The Effect of Superstar Firms on College Major Choice∗

Darwin Choi

Chinese University of Hong Kong

dchoi@cuhk.edu.hk

Dong Lou

London School of Economics and CEPR

d.lou@lse.ac.uk

Abhiroop Mukherjee

Hong Kong University of Science and Technology

amukherjee@ust.hk

First Draft: February 2017This Draft: May 2020

∗We thank Ashwini Agrawal, Li An, Josh Angrist, Tania Babina, Nick Barberis, Pedro Bordalo, JamesChoi, Daniel Ferreira, Xavier Gabaix, Nicola Gennaioli, Robin Greenwood, David Hirshleifer, CarolineHoxby, Dirk Jenter, Theresa Kuchler, Daniel Metzger, Per Ostberg, Paige Ouimet, Steve Pischke, Christo-pher Polk, Wenlan Qian, Antoinette Schoar, Kelly Shue, David Solomon, Johannes Stroebel, Geoff Tate,Constantine Yannelis, and seminar participants at Australian National University, CEMFI Madrid, ChineseUniversity of Hong Kong, Deakin University, Durham Business School, Hong Kong Polytechnic University,Hong Kong University of Science and Technology, KAIST, London School of Economics, Loughborough Busi-ness School, Monash University, New York University, Peking University, Tsinghua PBC School of Finance,Tsinghua SEM, Toulouse School of Economics, University of Hong Kong, University of Oxford, University ofToronto, Yale University, 2017 Adam Smith Finance Conference, 2017 European Finance Association AnnualMeeting, 2017 Stanford Institute for Theoretical Economics (SITE) Labor and Finance Meeting, 2017 Laborand Finance Group Meeting, 2018 American Economic Association Annual Meeting, 2018 CEPR SpringSymposium in Financial Economics, 2018 NBER Summer Institute Education/Labor Studies, 2018 MiamiBehavioral Finance Conference, 2019 American Finance Association Annual Meeting for helpful commentsand suggestions. All remaining errors are our own. We are grateful for funding from the Paul Woolley Centerat the London School of Economics.

The Effect of Superstar Firms on College Major Choice

Abstract

We study the effect of superstar firms on an important human capital decision—collegestudents’ major choice. Salient occurrences of superstar performers in an industry arefollowed by a significant rise in the number of college students choosing to major inrelated fields. This cohort effect remains significant after controlling for lagged indus-try returns and wages. Students’ tendency to follow superstars, however, results inlower real wages earned by entry-level employees when these students enter the jobmarket. Further evidence from the National Survey of College Graduates shows thatthis adverse impact on career outcomes can last for decades.

JEL Classification: D81, D91, G40, I23, J24

Keywords: Salience, Human Capital Investment, Skewness, College Major Choice

1 Introduction

Anecdotal evidence in the popular press points to an interesting link between college students’

major choice and the presence of superstar performers—a small number of firms that have

done exceptionally well in the recent past—in related industries. For example, Stanford

Daily reports that the number of Stanford undergraduate students that declared a Computer

Science major in 2013 was nearly four times that in 2006, possibly due to the high-profile

successes of a handful of mobile-app and social-media companies such as Facebook. A New

York Times article on June 15, 2011 indeed argues that “students are flocking to computer

science because they dream of being the next Mark Zuckerberg.”

One possible account for this link is that college students’ attention is drawn to—and their

expectations and decisions shaped by—salient occurrences of superstar performers.1 For one

thing, superstar firms and entrepreneurs attract disproportionate media coverage and social

attention: the story of Mark Zuckerberg, one of the youngest self-made billionaires, is a

constant talking point in the popular press and on college campuses. In addition, the occur-

rences of superstar firms and entrepreneurs are often accompanied by extreme payoffs: Mr.

Zuckerberg is consistently named one of the world’s wealthiest and most influential individ-

uals. A combination of these two salient features—disproportionate social attention coupled

with extreme payoffs—make superstar firms particularly attractive to college students when

deciding their human capital investment.2

The importance of salience in driving human decisions in light of limited cognitive re-

sources has long been of interest to psychologists. As Taylor and Thompson (1982) put it,

“salience refers to the phenomenon that when one’s attention is differentially directed to one

portion of the environment rather than to others, the information contained in that portion

will receive disproportionate weighting in subsequent judgments.” Kahneman (2011) goes

on to argue that “our mind has a useful capability to focus on whatever is odd, different or

unusual.” Bordalo, Gennaioli, and Shleifer, in a series of papers, formalize and apply the

core idea of salience to various economic settings—e.g., in the context of consumer choice,

financial investments and asset prices, and the judicial process.

1Note that “superstar firms” in our context refer to a handful of firms with exceptional recent performance.This definition differs from the one often used in the industrial organization literature, where “superstarfirms” refer to the largest firms (in terms of market value, employment or sales) in an industry. For example,a superstar performer in the tech sector in the early 2010s was Facebook, while the largest firm in the techsector in the same period was Microsoft, which was struggling in its competition against Apple and Google.

2While for most part of the paper, we are agnostic to which of the two features—one that is relatedto belief errors and the other to preferences for skewed payoffs—is driving our results, we provide somesuggestive evidence to differentiate the two in Section 5.2.

1

Applying this insight to the setting of college major choice, we hypothesize that the salient

occurrences of superstar firms and entrepreneurs—albeit non-representative of the whole

industry—can play a substantial role in shaping students’ expectations and decisions. This

channel is particularly relevant given the substantial search costs faced by college students

when choosing major fields, with respect to future job prospects, income, and fit (see, e.g.,

Hoxby, 2004; Hastings, 2016; Huntington-Klein, 2016). Simply put, our paper empirically

analyzes the impact of salience on one of the most important and irreversible decisions in

life—the choice of education and human capital investment.

To operationalize our empirical analysis, we take the following steps. First, we mea-

sure the occurrence of superstar performers (as well as super-losers) in each industry by the

cross-sectional return skewness in that industry.3 In our baseline results, we use employment-

weighted skewness to give more weight to more important/visible firms in the industry.4

Positive cross-sectional skewness indicates that, holding the industry’s average return and

return volatility constant, a small number of firms in the industry have performed exception-

ally well; these salient examples of extreme successes may draw college students to related

majors. Negative cross-sectional skewness, on the other hand, indicates that a small number

of firms in the industry have done exceptionally poorly, which likely turn students away from

related fields.

To provide more intuition for cross-sectional return skewness, we also construct an ap-

proximate measure of skewness, tailN : |topN − median| − |bottomN − median|, i.e., the

distance between the top N firms in the return distribution and the median firm, minus the

distance between the bottom N firms and the median firm.5 Take N = 1 for example, if the

best performing firm does spectacularly better than the median firm in the industry while

the worst performing firm only mildly worse than the median firm, this alternative skewness

measure will turn out highly positive. In other words, this measure is designed to capture

the presence of a very small number of extreme winners and/or losers in the industry. As we

show later, all our main results go through with this approximate (more intuitive) definition

of skewness.

3Unlike the standard deviation, skewness of a distribution can be both positive and negative. For refer-ence, the normal distribution has zero skewness—for each positive outcome, there is an equal chance of anequal-sized negative outcome.

4Employment-weighted skewness assigns a weight that is proportional to employment size to each firmin calculating the skewness of the return distribution. We do this for two reasons. First, larger firmsare more visible and therefore should carry a higher weight in decisions. Second, other variables in ouranalyses, including average wage, average industry return and return volatility, are also weighted by the sizeof employment. Our results are also robust to using firm-size-weighted skewness.

5We exclude firms below the 50th percentile of the size distribution when selecting the top and bottomN firms in each industry to ensure that we use only large, visible firms.

