixed effect in pay regressions as a measure of CEO ability, I also show that
firms with nonroutine task workers hire high ability CEOs and that CEOs in these firms matter
more for profitability and stock returns. Taken together, the evidence suggests that managing
human capital increases the importance of the CEO through synergies between managers and
nonroutine workers.
This paper contributes both to the literature on human capital and the literature on CEO pay.
One of the central contributions of this paper is to link these two literatures together by provid-
ing evidence that some of the increase in CEO pay is due to the rise of nonroutine workers. Thislink sheds light on the longstanding debate on the optimality of CEO pay. Critics of current CEO
pay practices argue that CEOs have captured the board of directors and are consequently pay-
ing themselves too much at shareholders expense (Bertrand and Mullainathan (2001), Bebchuk,
mottos, and shared goals can be viewed as an attempt to define what matters to the company. This helps nonroutineemployees understand where to focus their efforts.
5
8/14/2019 jkotter-jmp.pdf
7/67
Fried, and Walker (2002), Bebchuk and Fried (2003), Bebchuk and Grinstein (2005), Bebchuk, Grin-
stein, and Peyer (2010)), but other research suggests that the rise in pay is a result of optimal con-
tracting in the market for CEO talent (Gabaix and Landier (2008), Murphy and Zabojnik (2004),
Frydman (2005), Kaplan and Rauh (2010), Cremers and Grinstein (2011), Falato, Li, and Milbourn
(2011)) While both explanations might be relevant to the cross-sectional differences in CEO pay,
neither approach is fully consistent with the long-run time trendpay remained relatively stag-
nant from the 1940s to the early 1970s when it began to grow rapidly (Frydman and Saks (2010)).
My results help to resolve this puzzle with the insight that it is not the overall growth of the
firm, but the combination of the growth in size and the growth in skilled labor, that leads pay
packages to optimally increase. This type of firm growth begins in the 1970s with the drastic fall
in the price of computer capital. Consequently, skill-biased technological change can reconcile
the evidence presented in Frydman and Saks (2010) with the Gabaix and Landier (2008) model.
The dynamic model of Lustig, Syverson, and Van Nieuwerburgh (2011) is closely related to
this paper. Lustig et al. (2011) present a theoretical model that shows that optimal manager com-
pensation increases in response to technological shocks. My paper provides empirical evidence
that generally supports Lustig et al.s (2011) model, with the caveat that the causal force that
increases pay in this paper is skill-biased, and not general, technological change.
This paper also adds to the vast literature that explores the effects of human capital. Ciccone
and Papaioannou (2009) and Glaeser et al. (2004) are two among many papers that argue that
human capital is a key component of economic growth. Ciccone and Papaioannou (2009) sug-
gests that growth occurs most quickly in industries that rely on educated workforces. My results
suggest that growth is especially likely to occur in educated workforces that are led by a manager
that focuses on the human capital in the workforce. It is the combination of human capital withmanager focus that results in large synergistic gains.
Finally, this paper also adds to the growing literature that explores the connections between
labor and corporate finance. Agrawal and Matsa (2012) explores how the risk aversion of the
workforce affects firm leverage decisions, Pratt (2011) uses a structural model and Kim (2011)
6
8/14/2019 jkotter-jmp.pdf
8/67
uses Census establishment level data o explore how firm specific human capital affects lever-
age, Acharya, Baghai, and Subramanian (2012) shows that protecting employees from wrongful
discharge spurs innovation, and Brown and Matsa (2012) provides evidence that employees are
aware of the financial condition of firms and avoid seeking employment at firms with question-
able financial health. My paper broadly fits into this literature by showing that employees matter
for firm outcomes. To maximize firm value, shareholders must correctly motivate managers to
focus on the human capital of the firm.
The rest of the paper is structured as followed. Section I provides a brief theoretical motiva-
tion, Section II describes the data and important stylized facts concerning CEO pay and focus,
Section III describes my methodology, Section IV discusses the results, and Section V concludes.
Additional results are found in the Appendices.
I. Theoretical Motivation
This section introduces the theoretical intuition for the relationship between the role of the
CEO and the human capital of the workforce. A more formal model is developed and discussed
in Appendix B.
A. The Rise of Nonroutine Work
The computer revolution began in the early 1970s with the invention of the microprocessor,
read access memory (RAM), and the personal hard drive. A large literature in labor economics
shows that this shock was skill-biased in the sense that the technology increased the productivity
of high human capital workers relative to low human capital workers (Katz, Krueger, et al. (1998),
Bekman, Bound, and Machin (1998), Bresnahan, Brynjolfsson, and Hitt (2002), Autor, Levy, and
Murnane (2003)). To conceptualize the changes in the human capital of the workforce, I adopt
the task framework developed by Autor et al. (2003) and described in Table I.6 A workers job
is made up of a set of tasks (e.g., a professor teaches, researches, and grades exams). Tasks are
6This theoretical framework has been expanded by Autor, Katz, and Kearney (2006) and Acemoglu and Autor(2010).
7
8/14/2019 jkotter-jmp.pdf
9/67
defined as routine if they can be successfully completed by following explicit rules (e.g., grading
a multiple choice test). Nonroutine tasks require some degree of flexibility, innovation, and/or
interpersonal communication (e.g., research).
Within routine and nonroutine tasks, it is helpful to further subdivide tasks into cognitive
and manual. Computers are strongly complementary with nonroutine cognitive tasks such as re-
search (easy access to computer processing power increases the research productivity of a profes-
sor), but not particularly complementary with nonroutine manual tasks such as janitorial work.7
On the other hand, computers can substitute for both routine cognitive and routine manual work
(e.g., computers are capable of grading multiple choice tests and installing windshields on an
automobile assembly line).
[Table I about here.]
Under this framework, it is clear that a decrease in the price of computer technology causes
firms to substitute technology for routine workers. Since technology and nonroutine workers are
complementary, firms also demand more nonroutine workers. This shifts the composition of the
workforce toward high human capital labor. How does this compositional shift affect managers?
B. Managing Human Capital
CEOs perform many roles, but management scholars suggest that these roles can be summa-
rized in four major areas: managing people, operations, innovation, and external stakeholders
(Hart and Quinn (1993)). For simplicity, I focus on managing people and managing operations.
Managing people includes tasks such as hiring decisions, retention practices, creating a corporate
culture or vision, and personal interactions. Managing operations includes tasks focused on the
processes of production such as maintaining and deploying physical capital, managing supply
chains, and implementing procedures. Suppose that a CEO has a fixed amount of time to split
7This is true given the current capabilities of computers. It is entirely plausible that advances in technologies suchas Artificial Intelligence will allow computers to substitute for additional types of nonroutine labor in the future.
8
8/14/2019 jkotter-jmp.pdf
10/67
between effort on peoplehand effort on operationsf. Further assume that production,F(), is a
function of labor and CEO effort and that
F(n,h,f)
hn >0 andF(n,h,f)
fn =0, (1)
wherenrepresents the percentage of the labor force that is nonroutine. Equation 1 says that
the marginal benefit of managing people is higher for nonroutine workers and that the marginal
benefit of managing operations is independent of nonroutine workers.8 This implies a comple-
mentarity between managers and nonroutine workers, which I discuss below. Given Equation 1,
an increase in nonroutine workers has two effects: 1) it raises the marginal benefit of managing
people, which makes the CEO exert more effort on people and 2) it raises the cost of effort for the
CEO to manage operations (due to his fixed bank of time, see Holmstrom and Milgrom (1991))
which leads the CEO to exert less effort on operations. As long as the marginal gain from the
increased effort on people is greater than the marginal loss from reducing effort on operations,
the CEO will shift his focus to managing people. This is more likely to be the case as human
capital increases in importance to firm production.