2

Second, we focus on the set of science and engineering majors (e.g., computer science

vs. chemical engineering) that can be easily mapped to one or more industry sectors (e.g.,

information technology vs. pharmaceutical).6 Third, since college students can declare their

major by the end of the sophomore year, we focus on industry return skewness measured in

years t-7 to t-3 prior to the graduation year (i.e., from their junior year in high school to

sophomore year in college) to explain college major distribution at graduation in year t.7

Our empirical results reveal a strong positive relation between the occurrence of superstar

firms and enrollment in related college majors at the cohort level. Using college enrollment

data from the Integrated Postsecondary Education Data System (IPEDS), we find that a

one-standard-deviation increase in within-industry (cross-sectional) return skewness in years

t-7 to t-3 is associated with a 14.07% (t-statistic = 4.60) increase in the number of graduates

in related major fields in year t. This result is after controlling for the average industry

return and return volatility measured over the same period, as well as the lagged average

wage, and time and major fixed effects. Also, this result is stronger among elite universities

in states where the superstars’ industry has a substantial presence (e.g., Stanford graduates

and superstar Tech firms), potentially because students from top universities are more likely

to associate themselves with extreme successes.

Such a positive association between college major enrollment and within-industry return

skewness can be consistent with both demand- and supply-based explanations. First, oc-

currences of superstar firms may be indicative of brighter industry prospects, and students

choose related major fields in anticipation of more and better job opportunities—a chan-

nel that works through labor demand. Second, although the presence of superstar firms is

uninformative about future industry prospects, students attach a disproportionate weight

to these salient yet non-representative observations when choosing their majors—a channel

through labor supply.8 To empirically evaluate the relative importance of labor demand vs.

supply in driving our result, we simultaneously examine the wage and number of employees

(the price-quantity pair) at the time when the cohort enters the job market. The granularity

of employment data from Bureau of Labor Statistics (BLS) allows us to focus squarely on

the wage and employment of entry-level positions that require a college degree, and contrast

them with those of job positions that require prior experience.

6We exclude physical sciences and humanities majors (e.g., physics, history) from our analysis, becausestudents choosing these majors are unlikely to be targeting any specific industry jobs; for example, duringour sample period, most physics majors later become high school teachers.

7Our results are also robust to other return windows, e.g., t-8 to t-3 and t-6 to t-3.8A related possibility is that within-industry return skewness is informative of future industry prospects,

but students tend to overreact to this information; consequently, shifts in labor supply outweigh shifts inlabor demand.

3

Our results reveal a relatively larger outward shift in labor supply than in labor demand

with the occurrence of superstar firms. After controlling for year and major fixed effects, a

one-standard-deviation increase in industry return skewness in years t-7 to t-3 is associated

with a 1.11% (t-statistic = 3.00) drop in the real wage earned by entry-level employees in

related majors in year t.9 In contrast, within-industry return skewness in years t-7 to t-3

has little effect on the average wage of advanced positions in year t, consistent with the

fact recent college graduates do not compete for these types of jobs. Moreover, the effect

of industry return skewness on the average entry-level wage declines with the “versatility”

of the major, defined as the diversity of employment opportunities for graduates from that

major across different industries. This is because an increase in student supply is now shared

among many industries, leading to a smaller wage decline in each of them.

Meanwhile, the effect of lagged industry skewness on the number of entry-level employees

in year t is indistinguishable from zero. This is consistent with the view that labor demand

is relatively inelastic in the short run, as it takes time for firms to expand their operations.

An increase in labor supply (in the form of a larger number of college graduates coming

out of related fields) thus lowers the average wage earned by entry-level employees without

affecting the size of employment.

To understand the long-term impact of college students’ major choice on their career

outcomes, and to provide more granular evidence for our hypothesis, we draw on individual-

level data from the National Survey of College Graduates. The average respondent in this

sample is 43.9 years old, or roughly 20 years out of college; this allows us to examine income

at the individual level over a long period of time. Our analysis reveals that after controlling

for a range of fixed effects—e.g., for different demographic characteristics (like age and

gender), survey-year, industry, and major—a one-standard-deviation increase in industry

return skewness associated with the respondent’s declared major forecasts 0.99% (t-statistic

= 3.81) lower real earnings (including wages and bonuses) by the respondent at the time

of the survey. Such skewness is also associated with a 4.33-percentage-points (t-statistic =

2.97) higher likelihood that the respondent no longer works in an industry related to her

college major. These results suggest that an outward shift in labor supply not accompanied

by a similar shift in labor demand can have a long-lasting adverse impact on individuals’

career outcomes.

9Our empirical design is to compare the average market-adjusted entry-level wage of the same majoracross different cohorts. An alternative design would be to compare a student with his counterfactual-self,had he chosen a different major. An obvious issue with this second approach is that individual major choicecrucially depends on unobservable personal characteristics, such as ability and interest in the subject-matter.This is less of a concern at the cohort level—so long as the average ability and interest of each cohort doesnot vary systematically over time. We revisit this issue in Section 4.5.

4

As we have discussed earlier, a labor-demand-based explanation is unlikely to account

for the entirety of our results: a) industry return skewness has a significantly positive effect

on college students’ major choice, b) a negative effect on entry-level wages in related majors,

and c) an insignificant effect on entry-level new hires. Still, a potential concern is that there

may be a selection effect driving our findings. When a larger number of students select into a

major, it is conceivable that the quality of the marginal students in the major declines and/or

the matching quality between the marginal students’ skills and those required for the major

declines. First and foremost, this alternative story does not provide an explanation for why

students are attracted to major fields with superstar performers. This selection mechanism

is also inconsistent with two additional observations. First, as discussed earlier, the effect

of industry return skewness on college major choice is stronger among elite universities in

states where the superstars’ industry has a substantial presence; so it is unlikely that industry

superstars mostly attract less able students. Second, and along a similar line, the effect of

industry skewness on subsequent entry-level wages is larger for occupation codes that pay

higher wages within a major (e.g., a larger effect on the wages of software developers than

on database administrators for Computer Science majors).

Finally, an important premise in our empirical design is that cross-sectional return skew-

ness of an industry reflects salient occurrences of superstar firms in that industry. We

corroborate this assumption by using a more direct measure of salient events—skewness in

media coverage and tones. To this end, we calculate a news-tone score for each firm in

every year. News skewness of an industry is then defined as the cross-sectional skewness of

news-tone across all firms in the industry. Intuitively, a positive (negative) news skewness

measure indicates that, all else equal, a few firms in the industry have received a dispropor-

tionate amount of positive (negative) media coverage. Not surprisingly, the news skewness

measure is strongly and positively correlated with contemporaneous stock return skewness.

Moreover, when we repeat our analysis to forecast future college major enrollment, we find

that a one-standard-deviation increase in news skewness in years t-7 to t-3 is associated with

a 13.26% (t-statistic = 2.75) increase in the number of graduates in related majors in year t.

This news-based skewness measure also negatively forecasts future entry-level wages with a

point estimate of -1.52% (t-statistic = -2.92), and is unrelated to the number of entry-level

employees when these students graduate.

The remainder of the paper is organized as follows. Section 2 provides a background and

literature review. Section 3 describes the data we use. Section 4 reports the main results

of our empirical analyses. Section 5 reports additional tests and robustness checks. Finally,

Section 6 concludes.

5

2 Background and Literature Review

This paper provides evidence for a growing literature on the impact of salience on human

decision making (Bordalo, Gennaioli, and Shleifer, 2012; Gabaix, 2014). Bordalo, Gennaioli,

and Shleifer (2013a) examine the effect of salience on human decisions in the context of

consumer choice, Bordalo, Gennaioli, and Shleifer (2013b) study its effects on asset prices,

and Bordalo, Gennaioli, and Shleifer (2015) examines judicial decisions. Chetty, Looney, and

Kroft (2009) also study the effect of salience on tax decisions. In neuroeconomics, Fehr and

Rangel (2011) show that subjects evaluate choices by aggregating information about different

attributes, with decision weights influenced by attention. Recent theoretical research has also

developed novel ways of incorporating limited attention in economic decisions, for example

through sparsity models (Gabaix, 2014) where agents build a simplified model of the world

only taking into account a subset of the important variables. While none of these papers

have examined the role of salience in education choice, like we do here, it is perhaps a natural

application—given the complexity of the search process for information about job prospects

(e.g., Stigler, 1961; 1962).

Our result that the occurrences of superstar firms in an industry forecast worse job

outcomes for graduates from related major fields can be consistent with both preference-

and belief-based explanations. On the preference side, Rosen (1997) presents a model of

preferences for skewness based on state-dependent utility. In our context, students with a

preference for skewed payoffs may be willing to accept a lower average wage for a small

chance of having significantly better job opportunities. A preference for skewness is also a

central theme in the non-standard utility literature, e.g., prospect theory (Kahneman and

Tversky, 1979). Most prominent applications of this theory, however, have been in finance

(e.g., Barberis and Huang, 2008). A different explanation for our documented pattern is that

it reflects students’ mistaken beliefs: students who do not have the capacity or resources to

go through detailed industry employment records might base their income expectations on

just a few salient, non-representative observations. Theory and evidence on such mistaken

beliefs leading to oversupply can be found as far back as in Kaldor (1934), or more recently,

in Greenwood and Hanson (2015), although in contexts different from ours. While for most

part of the paper, we are agnostic to which of the two channels—belief errors or preferences

for skewed payoffs—is driving our results, we attempt to differentiate the two through a

customized survey in Section 5.2.

Fluctuations in the labor market and their relation to prevailing economic conditions

have been a focal point of academic discussions at least since the post-War era (e.g., see

6

Blank and Stigler, 1957; or Arrow and Capron, 1959 on the labor market for engineers).

The shortage of engineers in the late 1950s and early 1960s was followed by a surplus in

the late 1960s and early 1970s, and again a shortage in the latter half of the 1970s. This

motivated Freeman (1971) to apply cobweb theory to labor demand and supply decisions in

the market for new graduates. Freeman (1975, 1976) then assesses the predictions of cobweb

theory in the market for new engineers and lawyers. Cobweb theory exploits the time delay

between the decision to enroll in a major and the subsequent entrance into the labor market.

Under this view, a lower salary in major M in year t attracts fewer freshmen and, ultimately,

produces fewer graduates—but this change in supply manifests itself in the labor market

four years later. Consequently, starting wages in year t+4 become higher, affecting in turn,

major choice of the freshman cohort of that year, hence producing endogenous cycles in both

enrollment and wages. While our paper also a) acknowledges the time lag between major

choice and graduation and b) builds on the premise that students form expectations based on

stale, non-representative information, our focus is different: we are interested in the effect of

salient occurrences of superstar firms in an industry on the distribution of major enrollment,

while controlling for lagged industry wages and returns.

Our paper also relates to the literature on the impact of sectoral booms and busts on

workers’ choices and outcomes. Oyer (2008) shows that MBAs are more likely to become

investment bankers if the financial sector is expanding while they are in school. Gupta and

Hacamo (2018) show that the 2000s boom in finance-related industries attracted engineers

to the financial sector. Charles, Hurst, and Notowidigdo (2018) show that the housing boom

reduced educational attainment because individuals dropped out of school to work in the

housing sector. Hombert and Matray (2019) use data from France to show that skilled

workers entering the Information and Communication Technology sector during the 1990s

tech boom experience lower earnings over the longer term. None of these papers examine

the role of a few salient superstar firms within an industry attracting students into certain

sectors, which is our focus here.

Finally, we also contribute to the vast literature on students’ education choice and on

career outcomes (Hoxby, 2003 provides a comprehensive review).10 Most prior studies on

college major choice use a rational expectations framework in which students form expec-

tations of future earnings using statistical modeling and Bayesian updating. Berger (1988)

10See also James, Alsalam, Conaty, and To (1989), Altonji (1993), Sacerdote (2001), Avery and Hoxby(2004), Hoxby (2004), Bhattacharya (2005), Dickson (2010), Blom (2012), Goldin (2014), Lemieux (2014),Stinebrickner and Stinebrickner (2014), Zafar (2013), Bordon and Fu (2015), Altonji, Arcidiacono, andMaurel (2016), Arcidiacono, Hotz, and Kang (2012), Fricke, Grogger, and Steinmayr (2015), Wiswall andZafar (2015), among others.

7

is an early example of this. Subsequent research complements this approach (e.g., Altonji,

1993) by incorporating uncertainties (e.g., uncertainties about ability, preference and aca-

demic progress) to the baseline model. Our paper contributes to and deviates from this

literature by examining the role of salient events—occurrences of superstar firms—in deter-

mining college students’ expectations and major choice. More broadly, our results speak to

the literature on human capital investment. Given the near irreversibility of human capital

investment at the college level, our results suggest that salient extreme events have a large,

permanent impact on individuals’ lifetime income.

3 Data

Our data on college enrollment are obtained from the Integrated Postsecondary Education

Data System (IPEDS) and the National Science Foundation (NSF). NSF uses IPEDS Com-

pletions Surveys conducted by the National Center for Education Statistics (NCES) and

reports the annual number of bachelor’s and master’s degrees in science and engineering

fields. A list of the major fields (which can be readily mapped to industry sectors as de-

scribed below) is presented in Online Appendix Table A1. These degrees were conferred

between 1966 and 2017 by accredited institutions of higher education in the US, which in-

cludes the 50 states and the District of Columbia. Starting in 2001, IPEDS also provides

the annual number of graduates from each institution.11

We map the list of science and engineering degrees from Appendix Table A1 to 3-digit

NAICS industry codes, shown in Online Appendix Table A2. This is modified from the merge

of two maps: the 2010 Classification of Instructional Programs (CIP) to the 2010 Standard

Occupational Classification (SOC) Crosswalk and the 2010 SOC to the 2012 NAICS map.

Our mapping aims to capture the industries that students likely follow when they decide on

their majors. Each major field can correspond to multiple industries: e.g., a degree in Health

is linked to Ambulatory Health Care Services (NAICS = 621), Hospitals (NAICS = 622),

Nursing and Residential Care Facilities (NAICS = 623), and Social Assistance (NAICS =

624). Each industry code can also be mapped to several major fields: for example, Petroleum

and Coal Products Manufacturing (NAICS = 324) is associated with degrees in Chemical

Engineering, Industrial and Manufacturing Engineering, Materials Science, and Mechanical

11For the school-level analysis, we sum the number of first majors and second majors to get the numberof graduates. Between 1987 and 2000, IPEDS only reports the number of first majors from each school. Wedo not use these years as they are not comparable to the post-2001 data.

8

Engineering. 12

We start by calculating the employment-weighted cross-sectional return skewness within

each industry.13 Since students may follow multiple industries when deciding their majors, we

then take the equal-weighted average skewness across all industries associated with the major.

Our approach assumes that students are equally exposed to salient events in all industries

related to the major. We measure all industry-level variables (e.g., mean and volatility of

returns) in the same way. Our results are robust to using weighted-average skewness across

industries based on industry total market capitalization or total employment, as well as other

ways to aggregate industry-level variables (e.g., focusing on the industry with the maximum

absolute skewness or computing one skewness measure by first pooling firms across all related

industries).

Aggregate wage and employment data between 1997 and 2017 are available from the

Bureau of Labor Statistics (BLS) through the Occupational Employment Statistics (OES)

program. Wage is defined as straight-time, gross pay, exclusive of premium pay. Wage

and employment data are reported at the SOC code level in each industry. BLS provides

projections of the job requirement (degrees and approximate number of years of experience

required) of many SOC codes. We make use of the CIP-SOC Crosswalk and the BLS

projections to define entry-level jobs for graduates of each major. These are jobs that are

suitable for students of a particular major and that require a bachelor’s degree but does not

require prior work experience. Note that the aggregate wage and employment for each major

is calculated from the employment-weighted wage and total employment across all relevant

SOCs. Each SOC (e.g., Computer Programmers) can appear in multiple industries, some of

which (e.g., Financial Services) can be different from our major-NAICS map in Appendix

Table A1. In other words, our measures of aggregate wage and employment are derived from

the actual employment data and do not rely on the major-NAICS map.

We also use the National Survey of College Graduates sponsored by the NSF and con-

ducted by the Census Bureau. From the survey, we can obtain information about survey

respondents’ graduation year, major, demographics, total earnings (including variable com-

pensation, e.g., bonuses), and employment.14

12Two of the majors, Computer Sciences and Electrical Engineering, are mapped to the same set ofindustry codes as they are closely related.

13The employment-weighted return skewness is given by∑ni wi(

ri−r̄σ̂ )3∑n

i wi, where n is the number of firms

in the industry, wi is the number of employees in firm i, ri is the return of firm i, and r̄ and σ̂ are theemployment-weighted mean and standard deviation of return, respectively.