This shift in focus depends on the complementarity between managers and nonroutine work-
ers. There are two types of complementarities that are likely to be important. First, managing
nonroutine workers might involve production complementarities. Many nonroutine tasks, such
as innovation and teamwork, are highly dependent on the working environment (Finkelstein,
Hambrick, and Cannella (2008)). By focusing effort on creating a positive working environment
(through retention and recognition processes and providing workers with appropriate flexibil-
ity) a CEO might significantly improve the innovation potential of nonroutine employees. While
similar focus might also increase the productivity of routine employees, the productivity gains
are likely to be much lower.
With a production complementarity, an increase in the human capital of the workforce in-
8Technically, all that is needed is that the increased benefit from focusing effort on one more nonroutine workeris greater than the increased benefit of focusing on operations when the company adds one more nonroutine worker.
9
8/14/2019 jkotter-jmp.pdf
11/67
creases the marginal product of the CEO. Then, if we consider a competitive market for CEO
talent such as Gabaix and Landier (2008) or Lustig et al. (2011), firms optimally increase the level
of pay. In essence, because the value of talented CEOs is highest at the firm with the skilled
workers (holding size constant), these firms pay more to attract talented CEOs.
The other type of complementarity that exists between managers and nonroutine workers is
cost based. Nonroutine work, by definition, is nebulous. Since the steps to succeed at a nonroutine
task are not clear, it is quite costly for a nonroutine worker to figure out how to exert effort. A
good manager can help clarify the task by narrowing the set of options that the worker needs to
consider; this reduces the cost of effort for the employee. This type of complementarity is not
likely to exist for routine tasks where the procedure is already clearly defined.
As described in Edmans et al. (2011), this cost based synergy also leads to an optimal in-
crease in CEO pay. The CEO makes his effort choice taking the effort of the employees as given.
Consequently, he ignores the effect that his own effort has on the employees cost of effort (and
subsequently, the employees effort choice). As a result, from the perspective of the shareholders,
the CEO exerts too little effort. To correct for this, shareholders subsidize the CEOs effort with
larger pay.
Both classes of models predict that an increase in nonroutine workers leads to an optimal
increase in CEO pay; however, the models do have a few different predictions. The production
complementarity model, combined with the market for CEO talent, predicts an increase in the
level of pay but has no clear prediction on the slope of pay. The cost complementarity framework,
in contrast, predicts an increase in both the level of pay and the power of the incentives. While
the cost complementarity framework predicts that the effect of nonroutine workers is constant
across firm size, the CEO talent framework suggests that the effect is increasing in firm size. Iam not able to clearly differentiate between these two models in the data, but I provide evidence
that suggests both types of complementarities exist.
10
8/14/2019 jkotter-jmp.pdf
12/67
II. Data and Stylized Facts
To test the effect of changes in the human capital of the workforce on CEO pay, I follow Autor
et al. (2003) to develop a time-varying industry level measure of workforce task composition. Here
I briefly review the most salient features of this measure; additional details of its construction are
available in the Data Appendix of Autor et al. (2003) and Acemoglu and Autor (2010).
The task composition measure is based on occupations. I use the Fourth (1977) Edition and
Revised Fourth (1991) edition of the U.S. Department of Labors Dictionary of Occupational Titles
(DOT) to classify the occupation in the Census along two task dimensionsroutine and nonrou-
tine task intensity. To define routine task intensity, I take the average of finger dexterity (routine
manual tasks) and set limits, tolerances, and standards (routine cognitive tasks). Nonroutine task
intensity is the average of math aptitude (analytical thinking) and direction, control, and plan-
ning (managerial and interpersonal tasks). These measures are ordinal rankings that range from
0 to 10. The DOT task intensities are based upon first-hand observations of workplaces using
guidelines produced by a panel of experts from the National Academy of Sciences.
To illustrate these tasks, consider the following examples. Textile production line workers
have high finger dexterity; clerks have high set limits, tolerances, and standards; computer pro-
grammers have high math aptitude; and sales people have high direction, control, and planning.
I match the DOT nonroutine and routine task intensities by occupation to each person in the
Combined Current Population Survey May and Outgoing Rotation Group samples (May/ORG
CPS) from 1973 to 2010. The CPS is a monthly survey of about 50,000 households administered
by the Census Bureau; it is designed to reflect the composition of the civilian non-institutional
U.S. population. To merge the DOT measure with the CPS, I create an occupational classification
that is consistent across the sample. Then using each employed worker aged 18 to 64 I create an
average task intensity by industry (k) for each task and year,
taskk,t =
iktaski,ti,thi,t
iki,thi,t. (2)
11
8/14/2019 jkotter-jmp.pdf
13/67
The average is weighted by the CPS weight, , and the number of hours worked,h, so that
the measure represents the average task intensity for a full time worker in industryk. To obtain
compatibility in census industry codes across years, I use the crosswalk developed by Autor et al.
(2003) to aggregate to 140 consistent census industry codes.
Although using the DOT provides the most-complete time series of job task requirements in
the US, there are several limitations. Importantly, while the occupation task measure is updated
in 1991, the majority of the time variation in my sample results from the changing composition
of occupations within industries. Undoubtedly, the change over time of skill requirements within
occupations is an important, omitted source of variation. Also, ideally I would like a measure of
the skill of the firms employees, but this variable only measures employee skill at the industry
level. These deficiencies likely reduce the precision of my analysis.
Figure 2 graphs the median of these measures across all occupations from 19732009. Since
the value of these indices has no intrinsic meaning, to make the magnitude of the changes inter-
pretable I scale the measures by the distribution of tasks in 1973. I choose 1973 as the base year
primarily because it was the earliest CPS data examined by Acemoglu and Autor (2010); how-
ever, it corresponds well with the beginning of the computer revolution and should reflect the
distribution of tasks before substantial computerization occurred. Industry routine task intensity
is flat throughout the 1970s, but steadily decreases beginning in the 1980s. Industry nonroutine
task intensity increases slowly during the early 1970s, and then increases rapidly toward the end
of the 1970s and early 1980s. By 2009, the median nonroutine task intensity is at the seventieth
percentile of the 1973 distribution. In sum, the nature of work has shifted to become much more
focused on nonroutine tasks.
[Figure 2 about here.]
This secular trend in employee skill fits well with the trend in CEO pay. CEO pay has in-
creased dramatically since the 1970s, whether measured in absolute or relative terms. Figure 3
shows that median real CEO compensation increased from about one million dollars in 1984 to
about three and half million dollars in 2010. Figure 3 also reveals that the evolution of nonroutine
12
8/14/2019 jkotter-jmp.pdf
14/67
task intensity follows a very similar trend. Given that reverse causality is not likely, this figure
suggests that either increasing employee skill explains part of the increase in CEO compensation
or some other driving force determines both employee skill and executive pay. The most plau-
sible forces that could affect both pay and employee skill levels are macroeconomic conditions.