14The survey is conducted every two years since the 1970s. Data are available online beginning 1993. Werequire information on respondents’ majors, which is publicly available in the years 1993, 2003, 2010, 2012,

9

Our news sentiment data are from RavenPack News Analytics, which quantifies posi-

tive and negative contents of news reports. We focus on the Event Sentiment Score (ESS)

constructed by RavenPack. ESS is determined by matching stories typically categorized by

financial experts as having positive or negative financial or economic impact. It ranges be-

tween 0 and 100, where 50 represents neutral sentiment, and is available between 2000 and

2017.

We present summary statistics for our variables of interest in Table 1. The median number

of bachelors in each major is 8,344 students per year, with males contributing approximately

81.5% of that number. We define industries at the 3-digit NAICS level. On average, our

industry returns are positively skewed in the cross-section, with a median annual skewness of

0.86. The median cross-sectional skewness in news tone, measured from the Ravenpack news

analytics data, is 1.46. The employment-weighted average entry-level wage for workers with

a bachelor’s degree in science and engineering has a median of $56,889 (in 1997 dollars).15

The median annual change in the log number of employees in these positions is about 2.3%,

and the median change in the number of employees is 4,395.

4 Main Results

This section presents the main results of our paper. We start by examining the relation

between the occurrence of superstar firms in an industry and the subsequent number of

students choosing related major fields. Note that in contrast to standard major choice

regressions typically run at the individual level using survey data, we are mostly interested

in variation at the cohort level, as a function of industry characteristics.

4.1 Number of Graduates in Different Majors

Our main hypothesis is that while deciding upon a major, students get disproportionately

attracted to industries where a small number of superstar firms have been performing ex-

ceptionally well. For example, as mentioned in the Introduction, the salient success story

of Facebook helps draw students to the Computer Science major. More generally, given the

substantial cognitive costs faced by college students in figuring out job prospects, income,

2015 and 2017.15Note that we do not have data specifically on the first year of employment; the wage and employment

figures include seasoned workers who are still in these entry-level positions and have not been promoted.

10

personal interest in various industries, students’ expectations—and hence major choice—are

disproportionately influenced by salient, easy-to-recall events.

To analyze the effect of occurrences of superstar performers (measured by cross-sectional

returns skewness) on major choice decisions at the cohort level, we estimate the following

regression equation:

log(bachelori,t) = α + βLaggedSkewi,t−3 + γXi,t−3 + µi + τt + εi,t, (1)

where log(bachelori,t) is the natural logarithm of the number of graduates in major category

i in year t (t refers to the calendar year of graduation, all other time variables are expressed

with respect to this; i.e., t-k refers to k years before graduation). LaggedSkewi,t−3 is the

cross-sectional skewness of returns relevant to that major, calculated using returns data

from t-7 to t-3). Note that our skewness measure is lagged to reflect that extreme salient

events can only affect major choice if they occur before the major is decided, which for most

students is, at the latest, their sophomore year in college.16 This measure yields a highly

positive number if a handful of firms have done exceptionally well in that period in industries

related to that major.

Xi,t−3 is a vector of controls, suitable to our setting analyzing major choice at the cohort

level. µi and τt are major and time (year) fixed effects, respectively. Our vector of controls

includes the average return of firms in related industries between t-7 and t-3, to control

directly for the average firm performance in an industry. We also include a measure of

volatility of firm performance. These two controls ensure that our skewness measure picks

up differences between industries that have performed similarly in the recent past, and have

had similar stability of performance. Next, we include as a control the average size (market

capitalization) of firms in that industry. This is to account for any effect that might arise

from the presence of a few large firms—as opposed to a collection of smaller firms—in an

industry. We also account for the average firm age, and industry valuation ratio (book-to-

market, B/M), which are typically regarded as proxies for firm growth within an industry.

The inclusion of major fixed effects ensures that our identification of the coefficient of interest,

β, comes from changes in the number of graduates, not its level. Inclusion of time fixed effects

purges out any market-wide fluctuation from our estimate.

If our hypothesis—that college students’ major choice is influenced by superstar firms—is

16Further, Table A3 in the Online Appendix conducts a test using skewness measured in years t-2 and t-1,i.e., which are likely after major declaration (and hence only reflect those who switch majors). As expected,LaggedSkew in these last two years has a substantially more muted effect on major choice, reflecting thefact that switching majors is far less popular than sticking to a declared major.

11

indeed true in the data, we expect to see skewness positively predicts the number of graduates

in related major fields in the future. That is, the coefficient on the LaggedSkew measure in

the major choice regression, β, should be positive. We present these results in Table 2. As we

can see from column 1, LaggedSkew predicts major choice strongly, even after controlling

for the average return in the industry and its cross-sectional dispersion. For example, a

one-standard-deviation increase in employment-weighted LaggedSkew is associated with an

increase in the number of students majoring in related fields by 13.92% with a t-statistic of

4.61 (all explanatory variables are standardized for ease of comparison).

In column 2 of the same table, we further control for the average wage earned by each

major (LaggedLogAverageWage). This is to examine whether the empirical relation we

document above is distinct from the cycles of college major enrollment, and future employ-

ment and wages as described in Cobweb models. Consistent with prior literature on Cobweb

theory (e.g., Freeman, 1976), we control for the three-year lagged average real wage for each

major. (Our results are robust to controlling for lagged real wages measured over different

horizons.) As our results show, the past average wage does indeed significantly predict major

choice. But, importantly, our skewness measure continues to retain its predictive power—

both in terms of economic magnitude and statistical significance, consistent with the view

that its relation to major choice is distinct from that of the past average wage.

In column 3 of Table 2, we examine school-level data from four-year universities. We

focus on 262 schools that offer at least 6 out of our 11 majors and study whether the effect

of superstar firms is any different among top schools, especially those located in states with

significant presence of related industries (e.g., California for Tech jobs). To this end, we add

an interaction term for schools that are in the top 50 in the US News 2018 Best Colleges list

and are located in the state that hires the most people in related industries. Our results show

that top school students are especially drawn by local superstar performers. Moreover, we

also find that the relationship between log(bachelor) and LaggedSkew holds even if we drop

the Tech Bubble period (shown in Panel B of Online Appendix Table A4). Further, these

results are also robust to using cohort-level data on Masters degrees instead of Bachelors

(shown in Panel A of Appendix Table A4).

Finally, we would like to highlight two key aspects of our empirical design. First, many of

our majors can be stepping stones to careers in multiple industries and choosing to graduate

with a particular major does not necessarily limit the student to work in the industry most

closely related to it. For example, Computer Science graduates can also work as librarians.

All we assume for our analysis is that at the time the student chooses to major in Computer

Science, he is much more interested in a career in the Computing or Tech industry than he

12

is interested in librarianship.

Second, we use skewness in stock returns, rather than skewness in firm size, to reflect

the fact that students are more likely to be attracted by what is “exciting” at that time.

This notion of salience is more closely captured by a handful of superstar firms performing

exceptionally well and thereby capturing public imagination and media attention, and is less

so by the presence of a few dominant firms in the industry. Note that even though such

exciting firms may lead a student to choose a major field, there is often a small chance of

actually working for these dream companies. For example, even though Facebook’s success

was (and perhaps still is) capturing social attention from college campuses to movie studios,

the firm accounted for a tiny fraction of all jobs for Computer Science majors. We revisit

this issue in the final section of the paper.

4.2 Effects on Entry-Level Wages and Employment

The fact that students are drawn to industries with superstar performers can be consistent

with both labor-demand- and supply-based interpretations; that is, students are attracted to

these majors either because a) they rationally anticipate improving job prospects in related

industries, or b) they are simply drawn by extreme, salient events that are in fact uninforma-

tive about future job opportunities (or less informative than they expect). To examine the

relative importance of labor demand vs. supply channels, we simultaneously examine two

quantities—wages (inflation-adjusted) and employment. By examining the price-quantity

pair, we can disentangle the relative shifts in the labor supply vs. demand curves.