Consequently, I control for business conditions, unemployment, and recessions in my analysis.
[Figure 3 about here.]
Frydman and Saks (2010) provide the most extensive evidence on executive compensation
prior to the 1980s. Using a hand collected sample of large, publicly traded firms, they show that
executive pay was relatively flat from 1936 to the early 1970s, when pay levels began to rise. Pay
levels grew throughout the 1970s, and exploded beginning in the mid 1980s. The ratio of median
executive pay to average worker pay also began to rise in the 1970s, but surprisingly declined for
the three decades prior to that.
The framework of this paper suggests a potential explanation for Frydman and Sakss (2010)
results. In 1971, Intel began marketing the Intel 4004the worlds first microprocessor. Two
years later, IBM introduced the first modern hard drive, known as the Winchester" drive. This
computer revolution marked the beginning of a seismic technological shift. Over the following
four decades, the price of computer capital plummeted.
Perhaps the most extreme example of falling prices is for storage space. In 1981, Morrow
Designs sold a 26 megabyte hard drive for $5,000 ($193,000 per gigabyte of storage). In 2010,
Western Digital sold a 1 terabyte hard drive for $71.42 ($0.08 per gigabyte of storage). Over the
course of 30 years, the price of storage had fallen by a factor of 2,412,500! Figure 4 shows this
decline in storage costs over time.
[Figure 4 about here.]
The fact that the computer revolution began at the same time as the trend break in CEO pay
presents a particularly parsimonious explanation for the growth rate in CEO pay. My analysis
13
8/14/2019 jkotter-jmp.pdf
15/67
utilizes the computer revolution as an exogenous shock to labor force to estimate the effect of
human capital on CEO pay. My analysis does not quantify the effect of the technology shock
itself on CEO pay. It is entirely plausible that technology also directly increases the level of CEO
pay; Kaplan and Rauh (2010) presents some evidence suggesting that this might be the case. My
difference in difference estimates eliminate this effect.
To proceed with the firm-level analysis, I develop a concordance to match the four-digit Com-
pustat historical SIC code to the Census industry code. I then match the employee skill measures
to each firm based on the firms Census industry. The quality of this match rests both on the
accuracy of the concordance and the extent to which the four-digit SIC code accurately reflects
the business activities of the firm. If a firm does not update its SIC code over time, the historical
SIC code might not accurately describe the current firm. Additionally, I use the primary SIC code
of the firm, which typically represents the industry of the segment of the firm with highest sales.
For large, diversified firms this might not be a very accurate representation of firm activity. Con-
sequently, this matching process adds noise to the employee skill variables. However, there is no
particularex antereason to believe that this noise biases the results in a particular direction.
My sample includes all CEOs in Execucomp from 1992 to 2010 and a sample of CEOs in large
publicly traded firms from 1984 to 1991 used in Yermack (1995). I merge this sample of CEOs with
firm-level accounting data from Compustat and stock return data from CRSP. The final sample
includes 3,157 firms from 1984 to 2010. To check if the measure of employee skill seems reason-
able, I sort firms into skill quintiles each year and then examine the industry distribution by skill
quintile. Table II shows the ten most frequently observed Census industries in the highest and
lowest employee skill quintiles. The industry task measure appears to accurately capture indus-
tries that are commonly viewed as skilled. Banks, pharmaceuticals, and computers are amongthe most frequently observed high skill industries, while eating places, steel works, and trucking
are among the most frequently observed low nonroutine task industries. The concentration of
nonroutine tasks among industries appears to be higher at the upper end of the distribution, as
the top 10 industries make up 85% of the sample of the highest quintile observations but only 39%
14
8/14/2019 jkotter-jmp.pdf
16/67
of the sample of the lowest quintile observations. Four of the top 10 (roughly 54%) upper quintile
industries are from the financial sector, suggesting that some of the large compensation packages
observed in this sector might be due to skill-biased technological change.
[Table II about here.]
As an additional robustness check, I run a regression of nonroutine skill on the average wage
for the firm. Wage data is only available for about one third of my sample. There is a positive
and statistically significant correlation between average wages and nonroutine skill. Since high
skill employees should receive high wages, this suggests that my measure of human capital is
reasonable.
Finally, I create a measure of manager focus using text analysis of 10-K reports. For each
year that a firm files a 10-K with Edgar, I extract the management discussion and analysis section
(MD&A). I randomly select 100 MD&A sections and carefully read them to classify content based
on four areas: people, operations, innovation, and external stakeholders. Management literature
suggests that the role of the CEO can be summarized in these four areas (Hart and Quinn (1993)).
I further divide external stakeholders into customers, competitors, and shareholders. Through
reading these randomly selected MD&As, I create lists of words that correspond to each of these
categories. For example, people related words include employee, staff, and labor. I further check
my list of words against other lists used in other financial text analysis (Li (2010), Loughran and
McDonald (2011)). For each category, I count the number of words from that category that are
used in the MD&A and scale each count by the total number of words in the MD&A. Figure 5
shows the average of these measures over time. Note that although the absolute level of oper-
ations words far exceeds people words, there is a clear increase in the focus on people and a
decrease in the focus on operations over time.
[Figure 5 about here.]
One concern is that these trends might simply reflect changes in the structure of the economy
(i.e., a switch from a manufacturing to a service economy). Figure 6 shows the evolution of people
15
8/14/2019 jkotter-jmp.pdf
17/67
words across 9 Fama-French Industries (I exclude the 10th industry, other). While the pattern
is stronger in some industries than others,
[Figure 6 about here.]
In sum, this paper builds on the following four important stylized facts.
1. The computer revolution, a massive technological shock, began in the early 1970s and con-
tinues to the present.
2. This shock changed the nature of work. Routine tasks were computerized and labor shifted
toward performing nonroutine tasks.
3. CEO pay, which had remained relatively stagnant for the 40 years prior to 1970, exploded.
4. Since 1994, managers have increased their focus on the people of their organization as
compared to the operations.
While the first three items are well known in the literature, this paper contributes the first
evidence of a shift in the focus of the executive management team.
III. Methodology
Consider the following model of CEO pay:
lnij t = t+nonroutinekt1+ Xij t1+ ij t, (3)
wherei indexes individual CEOs, j indexes firms,k indexes industries, andt indexes time. is
total real CEO compensation and I am interested in , the effect of nonroutine labor on CEO pay.
What is necessary to interpretas a causal effect? It must be the case thatijtis uncorrelated with
nonroutinekt1. Unfortunately, that is almost certainly not the case. In particular, ij tincludes all
characteristics of the CEO that we have been unable to control (e.g., ability). I expect that CEO
ability is correlated with the skill of the workforce, perhaps because high ability CEOs are better
able to attract high skilled workers. One strategy for dealing with unobserved CEO ability is
16
8/14/2019 jkotter-jmp.pdf
18/67
to include CEO fixed effects. In this case, though, CEO fixed effects likely does not solve the
problem. As technology changes, CEOs likely learn about how best to utilize both technology
and nonroutine employees. This learning will not be captured by a CEO fixed effect and is almost
certainly correlated with the level of nonroutine labor that the CEO chooses.
My solution is to implement a difference in difference estimator that mimics the natural ex-
periment of randomly assigning workforces of different human capital levels to CEOs. I take the
level of routine task intensity in an industry in 1973 as exogenous; this reflects the task composi-
tion of the labor force before most firms had access to personal computer technology. Firms that
are highly exposed to routine labor in 1973 experience a large shock to the composition of the
labor force as computers replace routine workers. In contrast, firms that have little exposure to
routine labor in 1973 experience small changes in the task composition of the workforce.
I estimate the following difference in difference model:
lnij t =1t+ 2routinek1973 + 3 t routinek1973+ Xij t1+ ij t, (4)
where3represents the causal effect of changes in nonroutine labor on CEO pay. This effect is
identified through variation in the workforce exposure to the computer revolution. Conceptually,
this approach compares the growth rate over time in CEO pay at a firm that was highly dependent
on routine labor before the computer revolution (e.g., a manufacturing or finance firm) to the
growth rate of CEO pay at a firm that did not depend on routine labor (e.g., a pharmaceutical
firm or a hair salon). This estimated effect can be interpreted in a causal sense as long as there is
nothing that systematically affects both CEO pay and the probability of having high routine labor
exposure before the computer revolution. Given the variety of industries that form my treatment
group, this seems unlikely to be the case. In addition, I include industry fixed effects and CEO
fixed effects in some specifications. Note that identification of this model with CEO fixed effects
relies on CEOs moving across firms over time.
The following control variables are included inXijkt. We know that pay is increasing in firm
17
8/14/2019 jkotter-jmp.pdf
19/67
size; I include the natural logarithm of firm revenue ln(Revenue)to control for this relationship.
I includeTobins Qto capture the effect of growth opportunities on CEO pay. Income to assetsis
measured as earnings before interest and taxes divided by total assets. Shareholder returnis the
previous fiscal-year cumulative return on the stock. I expect CEO pay to be positively related
to these measures, since shareholders want to incentivize good performance. Std. dev. return
is the standard deviation of daily stock returns calculated over the previous fiscal year, and is a
proxy for the riskiness of the firm. Since CEOs are risk averse, they require higher compensation
for managing risky firms; consequently, pay should be positively related to Std. dev. return.
Age is the executives age, and Tenureis the length of time that the executive has worked for
her current firm. The model predicts that pay is positively related to these variables because
CEOs become more effective with experience. Executive-level information is from Execucomp or
Yermack (1995), firm accounting information is from Compustat, and stock returns are taken from
CRSP. To control for possible macroeconomic factors that might simultaneously determine both
CEO pay and employee skill, I use three variables. Recessionis a dummy variable equal to one if
the end of the fiscal year occurred during a recession as classified by the NBER. Unemployment
is the unemployment rate obtained from the Bureau of Labor statistics. Business conditionsis an
index published by the Federal Reserve that incorporates several factors including GDP, interest
rates, and stock market returns. Higher values of this index represent conditions that are more
favorable to businesses. I expect that CEO pay is procyclical, so it should be positively related to
Business conditions, and negatively related torecessionand unemployment.
I match CEO pay in fiscal year tto all other variables measured at year t 1 This ensures that
the firm and macroeconomic variables are known at the time when the CEO contract is finalized.
All nominal quantities are converted to millions of 2005 dollars using the GDP deflator of theBureau of Economic Analysis. Continuous variables are winsorized at the 1% level. The variables
used in this study are summarized in Table III.
[Table III about here.]
Equation 4 suggests one other problem with this estimation. The independent variable of
18
8/14/2019 jkotter-jmp.pdf
20/67
interest, the level of nonroutine tasks in 1973, is measured at the industry level. The dependent
variable, CEO pay, is measured at the firm level. This means that errors in measuring nonroutine
tasks will be correlated across industries. In addition, this correlation is likely to be time varying.
As noted previously, the nonroutine task measure captures task changes due to shifts in occupa-
tions within an industry, but it does not measure shifts in within occupations. This implies that
the task measure gets noisier as time goes in. To address the time varying industry correlation
in the error term, I cluster the standard errors by the interaction of industry and year.
IV. Results
A. Univariate results
I first look at the univariate properties of executive compensation and employee skill. Table IV
compares the mean of key variables used in this paper by high and low nonroutine task inten-
sity firms. In each year, I sort forms into quintiles based on employee skill. High skill firms are
defined as firms in the top two quintiles, while low skill firms are defined as firms in the bottom
two quintiles. I compare the means and report the two-sample t-test on their difference. Table IV
shows that low skill firms have 28 log points lower mean expected pay and a ratio of relative paythat is nine times less than high skill firms, and these mean differences are significant at the 1%
level. Low skill firms are also smaller than high skill firms, with mean revenue about 12% less.
Somewhat surprisingly, low skill and high skill firms look fairly similar along other dimensions.
Although the mean differences are statistically significant for several variables, there is little to no
economically significant difference between high and low skill firms in profitability, shareholder
return, risk, and CEO age and tenure. Table IV does reveal some differences in governance envi-
ronment, but it does not appear that high skill firms have systematically worse governance than
low skill firms.
[Table IV about here.]
19
8/14/2019 jkotter-jmp.pdf
21/67
It is well known that CEO pay varies with firm size, and Table IV suggests that high skill firms
are significantly larger than low skill firms. I explicitly control for firm size in the multivariate
results, but to get a clearer sense of the univariate differences in CEO pay between firms of differ-
ent skill levels, I employ the following procedure. In each year from 1984 to 2009, I first sort firms
into quintiles based upon revenue. Within each size quintile, I then sort firms into quintiles based
upon nonroutine task intensity. This double sort helps account for the fact that larger firms pay
their managers more. I then compare the median CEO pay across each bin. The results of this
procedure are shown in Figure 7. Across each size quintile, executive compensation is increasing
in nonroutine task intensity. Moving from the lowest skill quintile to the highest skill quintile
raises median pay by about $400,000 (about 33%) for the smallest firms, but raises median pay
by about $6 million (about 48%) for the largest firms. This positive interaction between nonrou-
tine labor, firm size, and CEO pay is consistent with theoretical predictions of production-based
complementarities between CEOs and nonroutine workers.
[Figure 7 about here.]
How does the difference in CEO pay vary over time? Using the same double sort procedure
as described above to control for size effects, I define high skill firms as firms in the top two skill
quintiles and low skill firms as firms in the bottom two skill quintiles. The double sort procedure
ensures that the size composition of high and low skill firms is roughly equivalent. I calculate the
difference in median CEO pay between high and low skill firms for each year between 1984 and
2010 and graph the result in Fig. 8. Throughout the 1980s, there is no difference in pay between
high and low skill firms. Beginning in the early 1990s, the difference in median pay increases.
At the peak difference in 2000, high skill firms pay their CEO $1.7 million more than low skill
firms. During the rest of the 2000s, the difference oscillates around $1 million. Fig. 8 also shows
the difference in medianemployee skillbetween high and low skill firms. The difference in skill
follows a similar pattern, suggesting that skill-biased technological change might explain the
stylized fact that the dispersion in CEO pay has increased over the past three decades.
20
8/14/2019 jkotter-jmp.pdf
22/67
[Figure 8 about here.]