Note that in these wage tests, the skewness measure is still based on related industries that

students likely think about while making their major choice decisions; yet the wage data are

from actual occupations that absorb graduates from various majors according to the major-

occupation-code map, CIP 2010 to SOC 2010 Crosswalk, provided by the National Center

for Education Statistics (NCES). Note that each occupation code can appear in multiple

industries (e.g., computer programmers in the Tech industry vs. financial services industry);

we take an employment-weighted average across all industries. In our baseline result, we

focus on entry-level employment and wages for jobs that require a bachelor’s degree but no

prior work experience.17

We examine what happens to job opportunities at the time of graduation of our year t

17BLS provides details of job requirements (degrees and the approximate number of years of experiencerequired) for each SOC code.

13

cohort in each major as a function of the presence of a few related firms having performed

exceptionally well. Specifically, we estimate the following regression equation:

log(annual wagei,t) = α + βLaggedSkewi,t−3 + γXi,t−3 + µi + τt + εi,t, (2)

where annual wagei,t is the employment-weighted average entry-level wage for students in

major i in year t. LaggedSkewi,t−3 is our employment-weighted salience measure of indus-

tries related to major i, as explained earlier (our results are robust if we use size-weighted

LaggedSkew instead). We include the same set of controls as in Table 2.

We also control for major and time fixed effects in our regressions, so one way of thinking

about our empirical design is that we compare the average market-adjusted entry-level wage

of the same major across different cohorts, as a function of lagged industry return skewness.

An alternative design would be to compare a student with his counterfactual-self had he

chosen a different major. We do not go down this alternative route as individual major

choice crucially depends on unobservable personal characteristics and interests. This is less

of a concern at the cohort level—as long as the average ability and interest in each cohort

does not vary systematically over time.

Table 3 reports results on wages and employment. In Panel A, Column 1, we examine

entry-level wages as a function of lagged industry skewness. LaggedSkew is significantly and

negatively associated with future entry-level wages. In terms of the economic magnitude,

a one-standard-deviation increase in lagged industry skewness is associated with a 1.11%

(t-statistic = 3.00) lower real wage for entry-level jobs requiring a bachelor’s degree.

In Column 2, we examine net new hires—i.e., year-on-year changes in the number of

employees in entry-level jobs related to the major—as a function of lagged industry skewness.

Given that our dependent variable is already defined in changes, we do not include major-

fixed effects (we are interested in the change in net new hires, not its growth rate). As can

be seen from Column 2, there is no significant relation between LaggedSkew and future new

hires. This suggests that even though salient events drive more students to related major

fields, entry-level job positions do not immediately expand to absorb these extra graduates,

consistent with labor demand being relatively inelastic in the short-run. In Online Appendix

Table A5, we show that industry turnover, defined as total separations minus hires (as a

percentage of employment), is also not related to LaggedSkew. This indicates that job

separation risk is not associated with salience-driven major choice.

In Columns 3 and 4, we add three additional controls. First we add the average number of

14

bachelors graduating in related majors in years t-1 to t-2, to account for the effect of delayed

absorption of the previous two years’ graduates in that industry. Next, we add the ratio

of male to female graduates, to account for changes in gender balance of graduates/entry-

level jobs. Finally, we also add the lagged average wage for related majors, to examine—

as in the previous table—whether our results on wages and employment are also distinct

from predictions of a Cobweb model. LaggedSkew continues to predict entry-level wages

negatively, and continues to be statistically unrelated to net new hires.18

In Columns 5 and 6 we conduct a placebo test, that is, to repeat our analysis from the

preceding columns, but use data on advanced job positions that typically require a Doctoral

or Professional degree or substantial prior experience. If our results are driven by Skew

reflecting changes in demand for labor, through, e.g., changes in industry structure, we

should see a similar coefficient on LaggedSkew even with these advanced jobs. If, instead,

our results are driven by LaggedSkew affecting wages through salience-driven supply of

labor, this should mostly affect entry-level jobs, and we should not find any effect on wages

for advanced positions. Our evidence supports the latter. LaggedSkew has a coefficient

of 0.0007 (t-statistic = 0.12) in column 5, for example, relative to -0.0121 (t-statistic =

3.56) in column 3 (recall that LaggedSkew is standardized, so these coefficients are directly

comparable).

In Panel B, we check whether the effect varies across different occupation codes within

a major and across different majors. First, within each major-year, we separate occupation

codes into those offering average wages above and below the median in year t-2. The former

(latter) group represents higher-(lower-) skilled jobs within the major.19 We calculate the

average wage and change in the number of employees in each of these major-occupation

groups, and add an interaction term between LaggedSkew and Low, where Low indicates

the lower-wage group. Column 1 shows that LaggedSkew mostly affects wages of higher-

skilled occupations, while Column 2 shows that LaggedSkew is not systematically related to

the net new hires in either group. Together with the evidence that top school students react

more to local superstar firms (Table 2), these results suggest that the drop in average entry-

level wage is not due to an influx of less capable students. If anything, the evidence is more

consistent with the view that more capable students are drawn to majors with superstar

performers.

18There is some evidence of rationality in college major choice: a larger number of students choosing tomajor in related fields indeed is associated with higher entry-level wages at graduation, as indicated by thepositive coefficient on log(Number of Bachelors).

19The results are similar if we classify occupation codes into high- vs. low-skilled groups within eachindustry-year.

15

Next, we investigate the notion that the fungibility of employment opportunities is dif-

ferent across majors. Specifically, if graduates from a particular major have employment

opportunities in a variety of industries, then part of the excess labor supply can be shared

among the industries, leading to less downward pressure on wages in each industry. We test

this hypothesis using an interaction term between LaggedSkew and V ersatility, defined as

the diversity of employment for graduates from a particular major. Column 3 reveals that

industry skewness negatively affects real wages mostly for majors that have concentrated job

opportunities in a small number of industries. Again, we do not find statistically significant

differences in entry-level employment size in Column 4.

Overall, the evidence presented in Table 3 indicates that the presence of salient superstar

performers forecasts lower future wages, but does not lead to additional entry-level jobs.20

Put differently, students’ decision to follow superstars in their major choice does not seem

to benefit them upon graduation; if anything, it costs them in terms of starting out with a

lower entry-level salary.

Finally, in Online Appendix Section 1, we provide further evidence for the supply side

of human capital investment by linking time variation in the relative popularity/salience of

different industries to student major choice. More specifically, we exploit: (a) structural

breaks in industry valuation during the NASDAQ bubble in the late 1990s to identify su-

perstar industries; and (b) time variation in the viewership of one of the longest-running

TV series in the US, Law & Order, to gauge the popularity of the legal profession among

prospective students. As shown in Online Appendix Tables A7 and A8, industry salience

plays an important role in driving college students’ major choice.

4.3 A More Intuitive Measure of Superstar Performers

The main skewness variable we use is the standard definition in the literature, designed to

capture the presence of outliers in a smooth and continuous fashion exploiting the entire

distribution. To provide more intuition for skewness, we use an alternative, more discrete

definition of distribution asymmetry: the distance between the right tail of a distribution

and its median, minus the distance between the left tail and the median. More formally,

20In Online Appendix Table A6, we repeat the exercises in Tables 2 and 3 by using alternative ways toaggregate industry return skewness to the major level, specifically by focusing on the industry with themaximum absolute skewness (Panel A) and computing one skewness measure by first pooling firms acrossall related industries (Panel B). The results are by and large unchanged.

16

we define tailN = (|topN − median| − |bottomN − median|)/stdev.21 Take N = 1 for ex-

ample, the measure will show up highly positive if the best performing firm in the industry

does spectacularly better than the median firm while the worst performing firm only mildly

worse—which intuitively captures our definition of an industry with superstar performers.

We repeat our analyses in Tables 2 and 3 using tailN in place of cross-sectional skew-

ness. Again, the measure is constructed using returns data from t-7 to t-3. The results are

shown in Table 4.22 Column 1 shows that this alternative measure positively predicts the

number of bachelors as in Table 2. A one-standard-deviation increase in tailN forecasts a

15.30% (t-statistic = 4.71) higher number of graduates in related major fields. Column 2

shows that most of the effect is coming from students being attracted to superstars—firms

that do significantly better than the average—rather than being repelled by super losers.