The difference in difference approach described in Section III also lends itself to nonparametric
estimates. To estimate the effect of nonroutine labor on CEO pay, I split firms into treatment and
control groups based on the nonroutine worker intensity as of 1973. Firms in the top quartile
of routine tasks as of May 1973 form my treatment group, while firms in the bottom quartile of
routine tasks as of May 1973 make up my control group. I define the treatment period as post
1995. This is a convenient breaking point both because it is the approximate midpoint of the
sample and because the commercial version of the internet came online in 1995. Consequently,
firms experienced an additional large technology shock after 1995. The definition of the treatment
period is admittedly ad hoc; however, the estimates are robust to choosing alternative treatment
periods such as the first five years versus the last five years of the sample or the 1980s versus the
2000s.
Panel A of Table V reports the simple difference in difference estimate. Before 1995, on average
high routine firms paid their CEOs roughly $650,000 less than low routine firms. Pay grew much
faster at high routine firms than low routine firms, though, and post 1995 CEOs at low routine
firms actually made on average $200,000 more. The difference in difference estimate implies that
increases in human capital of the workforce increased CEO pay by $859,000; this difference is
statistically significant at the 1% level.
[Table V about here.]
The obvious problem with this simple difference in difference estimate is that it does not adjust
for differences in firm characteristics. In particular, Table IV shows that there is a significant
difference in size across treatment and control groups. Given that CEO pay is positively correlated
with firm size and that the treatment group of firms is systematically smaller than the control
group, it seems likely that the simple difference in difference estimate is an underestimate. I adjust
for this using a semi non-parametric kernel matching difference in difference estimate (Heckman,
Ichimura, and Todd (1998)). I first estimate the propensity score of being in the treatment group
21
8/14/2019 jkotter-jmp.pdf
23/67
based on log revenue, TobinsQ, income to assets, shareholder return, standard deviation of stock
returns, beta, CEO tenure and age. For each firm in the treatment group, I then choose a match
from the treatment group based on the propensity score weighted using the epanechnikov kernel.
I only match firms on the common support of the estimated propensity score.
Panel B of Table V reports the results of this estimator. Controlling for firm characteristics in
this way increases the estimate of the difference in difference effect to about $1.1 million. Mean
CEO pay increased by about $3.7 million dollars from 1984 to 2010, so the effect of changes in
human capital on CEO pay explains about 30% of the average increase in executive pay.
While this evidence is suggestive, it does not fully control for other differences between high
and low skill firms that might influence executive pay, nor does it control for global factors that
might simultaneously affect pay and employee skill. To control for these differences, I proceed to
the multivariate analysis outlined in Section III.
B. Multivariate results
Throughout this section, I estimate all models using both CEO expected pay (Execucomp
TDC1) and realized pay (Execucomp TDC2). The results are quantitatively and qualitatively sim-
ilar, so to conserve space I present estimates using only CEO expected pay.
Table VI presents estimates of Equation 4. This is the multivariable, continuous version of the
simple difference in difference estimates presented in the previous section. One benefit of this
approach as compared to the simple difference in difference estimate is that I do not need to take a
stand on the cutoff between treatment and control groups. Instead, I use the continuous variable
Original Routine Tasksas a proxy for the intensity of treatment. Higher values of this variable
represent more exposure to routine task workers in 1973, which implies that the technologyshock had a larger effect on the workers of the firm. As a result, the intensity of the treatment
varies positively with this variable. Another benefit is that I do not need to specify a pre and
post treatment period. Instead, I include a time trend so that the effect of the technology shock
grows across time. The difference in difference estimate is represented by the coefficient on the
22
8/14/2019 jkotter-jmp.pdf
24/67
interaction of the time trend and Original Routine Tasks. The estimate on this coefficient is 0.46
and is statistically significant at the 1% level. This estimate implies that for a firm at the average
level of routine intensity in 1973, the increase in human capital of the workforce caused CEO pay
to increase by 84%.9
[Table VI about here.]
Column 2 of Table VI estimates a more stringent version of Equation 4 that includes CEO
fixed effects. Given theoriginal routine tasksdoes not vary within a firm, the identification of the
difference in difference estimate in Column 2 comes from comparing the pay changes of CEOs
that switch firms. By controlling for unobserved CEO characteristics, the CEO fixed effect speci-
fication strengthens the causal interpretation of the estimate. The estimate should be interpreted
with caution, though, since the number of individuals that work as CEO for two separate firms
in my sample is quite small (around 234). With that caveat in place, the CEO fixed effects esti-
mator increases the difference in difference estimate to 0.68 (p-value
8/14/2019 jkotter-jmp.pdf
25/67
An alternative specification to estimate the regression difference in difference estimator is to
use year dummy variables rather than a time trend, that is to estimate
lnij t =
2010i=1984
2010j=1984
dij+2routinek1973+
2010i=1984
2010j=1984
dij routinek1973 + Xij t1+ ij t, (5)
wheredij =1 ifi =jand 0 otherwise. This places less structure on the way that the technology
shock propagates through time, but comes at the cost of making it more difficult to interpret
the economic magnitudes of the effect. The difference in difference estimates for each year are
contained in thevector. It is easiest to interpret these estimates graphically, Figure 9 graphs
along with the 95% confidence interval for each year from 1985 to 2010. The estimated effect of
nonroutine workers on CEO pay is around zero (or slightly negative) until 1994. After 1994, the
effect grows until around 2000 and then it effect stays relatively constant for the remainder of the
sample. This picture is consistent with the internet shock that began in 1994 and 1995 with the
commercialization of the web.
[Figure 9 about here.]
There is nothing mechanical in the estimation strategy employed in Table VI that ensures
firms actually experience a change in workforce composition. Rather, I am identifying off of the
potential for workforce change. If the effects reported above are really identified through the
channel I have proposed, I should only see effects for firms that actually experienced an increase
in human capital, i.e. firms that increased their nonroutine worker intensity. Table VII tests this
by re-estimating column 2 of Table VI for various subsets of firms. Column 1 limits my sample to
firms that actually increased their nonroutine task intensity from the beginning of the sample to
the end of the sample and that were also not always either in the lowest quartile or the highest
quartile of nonroutine task use. As expected, the difference in difference for this group of firms is
positive and similar in magnitude to the estimates in Table VI. The estimate is less precise because
the variation in the sample went down, but the result is broadly consistent with the human capital
24
8/14/2019 jkotter-jmp.pdf
26/67
channel. Column 2 repeats this exercise for firms that started and ended the sample in the top
quartile of nonroutine task use. These firms always had a lot of human capital, so if the results
are driven by changes in human capital our estimated effect should disappear. If, however, the
results are driven by technology (i.e. technology increases the scale of the CEOs effort which leads
to higher pay) the results should apply equally well to this group of firms. Consistent with the
human capital channel, the difference in difference estimate in Column 2 is 0.26 and statistically
indistinguishable from zero.
Columns 3 and 4 of Table VII perform a similar exercise of high tech firms and excluding
high tech firms. The technology shock the underlies my identification is plausibly exogenous to
most firms, but might be endogenous to high tech firms. Excluding these firms strengthens my
estimate, the estimate in Column 4 implies that the increase in nonroutine workers over the last
three decades caused CEO pay to increase by 285%. If we estimate the difference in difference
model with only high tech firms, the effect of human capital is again indistinguishable from
zero. This is consistent with the nonroutine worker channel, since high tech firms did not likely
experience much change in the composition of their workforce.
Taken together, these results suggest that the estimates in Table VI represent the causal effect
of changes in workforce nonroutine task intensity on executive pay.
[Table VII about here.]
One of the strongest empirical results in the literature, at least for data after 1970, is the pos-
itive relation between CEO pay and firm size (Gabaix and Landier (2008)). To help ensure that
firm size is not driving my main result, I re-estimate the main difference in difference specifica-
tion found in column 2 of Table VI including dummy variables for the firm size quintile. I interact
the difference in difference estimate with these dummy variables and plot the resulting estimates
in Figure 10. There is some evidence that the effect of human capital on CEO pay increases with
firm size. That is not unexpected; the CEO talent framework discussed in Section I suggests a
positive relation between size, employee skill, and CEO pay if their are production complemen-
tarities between CEOs and nonroutine workers. While Figure 10 does provide some evidence of
25
8/14/2019 jkotter-jmp.pdf
27/67
production complementarities, more importantly, it also reveals a strong effect of human capital
on pay across all firm size groups. The results in this paper do not appear to be driven by firm
size.
[Figure 10 about here.]
The difference in difference methodology used in this paper provides a useful setting to iden-
tify causal effects; however, it suffers from the fact that the measure of employee skill is fixed for
each firm. To take advantage of the heterogeneity in workforce skill across time, I implement a
two-stage least squares model. Autor et al. (2003) provide evidence that the fall in the price of
computer technology caused firms to replace routine workers with computers and hire additional
nonroutine workers. Motivated by that result, I use unanticipated computer price shocks as an
instrument for the level of nonroutine task intensity.
Specifically, I use the average retail cost of one mebibyte (1,048,576 bytes) of computer RAM
for my measure of computer prices.11 I choose this for my computer price series both because it is
available for the entire time series and because Intel released the first DRAM chip in 1970, which
corresponds to the start of the computer revolution. Of course, there is a strong downward time
trend in the price of RAM. To alleviate the effect of the time trend driving any results, I hindcast
a Moores Law model of computer prices. For each year t, I use data on RAM prices from 1970
to t-1 to estimate Moores Law. I then use this estimation to predict the price in t+1. I take the
difference between the actual price in year t+1 and my estimate as the RAM price shock. The
time series of RAM price shocks has no clear time trend.
In addition to RAM price shocks, I use the original level of routine tasks as an additional
instrument. For my identification to be valid, these instruments must be correlated with the level
of nonroutine task intensity. Table VIII Column 1 shows the first stage estimates. The RAM price
shock is negatively correlated with nonroutine labor and is highly significant. An unexpected
fall in computer prices (a negative shock) does lead to an increase in human capital. As expected,
industries with a high level of original routine task have a lower current level of nonroutine labor.
11This data is from John C. McCallum, and is available athttp://www.jcmit.com/memoryprice.htm.
26
8/14/2019 jkotter-jmp.pdf
28/67
This effect is not statistically significant, but the F-statistic on the combined instruments is 23,
suggesting that weak instruments are unlikely to be a problem.
For these instruments to be valid, they must also pass the exclusion criteriathat is, they
should be uncorrelated with CEO pay except through changes in employee skill. The original
level of routine tasks is not likely to be a problem; the estimates in Table VII suggest that changes
in CEO pay due estimated from the original level of routine tasks likely occur due to changes in
employee skill. Changes in computer prices are somewhat more problematic. While this variable
is plausibly exogenous to CEO pay, it is possible that changes in technology directly increase
the productivity of the CEO which then leads to higher pay irrespective of changes in the labor
force. As a result, these instrumental variable estimates are likely biased upward and should be
interpreted as an upper bound on the effect of nonroutine labor on CEO pay.
Table VIII reports the coefficients from the 2SLS estimate of
lnijkt = i+ kt1+ Xijkt1+ ,
whereijktis total pay for CEO iat firm j in industrykat timet, kt1 isNonroutine tasks
of industrykin the prior fiscal year, and other covariates are measured as of the previous fiscal
year. I instrument for Nonroutine tasksusing Original Routine Tasksand RAM price shock. The first
stage regression is shown in Column 1. Columns 2 and 3 show the second stage regression with
and without industry fixed effects. Note that becauseRAM price shockdoes not vary across firms
within a given year, it is not possible to include year fixed effects. Instead, I include half-decade
fixed effects. The results are similar if I do not include time effects and instead include macro
variable trends such as unemployment, GDP per capita, and a recession dummy variable. The
2SLS estimate of the effect of human capital on CEO pay in Column 2 is 6.23 and is significant at
the 5% level. This estimate implies that the average change in nonroutine skill from 1984 to 2010
caused CEO pay to double. Adding industry fixed effects increases the magnitude and statistical
significance of the estimate. Although identified in a completely different way, the magnitude
and statistical significance of these estimates is strikingly similar to the difference in difference
27
8/14/2019 jkotter-jmp.pdf
29/67
estimates presented earlier.
[Table VIII about here.]
The 2SLS approach also allows me to comment on the types of nonroutine tasks drive in-
creases in CEO pay. Using a recently updated database of occupational characteristics, 0*NET,
I can further classify nonroutine tasks into analytic tasks and interpersonal tasks. I re-estimate
the 2SLS regression from Table VIII using these two measures instead of nonroutine tasks. The
results are reported in Section C. All of the estimated effect appears to come from changes in in-
terpersonal employers skill. This seems broadly consistent with the source of synergies between
managers and employees relying on increased manager focus on employees.
Does managing human capital change the incentive structure for CEOs? The market for CEO
Talent theory suggests that the level of pay increases with production complementarities be-
tween managers and skilled labor, but is silent on incentive structures (See Section B). Edmans
et al. (2011) suggest that both the level of pay and the power of incentives increase with effort
cost complementarities between managers and workers. Table IX presents difference in differ-
ence estimates for various measures of CEO incentives. While column 1 shows that nonroutine
labor increases the cash component of CEO pay, column 2 reveals that the effect doubles for op-tions. These effects are both significant at the 5% level; they suggest that increases nonroutine
workers in the last thirty years have doubled cash pay, but quadrupled option pay. There is some
evidence that nonroutine labor changes the ownership of CEOs. Column 3 shows a positive, but
marginally significant estimated effect. The effect is large, though, as it implies that changes in
nonroutine have increased CEO pay by nearly 5%. Finally, Column 4 shows that managing hu-
man capital reduces tenure. The estimate implies that changes in nonroutine task workers have
caused a 3 year decline in average CEO tenure, though the estimated effect is only marginally
statistically significant. As a whole, the evidence suggests that managing nonroutine labor in-
creases the power of CEO incentives. This is inconsistent with manager rent extraction theories
and consistent with an increased importance of the role of the CEO. It also suggests that there
might be cost complementarities between skilled workers and managers.
28
8/14/2019 jkotter-jmp.pdf
30/67
[Table IX about here.]
The identification structure used in this paper makes it unlikely that the results are due to
manager power. As an additional check, I repeat the difference in difference and 2SLS estimates
while controlling for various proxies of governance and agency problems. The results are quan-
titatively and qualitatively similar and are reported in reported in Appendix C.
The evidence on CEO pay is consistent with synergies between managers and skilled employ-
ees. As additional evidence, I use the text-based measure of management focus. This measure is
created through a text analysis of the MD&A section of the firms 10-K and is designed to reflect
the percentage of section that is focused on people within the firm. Section II shows that firms
have increased their people focus over time. Now, I examine how employee skill affects manager
focus and how the combination affects firm value.