The evidence in Column 3 suggests that a one-standard-deviation increase in the return

asymmetry predicts 1.59% (t-statistic = 3.88) lower entry-level wages in related industries

when the cohort graduates. Column 4 shows that this wage effect—in line with the effect on

the supply of graduates in column 2—is also coming solely from superstars, and not super

losers. As before, tailN does not predict any change in entry-level employment size in related

industries.

4.4 Long-Term Effects on Income and Career Outcomes

One possible concern thus far is that although students’ attraction to superstar firms does

not benefit them immediately after graduation, their job prospects may improve in the

longer term. Another concern is that our wage measure in the previous section does not

adequately capture total earnings (which should also include bonuses and stock options). A

third concern is that while we show some results consistent with employment not expanding

in the short-term to keep pace with increased labor supply, we have not provided direct

evidence as to how the extra supply is eventually absorbed by the labor market, possibly

by other sectors not directly related to the chosen major. We explore all these issues by

examining long-term survey data from the National Survey of College Graduates (NSCG).

Respondents report their total earnings, which include other types of compensation in

addition to wages.23 The average age of the respondents at the time of the survey is 43.9

21We exclude firms below the 50th percentile of the size distribution when selecting the top and bottomN firms in each industry so that the measure is not dominated by small firms.

22In our baseline result, we pick N = 10 to reduce the noise in the measure, but as shown in OnlineAppendix Table A9, our results are robust to other choices of N (e.g., 1, 3, 5).

23In terms of respondent characteristics, 67.7% of the NSCG respondents are male and 71.7% are married.

17

years, so roughly 20 years out of college; thus the survey provides useful information on long-

term outcomes. Another advantage of the survey is that it contains information on whether

the graduate works in a sector which he/she considers typically unrelated to his/her field of

study (Job Outside Main Field of Study). If there is indeed an excess supply of graduates

looking for jobs in sectors with superstars, and these sectors do not expand employment to

keep pace with labor supply, as we see in Tables 2 and 3, we should expect to see a positive

correlation between LaggedSkew and the propensity to work in industries outside the field

of study.

Note that unlike the previous tables which were based on aggregate (cohort-level) data

on majors and industries, the survey data in the section are at the individual level. Our

regression specification employs non-parametric controls for a variety of determinants of

earnings and job choice. These include a host of fixed effects for majors, years, gender,

marital status, minority status, industries, survey cohorts, and age deciles.

Table 5 shows the results for respondents who are employed full-time (those who work

more than 35 hours in a typical week). The first two columns report results from panel

regressions with log(Earnings) as the dependent variable, while the last two columns report

results from Logistic regressions with a categorical dependent variable indicating whether the

graduate now works outside his field of study. Our results reveal that one-standard-deviation

higher LaggedSkew is associated with 0.66% to 0.99% lower earnings and 4.11–4.33% higher

propensity to work in a job outside the field of study. This is consistent with an expanding

literature on long-term effects of adverse initial labor market conditions on the lifetime

earnings of college graduates (see, for example, Oyer, 2006, 2008; Oreopolous et al. 2012).

Overall, our results suggest a relatively larger shift in labor-supply in response to salient

superstar performers in some sectors of the economy. In the short run, labor demand seems

relatively inelastic, so a sudden increase in labor supply lowers the average wage earned by

entry-level employees, without affecting the size of entry-level employment. This adverse

effect on earnings lasts for years/decades: graduates continue to earn lower income and have

a higher likelihood of taking up jobs in sectors unrelated to their fields of study.

4.5 Discussions of Alternative Explanations

In this section, we discuss a number of potential explanations for our evidence, and discuss

why none of these alternative mechanisms can explain the entirety of our empirical findings.

18

Changes in aggregate labor demand Perhaps the most obvious alternative interpreta-

tion of the data is that occurrences of superstar performers are indicative of brighter future

industry prospects, and students choose related major fields in anticipation of more and bet-

ter job opportunities—a channel that works through labor demand. As we have discussed

throughout the paper, this explanation is inconsistent with the findings that a) industry re-

turn skewness negatively forecasts entry-level wages when students enter the job market, and

b) high industry skewness is not followed by an increase in entry-level hiring. Further, we

show in Online Appendix Table A10 that industry return skewness is uncorrelated with a list

of observable measures of future industry performance, such as profitability and sales, which

should reflect changes in future labor market opportunities (see Online Appendix Section

2.3 for more details).

Composition of labor demand A more nuanced version of the labor-demand story is

that occurrences of superstar performers are associated with changes in the composition of

labor demand. More specifically, industry skewness forecasts higher demand for low-quality

employees but lower demand for high-quality employees, thus keeping the total employment

constant while driving down entry-level wages. This mechanism, however, cannot explain

why a larger number of students (especially those from elite universities) choose to major in

related fields in response to industry return skewness, knowing that future job opportunities

are worse.

Bargaining power Another explanation is that return skewness reflects changes in the

industry structure, which impacts firm bargaining power vis-a-vis employees. Specifically,

superstar performers gain market power as they grow, which they then exploit to negotiate

wages down. First, this mechanism does not explain the finding that industry return skewness

only negatively predicts entry-level wages, but not wages earned by occupations in the same

industry that require prior experience. Second, and more importantly, this mechanism also

fails to explain why more students are attracted to related major fields by superstar firms,

knowing that their job prospects are going to be worse.

Composition of college graduates Another possibility is that the wage result may

reflect “selection” on the supply side. When a larger number of students select into a major,

it is conceivable that the quality of the marginal students in the major, and/or the matching

quality between the marginal students’ skills and those required for the major, declines.

First and foremost, this alternative story again does not provide a rational explanation for

19

why more students are attracted to major fields with superstar performers. Moreover, this

explanation is inconsistent with two additional pieces of evidence discussed earlier: a) the

effect of industry skewness on college major choice is stronger for elite university students,

unlikely to be of lower quality; b) the effect of industry skewness on subsequent entry-level

wages is larger for high-wage occupation codes within each major, which also attract more-

productive, better-quality employees.

The counterfactual-self Our empirical design throughout the paper is to compare col-

lege students’ major choice and average entry-level wages at the cohort-level ; we implicitly

assume that cohort characteristics do not vary significantly with industry return skewness.

A potential concern with our approach is that students that are attracted by superstar firms

may in fact be better off relative to the counterfactual-self had he/she chosen a different

major. This welfare analysis is difficult to implement because individual major choice de-

pends crucially on unobservable personal characteristics, such as ability and interest in the

subject-matter. Still, our approach is valid as long as the average ability and interest of each

cohort does not covary systematically with related-industry skewness over time. Moreover,

our survey evidence (reported below) indeed points toward ex-post regret for not choosing

college majors optimally: graduates from superstar-dominated majors are more likely to

admit that they did not spend as much time as they should have on major choice.

5 Additional Tests

5.1 Other Measures of Extreme, Salient Events

Until now, we have measured salient occurrences of superstar firms in an industry based on

stock returns. This is not to suggest that high school students, or for that matter first and

second year college students, follow the stock performance of all firms on a regular basis to

be able to calculate or affected by stock return skewness. Indeed, we think of LaggedSkew

as a proxy for salient events taking place in related industries that draw students’ attention

and shape their expectations and decisions.

While there could be many types of prominent events that affect just a few firms, con-

tributing to LaggedSkew, one overarching outcome of any such event is media attention.

Thus, an alternative way of capturing extreme, salient industry events is to exploit the cross-

sectional skewness in media coverage and tones received by firms in the industry. In order to

20

measure media skewness, we use a media-coverage positivity score supplied by RavenPack

(which sifts through all news articles published by major financial news outlets starting in

2000). More specifically, we assign a score of −1 to 1 to each news article depending on

the positivity or negativity of the tone, following Dang, Moshirian, and Zhang (2015).24

For every firm, we then calculate the sum of all news scores in each year (similar to the

cumulative stock return in a year). Finally, we calculate the employment-weighted cross-

sectional skewness of this firm-level news tone measure for each industry in each year, labelled

News Skew.