Table X Columns 1 and 2 show the results of a linear regression of management focus on the
prior year level of nonroutine tasks. Both specifications include year fixed effects, industry fixed
effects, and the firm-level control variables used in Table VI. In addition, Column 2 includes CEO
fixed effects. The magnitude of both estimates is similar, though the CEO fixed effect estimate is
less precise. These estimates imply that a one standard deviation in nonroutine tasks leads to a
9-11% increase in people focus. Caution is warranted in interpreting these as causal estimates,
since the CEO has at least some impact on both the focus of the firm and the composition of the
workforce. CEO fixed effects help alleviate this problem, since identification comes from CEOs
that switch firms and then choose a focus based on the existing labor force of the new firm. Even
still, CEOs choose whether or not to take the new job, and that choice might be influenced by the
workforce of the new firm, so these results are best interpreted as strong correlations.
[Table X about here.]
Columns 3 and 4 provide evidence that focusing on people increases firm profitability and
value, but only in firms that have high human capital workforces. I estimate a linear regression
of total firm value and return on assets on people focus, nonroutine tasks, and the interaction
29
8/14/2019 jkotter-jmp.pdf
31/67
between people focus and nonroutine tasks. In both specifications I include CEO fixed effects. The
interaction term reflects the value of management focus on people in highly skilled firms. Column
3 shows that the effect of people focus on profitability is negative. However, the interaction term
is positive and marginally statistically significant, suggesting people focus improves profitability
in high human capital firms. The effect is economically large; evaluated at the mean levels of
people focus and nonroutine task, the coefficient implies a 1% increase in ROA which is large
relative to the unconditional mean of 8.5%. Column 4 shows a similar relationship between people
focus, nonroutine tasks, and total firm value. The coefficient on the interaction term is positive,
statistically significant at the 1% level, and economically meaningful. Again evaluated at the mean
of nonroutine tasks, a one standard deviation increase in people focus implies a 6% increase in
firm value. These estimates suggest both that there are important synergies between managers
and nonroutine workers and that shareholders have strong incentives to ensure managers to
focus on these synergies.
If CEO focus on high skilled labor truly has the potential to increase shareholder value as much
as the above estimates imply, models of the market for CEO talent suggest that high human capital
firms should hire the most talented CEOs (Gabaix and Landier (2008)). To test this prediction, I
estimate a CEO fixed effects regression of CEO pay on on industry fixed effects (at the Fama-
French 17 industry level), year fixed effects, and the natural logarithm of firm sales. Row 1 of
Table XI reports the F-test for the significance of the overall CEO fixed effects. Not surprisingly,
CEO fixed effects matter for CEO pay (the Fstatistic is 6.45 and significant at 1% level). In this
context, the CEO fixed effect can be interpreted as the talent or ability of the CEO. With that in
mind, Column 2 and 3 report the mean CEO fixed effect for firms in the bottom and top quartile of
nonroutine task intensity. Column 4 calculates reports the difference in means and the performsa two-sidedttest to determine if this difference is statistically different from zero. Interestingly,
high nonroutine task intensity firms hire high fixed effect CEOs, which is consistent with high
ability CEOs matching with high human capital firms. The difference in fixed effects implies that
CEO ability accounts for 20% of the difference between CEO pay at high and low human capital
30
8/14/2019 jkotter-jmp.pdf
32/67
firms, this difference is both economically large and statistically significant.
In this context, I can statistically estimate the fixed effect even if the CEO does not change
firms; the coefficient is identified off of variation in CEO pay within the CEOs tenure at a firm.
To be conservative, though, I limit the sample to the relatively small number of individuals in
my sample that work as CEO in multiple firms across time. For these 234 CEOs, fixed effects are
primarily identified by the CEO moving to a new firm. Using this sample, the estimated effect
actually increasesthe fixed effect of the average high skill firm CEO implies a compensation
level more than double the CEO of a low skill firm.
[Table XI about here.]
Rows 2 and 3 then estimate similar fixed effect regressions for firm profitability and stock re-
turns. In addition to the controls included for the CEO pay regression, the profitability regression
includes Tobins Q. The stock returns regression includes industry and year fixed effects and the
return on the market (value weighted CRSP return) so that the fixed effect represents industry ad-
justed excess returns. TheF-statistic indicates that CEO fixed effects matter for both profitability
and returns. In both cases, high human capital firms employ CEOs with higher fixed effects and
the difference is statistically significant from zero. The differences are large: using the movers
only sample, the difference in fixed effects explains a 6% increase in ROA and a 14% increase in
stock returns.
As a whole Table X and Table XI imply that there is a significant synergy between CEOs
and nonroutine labor. Talented CEOs that focus on this synergy can provide large returns to
shareholders. These potential synergy gains rationalize increased CEO pay.
V. Conclusion
At least since Schultz (1961) and Becker (1962), academic economists struggled to quantify the
effect of human capital. Much of the existing research focuses on the relationship between school-
ing and human capital accumulation or the role of human capital in explaining macroeconomic
31
8/14/2019 jkotter-jmp.pdf
33/67
growth. More recently, the finance literature has started to explore the value of employees as key
assets of the firm Edmans (2011), Rajan and Zingales (1998), Berk, Stanton, and Zechner (2010).
This paper examines how managers enhance the value of high human capital; consequently, this
work provides insight on one particular channel through which employees add value to the firm.
I show that the increase in human capital induced by the computer revolution fundamentally
alters the role of the CEO. The overall increase in human capital over the last two decades has
been accompanied by a shift in the focus of executive mangers away from the operations of the
firm and towards the people of the firm. This shift in focus is particularly strong for managers of
high human capital workforces, and the shift for these types of managers results in a significant
increase in firm value.
This evidence is consistent with large synergies between CEOs and nonroutine workers. Con-
sistent with Edmans et al. (2011), shareholders increase the pay of CEOs to induce managers to
focus on these synergies. These synergies appear to be large; talented CEOs at high human cap-
ital firms raise total firm value from between 6 to 14%. As a result, the growth in CEO pay at
nonroutine task intense firms can be justified by the increased value that comes from managing
human capital.
The substantial growth in CEO pay since the 1970s has led to significant academic interest
in the question of whether or not CEO compensation contracts are optimally set. While the
evidence in favor of managerial power is difficult to reconcile with the growth in CEO pay, the
empirical literature on optimal CEO contracts also fails to explain the sudden increase of CEO
pay in the 1970s after several decades of stagnant growth. Skill-biased technological change has
the potential to reconcile these differences. This paper provides the first direct empirical evidence
that some of the growth in executive pay is due to skill-biased technological change. I estimatethat the computer revolution led to an increase in nonroutine employees that that approximately
doubled the level of CEO pay and explains around half of the actual increase in CEO pay over the
last three decades. These results are robust to controlling for the firms governance environment,
and suggest that a substantial portion of the increase in CEO pay over the past three decades
32
8/14/2019 jkotter-jmp.pdf
34/67
represents an optimal response to skill-biased technological change.
33
8/14/2019 jkotter-jmp.pdf
35/67
Appendix A. Variable Definitions
Table AI
Variable Definitions. This appendix defines the variables used throughout the paper.