Next, we run regressions similar to equation (1) but replace return skewness with the

media skewness measure discussed above. We report these results in Table 6. In column 1,

we find that media skewness also predicts major choice, with a similar economic magnitude

as the one reported in Table 2. A one-standard-deviation increase in News Skew in years

t-7 to t-3 is associated with 13.26% (t-statistic = 2.75) more graduates from related majors

in year t. Note that this estimate is statistically significant at the 5% level, despite the fact

that our sample size drops substantially due to the lack of media coverage data in the earlier

years.25

In columns 2 and 3, we examine the relation between media skewness measured in years

t-7 to t-3 and entry-level wages and net new hires in year t. One-standard-deviation higher

news skewness is associated with a 1.52% (t-statistic = 2.92) lower entry-level wage; once

again, there is no significant relation between our news-tone skewness measure and net new

hires. These results are consistent with our baseline in Table 3.

5.2 Belief Errors or Preferences for Skewed Payoffs

As discussed earlier, there are two related channels through which college students may be

drawn to industries with superstar performers. One possibility is belief errors: students may

form income expectations (or more generally, expectations of future successes) based on a

small number of non-representative but highly visible observations. For example, a student

24Following their paper, we calculate the weekly average score for each firm. Only news articles withrelevance = 100 (articles which can be ascertained as referring to a certain firm) and event novelty score =100 (novel stories reporting a categorized event within a 24-hour window) are counted.

25In Online Appendix Table A11, we zoom in on two specific types of salient events in the equity market:initial public offerings (IPOs) and firm defaults, both of which can attract substantial media and publicattention to the relevant industry. While IPOs are salient positive events, likely drawing students to relatedmajors, defaults are significant negative events, which are likely to turn students away. We find consistentevidence: IPO returns are positively related to future major enrollment while firm defaults are negativelyassociated with it.

21

may decide to study Computer Science because he is attracted by the salient, extreme success

of Facebook, and does not fully realize that a typical computer science graduate gets a far

less glamorous job, earning a much lower income. The other possibility is that students

indeed form rational expectations about future job opportunities but are happy to accept a

lower average wage for a small chance of hitting the jackpot—a preference for skewed payoffs.

Empirically, it is nearly impossible to differentiate between the two channels using avail-

able data, as we neither observe individuals’ income expectations, nor their preferences for

skewed payoffs. To shed more light on this issue, we design and conduct our own survey

of college graduates on the platform provided by SurveyMonkey (the survey was conducted

in July 2018 and was titled “Looking Back at College Major Choice”; the actual survey

questions are included in Online Appendix Table A12). We screen our respondents using

the following criteria: at the time of the survey, the respondent is a) a college graduate

with one of the NSF majors that we examine in the paper, b) between 21 and 65 years old,

and c) employed full time. Out of the 1,169 individuals that start the survey, our screening

procedure reduces the sample size to 516. In addition, we also screen respondents based

on whether they cared/worried about job market outcomes when they chose their majors,

leaving out those that chose majors based purely on academic interest.26 We then match

the survey data to industry return skewness and other control variables used in earlier tests,

leaving us with a final sample of 398 respondents with non-missing data.

In our sample, 50.12% of respondents are male and the median household income is in

the $75,000–$95,000 range. The median age of survey respondents is in the 30–44 years

range. (Demographic information, such as gender, age and income, is provided to us by

SurveyMonkey.) Table 7 reports results from this survey. To begin, we examine the con-

sistency between our survey evidence and empirical results reported earlier on labor market

outcomes. Column 1 in Panel A shows an ordered logistic regression with Household income

(in 8 buckets: 7 buckets of size $25,000, starting from $25,000, and the last bucket for in-

comes $200, 000+) as the dependent variable.27 Column 2 reports the marginal effects from a

logistic regression with a dummy dependent variable indicating whether the graduate started

his/her post-college career in an industry that he/she expected to work in when choosing

the major.

26Specifically, we asked, “How important was the availability of jobs or future income prospects in relatedindustries (where people with this major typically worked) in your major choice decision?” and only keptthose who answered “Most Important” or “Somewhat Important” (screening out those that chose “LeastImportant”).

27Note that due to the truncation of reported incomes at $200, 000, we do not use any measure of theactual skewness of incomes in our analysis.

22

We are interested primarily in the correlation between our stock return skewness measure

used throughout the paper and these career outcomes. Note that we never ask the survey

respondents about salient extreme events or industry return skewness: knowing their major

and the year they declared the major (which we ask directly in our survey), we can back

out the cross-sectional return skewness relevant to each respondent. In all regressions, we

control for major-, industry-, graduation-year-, gender-, and region-of-residence- fixed effects,

as well as fixed effects based on categorical answers to the question regarding how worried

the respondent was about paying back his student loans when choosing the major.

We find a significant, negative relation (Panel A, Column 1) between cross-sectional

return skewness in related industries prior to major declaration and respondents’ long-term

income. Columns 2 repeats the Job Outside F ield test from the National Survey of College

Graduates in Table 5. The result suggests that industry return skewness positively forecasts

the likelihood that the respondent takes up a job in a field unrelated to his major right after

graduation.28

In Panel B, we examine the possible factors that drive respondents’ decisions to follow

superstar firms. The dependent variable is the industry return skewness associated with the

chosen major of each respondent. The main independent variable of interest in Column 1

is a dummy variable, Expectation errors, based on a question of whether the respondent

thought his/her expectations of future job/income outcomes could have been more accurate

had he/she done a bit more research (Question 14).29 Column 1 of the panel shows that

this self-reported expectation-error variable is significantly and positively associated with

industry return skewness, suggesting that those who chose a major with salient superstar

performers are more likely to have made a mistake in their expectations.

The main independent variable of interest in Column 2 is Lottery preference, a dummy

variable that equals one if the respondent answers that he/she would have preferred a lottery-

like payoff with a small probability of earning a large income over a stable future income

stream had he/she been given that option in college (Question 15). We do not find any

significant relationship between lottery preferences and Skew.

28We also impute from a survey question (Question 9) whether respondents changed their employmentindustries, and find that a respondent with a higher LaggedSkew is indeed more likely to switch industries,although this result is not statistically significant.

29Given that many respondents took the survey long after college graduation, we use two precedingquestions to remind them of the conditions at that time. Specifically we ask: (a) the amount of researchabout job prospects he/she conducted before choosing the major (Question 12) and (b) the job conditionsupon graduation, i.e., how others in their cohort with the same major did right after graduation, relative tothe expectation.

23

The last column presents a horse-race between Expectation errors and Lottery preference.

Again, expectation errors are strongly correlated with skew-driven major choice, but not

lottery preferences. This evidence suggests that expectation errors, rather than an inher-

ent preference for skewness, is more likely to drive individuals’ decisions to chase superstar

performers in their college major choice.

6 Conclusion

This paper examines the effect of superstar firms on an important human capital decision—

college students’ major choice. Using cross-sectional skewness in stock returns or favorable

news coverage as proxies for the occurrences of superstar firms in an industry, we find that

they are associated with a disproportionately larger number of college students choosing to

major in related fields. Students’ tendency to follow superstars, however, seems to result in

temporary over-supply of human capital. In particular, we find that upon entering the job

market, students attracted by superstar performers earn lower real wages relative to their

peers with the same major but from different cohorts. Coupled with the finding that the

number of entry-level employees stays roughly constant, this result confirms the view that

industry labor demand is relatively inelastic in the short run; so a sudden increase in labor

supply lowers the average wage earned by entry-level employees without affecting employ-

ment size. Moreover, these adverse effects on career outcomes can last for years/decades:

cohorts drawn into major fields by superstar performers earn lower wages even 20 years after

graduation and have a lower propensity to work in jobs related to their college majors.

In sum, our paper is the first to examine the role of salient extreme events, such as

the occurrences of superstar firms, in driving one of the most important and irreversible

decisions in one’s life: human capital investment. Our results have implications for both labor

economists who study variation in individual education choice, as well as micro-economists

who emphasize the role of salience and skewed payoffs in human decision making. Finally,

our paper also points to an additional channel—education choice—through which superstar

firms/entrepreneurs may impact social welfare and economic growth.

24

References

Altonji, J. G., 1993, “The demand for and return to education when education outcomesare uncertain,” Journal of Labor Economics 11, 48–83.

Altonji, J.G., Arcidiacono, P. and Maurel, A., 2016, “The analysis of field choice in collegeand graduate school: Determinants and wage effects,” in Handbook of the Economicsof Education, Elsevier.

Arcidiacono, P., Hotz, V. J. and Kang, S., 2012, “Modeling college major choices usingelicited measures of expectations and counterfactuals,” Journal of Econometrics 166,3–16.

Arrow, K. and Capron, W., 1959, “Dynamic Shortages and Price Rises: The Engineer-Scientists Case,” Quarterly Journal of Economics, 73, 292–308.

Avery, C. and Caroline M. Hoxby, 2004, “Do and Should Financial Aid Decisions Affect Stu-dents’ College Choices?” in Caroline Hoxby, ed. College Choices: The New Economicsof Choosing, Attending, and Completing College. University of Chicago Press.

Bai, Jushan, 1997, “Estimation of a change point in multiple regression models,” Reviewof Economics and Statistics 79, 551–563.

Bai, Jushan, and Pierre Perron, 1998, “Estimating and testing linear models with multiplestructural changes,” Econometrica 66, 47–78.

Barberis, N. and M. Huang, 2008, “Stocks as Lotteries: The Implications of ProbabilityWeighting for Security Prices,” American Economic Review 98, 2066–2100.

Berger, M.C., 1988, “Predicted future earnings and choice of college major,” Industrial &Labor Relations Review, 41, 418–429.

Bhattacharya, J., 2005, “Specialty selection and lifetime returns to specialization,” Journalof Human Resources 15, 115–143.

Blank, David M. and George J. Stigler, 1957, “The demand and supply of scientific per-sonnel,” NBER Books.

Blom, E., 2012, “Labor market determinants of college major,” Yale University WorkingPaper.

Bordon, P. and Fu, C., 2015, “College-major choice to college-then-major choice,” Reviewof Economic Studies, 82, 1247-1288.

Bordalo, P., Gennaioli, N. and Shleifer, A., 2012, “Salience theory of choice under risk,”Quarterly Journal of Economics 127, 1243–1285.

Bordalo, P., Gennaioli, N. and Shleifer, A., 2013a, “Salience and consumer choice,” Journalof Political Economy 121, 803–843.

25

Bordalo, P., Gennaioli, N. and Shleifer, A., 2013b, “Salience and asset prices,” AmericanEconomic Review 103, 623–628.

Bordalo, P., Gennaioli, N. and Shleifer, A., 2015, “Salience theory of judicial decisions,”The Journal of Legal Studies vol. 44(S1), S7–S33.

Campello, M. and J.R. Graham, 2013, “Do stock prices influence corporate decisions?Evidence from the technology bubble,” Journal of Financial Economics 107, 89–110.

Charles, Kerwin Kofi, Erik Hurst and Matthew J. Notowidigdo, 2018, “Housing Booms andBusts, Labor Market Opportunities, and College Attendance,” American EconomicReview, 108, 2947-94..

Chetty, Raj, Adam Looney, and Kory Kroft, 2009, “Salience and Taxation: Theory andEvidence,” American Economic Review 99(4): 1145–77.

Dang, T.L., Moshirian, F. and Zhang, B., 2015, “Commonality in News Around the World,”Journal of Financial Economics 116(1), 82–110.

Dickson, L., 2010, “Race and gender differences in college major choice,” The Annals ofthe American Academy of Political and Social Sciences, 627, 108–124.

Fehr, Ernst, and Antonio Rangel, 2011, “Neuroeconomic Foundations of Economic Choice—Recent Advances,” Journal of Economic Perspectives 25(4): 3–30.

Freeman, R.B., 1971, “The Labor Market for College-Trained Man-power” (Cambridge:Harvard University Press).

Freeman, R.B., 1975, “A Recursive Model of the Market for New Lawyers,” The Review ofEconomics and Statistics, 57, 171–179.

Freeman, R.B., 1976, “A cobweb model of the supply and starting salary of new engineers,”Industrial & Labor Relations Review, 29, 236–248.

Fricke, H., Grogger, J., Steinmayr, A., 2015, “Does exposure to economics bring new majorsto the field? Evidence from a natural experiment,” NBER Working Paper.

Gabaix, X., 2014, “A sparsity-based model of bounded rationality”, The Quarterly Journalof Economics, 129(4), pp.1661-1710.

Goldin, C., 2014, “A grand gender convergence: Its last chapter,” American EconomicReview 104, 1091–1119.

Gupta, N. and Hacamo, I., 2018, “Superstar ( and Entrepreneurial ) Engineers in Finance.”Working Paper.

Green, T. Clifton and Byoung-Hyoun Hwang, 2012, “Initial public offerings as lotteries:Skewness preference and first-day returns,” Management Science 58, 432–444.

26

Greenwood, R. and Hanson, S.G., 2015, “Waves in ship prices and investment,” The Quar-terly Journal of Economics, 130, 55–109.

Hastings, J.S., Neilson, C.A., Ramirez, A. and Zimmerman, S.D., 2016, “(Un) informedcollege and major choice: Evidence from linked survey and administrative data,” Eco-nomics of Education Review, 51, 136-151.

Hombert, J. and Matray, A., 2019, “The Long-Term Consequences of the Tech Bubble onSkilled Workers’ Earnings,” Working Paper, Princeton University.

Hoxby, Caroline M., ed., 2003, “The Economics of School Choice,” Chicago: University ofChicago Press.

Hoxby, Caroline M., ed., 2004, “College Choices: The Economics of Where to Go, Whento Go, and How to Pay for It,” Chicago: University of Chicago Press.

Huntington-Klein, Nick, 2016 “(Un) informed College and Major Choice: Verification in analternate setting.” Economics of Education Review 53: 159–163.

James, E., Alsalam, N., Conaty, J. C., To, D. L., 1989, “College quality and future earnings:Where should you send your child to college?” American Economic Review 79, 247–252.

Kahneman, D., 2011, “Thinking, fast and slow,” Macmillan.

Kahneman, D. and A. Tversky, 1979, “Prospect Theory: An Analysis of Decision UnderRisk, Econometrica 47, 263–291.

Kaldor, Nicholas, 1934, “A Classificatory Note on the Determination of Equilibrium,” Re-view of Economic Studies, 1, 122–136.

Lemieux, T., 2014, “Occupations, fields of study and returns to education,” CanadianJournal of Economics 47(4), 1047–1077.

Oreopoulos, Philip, Till von Wachter, and Andrew Heisz, 2012, “The Short- and Long-Term Career Effects of Graduating in a Recession: Hysteresis and Heterogeneity inthe Market for College Graduates,” American Economic Journal: Applied Economics4, 1–29.

Oyer, Paul, 2006, “The Macro-Foundations of Microeconomics: Initial Labor Market Con-ditions and Long-Term Outcomes for Economists,” National Bureau of EconomicsWorking Paper 12157.

Oyer, Paul, 2008, “The Making of an Investment Banker: Stock Market Shocks, CareerChoice, and Lifetime Income,” Journal of Finance, 63, 2601–28.

Rosen, S, 1997, “Manufactured Inequality,” Journal of Labor Economics 15, 189–196.

Sacerdote, B., 2001, “Peer effects with random assignment: Results for Dartmouth room-mates,” Quarterly Journal of Economics, 116, 681–704.

27

Stigler, George J., 1961, “The economics of information,” Journal of Political Economy, 69,213–225.

Stigler, George J., 1962, “Information in the labor market,” Journal of Political Economy,70, 94–105.

Stinebrickner, T. and Stinebrickner, R., 2014, “A major in science? Initial beliefs and finaloutcomes for college major and dropout,” Review of Economic Studies 81, 426–472.

Taylor, S.E. and Thompson, S.C., 1982, “Stalking the elusive “vividness” effect,” Psycho-logical review, 89, 155.

Wiswall, M. and Zafar, B., 2015, “Determinants of college major choice: identification usingan information experiment,” Review of Economic Studies 82, 791-824.

Zafar, Basit, 2013, “College major choice and the gender gap,” Journal of Human Resources48.3: 545–595.

28

29

30

31

32

33

34

35

36

37