Variable Definition
ln(CEO pay) Natural log of the dollar value of salary, bonus, restricted stock granted, long-termincentive payouts, and the Black-Scholes value of stock-options granted. Source: Ex-ecucomp and Yermack (1995).
CEO Pay Sum of salary, bonus, restricted stock granted, long-term incentive payouts, and theBlack-Scholes value of stock-options granted (TDC1). Source: Execucomp and Yer-mack (1995).
Cash salary Sum of the CEOs salary and bonus.Source: Execucomp and Yermack (1995).Options The Black-Scholes value of stock options granted during the fiscal year.Source: Ex-
ecucomp and Yermack (1995).ln(Revenue) The natural logarithm of firm revenue (SALE).Source: Compustat.
TobinsQ The market value of equity (CSHO*PRCC_F) plus the book value of debt(DLTT+DLC+PSTKRV) minus the value of financial assets (CHE+RECT+ACO) di-vided by the total value of assets (AT) less financial assets.Source: Compustat.
Income to assets Income (INCOME) divided by total assets (AT).Source: Compustat.Shareholder return Fiscal-year cumulative return on the stock.Source: CRSP.Std. dev. return The standard deviation of daily stock returns calculated over the fiscal year.Source:
CRSP.Beta The firms CAPM beta calculated using the previous year of stock returns. Source:
CRSP.CEO Tenure The length of time, measured in years, that the executive has worked as CEO for her
current firm.Source: Execucomp and Yermack (1995).Age The executives age, measured in years.Source: Execucomp and Yermack (1995).Original Routine Tasks The industry intensity of routine task occupations as of 1973.Source: Current Popu-
lation Survey Merged Outgoing Rotation Group (CPS) and US Department of LaborsDictionary of Occupational Titles (DOT).
Nonroutine tasks The industry intensity of nonroutine task occupations measured as of May of thegiven year. This variable is transformed into percentile values corresponding to itsrank in the 1973 distribution of nonroutine tasks. Source: CPS and DOT.
Analytical skill The industry intensity of nonroutine tasks that require analytical skill to complete,measured as of May of the given year. This variable is transformed into percentilevalues corresponding to its rank in the 1973 distribution of analytical skill. Source:CPS and US Department of Labors O*NET (O*NET).
Interpersonal skill The industry intensity of nonroutine tasks that require interpersonal skill to com-plete, measured as of May of the given year. This variable is transformed into per-centile values corresponding to its rank in the 1973 distribution of interpersonal skill.Source: CPS and O*NET.
RAM price shock The unexpected change in annual computer RAM prices. For each year, prices arehind-casted using Moores Law and price data beginning in 1950. The price shock ismeasured as the regression error.Source:
Ownership The percent of firm equity owned by the CEO.Source: Execucomp and Yermack(1995).
Board Size The number of directors on the firms Board. Source: Thomson Reuters ownershipdatabase.
Continued on next page
34
8/14/2019 jkotter-jmp.pdf
36/67
Table AI Continued
Variable Definition
Pct. Indep. Directors The percentage of the firms directors that are classified as independent (i.e., neitherinside nor grey directors).Source: Thomson Reuters ownership database.
CEO is Chair A dummy variable equal to one if the CEO is also the chair of the board of directors.
Source: Execucomp and Yermack (1995).Institutional Ownership The percent of firm equity owned by institutional investors as reported on Form 13F.Source: Thomson Reuters ownership database.
Appendix B. Theoretical Model
In this paper, I define skill-biased technological change as the computer revolution described
in Section ??. I take the price of computer capital as exogenous to the firm. To understand the
connection between falling computer prices and executive compensation, I use the task frame-
work developed by Autor et al. (2003) combined with the CEO pay model laid out in Gabaix and
Landier (2008). I assume that production consists of a combination of four types of tasks: rou-
tine cognitive, routine manual, nonroutine cognitive, and nonroutine manual. Tasks are defined
as routine if they can be accomplished by an exhaustive set of programmable rules. Assembly
line work is an example of a routine manual task, while balancing a firms ledger is an exam-
ple of a routine cognitive task. Nonroutine tasks, in contrast, do not have an exhaustive set of
well-defined rules. An example of a nonroutine manual task is delivering packages, while an
example of a nonroutine cognitive task is designing a new vaccine. The nature of routine tasks
makes them particularly suited to be performed by a computer, while nonroutine tasks are not
easily completed by current levels of computer technology. This task framework is summarized
in Table I, which is reproduced from Autor et al. (2003).
This basic framework suggests that computers substitute for workers that perform routine
cognitive and manual tasks, but complement workers that perform nonroutine cognitive tasks.
For my purposes, I assume that nonroutine manual tasks and computers are neither strong sub-
stitutes nor compliments.12 I combine this framework with the exogenous decline in the price of
12Current technology is not yet advanced enough to substitute for most nonroutine manual tasks, though for atleast some tasks it is moving in that direction (e.g., computer-driven cars). Further, the extent to which computerscan complement manual tasks is naturally limited by the physical limitations of human employees (e.g., even with aG.P.S. system, a single UPS driver can only deliver so many packages in a day.
35
8/14/2019 jkotter-jmp.pdf
37/67
computers to create quasi-exogenous changes in executive compensation.
To understand the effect of falling computer prices on CEO pay, I introduce a simplified model.
I assume an aggregate production function of the form,
Q =F(n,r,c)(1 + A T),
F(n,r,c) = (r+ c)1n, (0,1),
A =F(n,), (B1)
whereTis the talent of the manager, Ais the firms organizational capital, rand nare routine
and nonroutine labor inputs, and c is computer capital. All inputs are measured in efficiency
units.F(n,r,c)is a Cobb-Douglas production function as in Autor et al. (2003). 13
The firms organizational capital,A, quantifies the effect of CEO talent on production. Orga-
nizational capital is a function of nonroutine laborn, i.e. the skill level of the firms workforce,
and CEO specific traits,, such as age, experience, and education. I assume that Ais increasing
in ; that is, I assume that some of the skills that make an effective CEO can be learned. I also
assume that
A
n >0. (B2)
Eq. B2 is the key assumption of the model. In words, there are positive synergies between
managers and nonroutine task employees. This can be viewed as a reduced form way of mod-
eling the effect of CEO effort on employees; the assumption implies that CEO effort increases
productivity (or reduces the cost of effort) of nonroutine employees more than routine employ-
ees. Since the validity of my empirical results rests on Eq. B2, it is important to carefully consider
the plausibility of this assumption. Why should CEO effort matter more to skilled employees? I
13Eq. B1 can be modified as in Gabaix and Landier (2008) to allow decreasing returns to scale to manager produc-tivity, i.e.Q =F(n,r,c)+ F(n,r,c) A T.is what Lustig et al. (2011) refer to as the span of control parameterof the manager; if CEOs have less effect on big firms than small firms, then
8/14/2019 jkotter-jmp.pdf
38/67
argue that the role of a manager is fundamentally different when managing routine tasks versus
nonroutine tasks. As a manager of routine tasks, the CEO is essentially the colonel of the firm
giving orders and ensuring that these orders are followed through. Managing nonroutine tasks
this way, though, is inefficient. Instead, a manager of nonroutine tasks acts as a coach, leading
his employees but allowing them freedom to find innovative solutions to the task at hand.
To illustrate the switch from colonel to coach, consider an academic professor. When the
profess