Job Hopping and Adverse Selection in the Labor Market∗

Xiaodong Fan† Jed DeVaro‡

November 21, 2019

Abstract

A model of employer learning (both symmetric and asymmetric) about worker ability from
job histories is constructed, and testable implications are derived to detect asymmetric learning
empirically. The model predicts that early-career bad job matches are particularly damaging
when learning is asymmetric. Analysis of NLSY79 data reveals that job hopping is associated
with lower wages for college graduates, controlling for measured ability, labor market expe-
rience, and current job tenure. Suggestive of asymmetric learning, the effect is strongest for
job tenures less than one year and for early-career workers, and mitigated when job hopping
severs matches that were formed during economic downturns.
KEYWORDS: job hopping, asymmetric employer learning, work history
JEL classification: D82, J31, J63

∗ We are grateful for invaluable comments from the editors (Rocco Macchiavello and Daniel Paravisini), three anony-

mous referees, Joseph Altonji, Hugh Cassidy, Hanming Fang, Michael Keane, John Kennan, Rasmus Lentz, Hodaka
Morita, Uta Schönberg, Christopher Taber, Michael Waldman, Alan Woodland, and seminar participants at University
of Wisconsin and University of New South Wales.
† Department of Economics, Monash University, Clayton, VIC, 3800 Australia. Email: xiaodong.fan@monash.edu.
‡ Departments of Management and Economics, College of Business and Economics, California State University, East

Bay, 25800 Carlos Bee Boulevard, Hayward, CA 94542-3066, USA. Email: jed.devaro@csueastbay.edu.
 “...Decide what you feel you’re worth and, once you’re satisfied with your salary level, try to
stay in one place for a while.” (Jessica Kleiman, “8 Pros and Cons of Job Hopping”, Forbes,
8/06/2012.)

1 Introduction
The typical career spans multiple employers, and switching employers after a short job tenure
is particularly common for early-career workers.1 Such “job hopping” is commonly thought to
create adverse selection in the labor market by signaling negative attributes (e.g., disloyalty, im-
patience, a short attention span, low productivity, and a high likelihood of future turnover) that
scare off prospective employers.2 This study’s first objective is to theoretically analyze the role of
job hopping as a signal of low worker quality that depresses wages. A model of employer learning
(both symmetric and asymmetric) about worker ability is constructed in which some workers job
hop.3 The model’s key idea is that when worker productivity reflects both ability and worker-firm
match quality, employers can infer ability using résumé information that predates the current job.
The study’s second objective is to empirically investigate the model’s testable implications. Evi-
dence is shown that employers do consider prior job hopping, and their use of this information
has implications for wages and aggregate efficiency.4
In each of the model’s three periods, worker productivity is the sum of ability and match qual-
ity. At the start of the career no employers observe these productivity components. Workers make
turnover decisions at the end of periods 1 and 2—either remaining with their current employers
or switching firms—recognizing the future signaling role of those decisions. Under asymmetric
learning, a worker’s current employer observes ability and match quality after one period, but
competing employers only observe the worker’s prior turnover decisions. Therefore, workers
make turnover decisions and self-select (or self-label) based on their own productivities. High-
ability workers switch jobs less often than low-ability workers, thereby signaling to competing
employers that they are not “lemons” (Akerlof, 1970). Longer tenures result. Thus, under asym-
1 According to a 2012 Bureau of Labor Statistics study the average American born between 1957 and 1964 held 11.3

jobs from age 18 to 46, and nearly half of those jobs were held from ages 18 to 24. Moreover, Topel and Ward (1992)
finds that a typical worker has 7 jobs during the first 10 years of labor-market experience.
2 Evidence suggests that job hopping creates a stigma that can hurt careers. A 2012 survey by the recruiting software

company Bullhorn finds that 39 percent of 1500 recruiters and hiring managers report “the single biggest obstacle for
an unemployed candidate in regaining employment is having a history of ‘hopping jobs,’ or leaving a company before one year of
tenure.” In case studies of 6 Chicago organizations, Bills (1990) finds that a) many employers use job history information
in making hiring or promotion decisions, and b) the most important component of those decisions is job hopping, which
can be detrimental for both middle-and-low-status positions. Similar results are found in a local restaurant industry
(Bills, 1999).
3 Symmetric learning means that all employers learn about a worker’s ability at the same rate. Asymmetric learning

means that a worker’s current employer has private information about that worker’s ability.
4 Evidence that the relevance of job hopping and short job tenures in the labor market have increased is suggested

in Farber (2008) using Current Population Survey data from 1973 to 2006, which shows that long-term employment
relationships have become less common for men in the private sector and that the same worker group has seen an
increase in “churning” (defined as the proportion of workers in jobs with less than a year of tenure). The 2012 survey
evidence from Bullhorn, cited in the second footnote, suggests that such churning sends a negative signal to employers
about prospective hires.

1
metric learning, competing employers interpret both current and previous job tenures as signals
of worker ability, making wage offers accordingly. In contrast, under symmetric learning, prior
job tenures have no effect on current wages after controlling for observed ability, tenure with the
current employer, and labor-market experience.
The model under asymmetric learning permits analysis of the welfare loss associated with
inefficient turnover, meaning the termination of high-quality matches or the persistence of low-
quality matches, neither of which would occur under symmetric learning. When learning is sym-
metric (or when ability is publicly observed), workers with poor matches switch firms, and these
separations are independent of ability. However, when learning is asymmetric and abilities are
not observable, high-ability workers with poor matches might find it optimal to stay with their
current employers (to avoid being seen as lemons by the market) while low-ability workers with
good matches might switch firms to take advantage of “the benefit of the doubt”. The logic of
the result under asymmetric learning follows from successive application of Greenwald’s (1986)
adverse-selection-turnover argument when multiple turnover decisions in the career are possible.
The model yields three testable implications. First, period-3 wages are independent of the
worker’s job history under symmetric learning, whereas under asymmetric learning job hopping
adversely affects wages. Under asymmetric learning, in period 3, workers who quit when period
1 ends are paid less than those who stay, even when they have identical labor-market experi-
ence and current job tenure. This period-3 wage difference between quitters and stayers, i.e., the
“job-hopping wage penalty”, reflects the signaling effect of job hopping: the market offers lower
period-3 wages to those workers who quit when period 1 ends, because those workers are be-
lieved to have lower ability than stayers. As a corollary, under asymmetric learning the standard
approach of including labor-market experience and current job tenure (but not prior work history)
on the right-hand side of a wage regression yields biased estimates of the returns to tenure and
labor-market experience.
Second, the job-hopping wage penalty under asymmetric learning is larger in magnitude for
early-career than for late-career workers. The prediction arises from a numerical comparative
statics analysis conducted under asymmetric learning. The magnitude of the period-3 job-hopping
wage penalty is increasing in the dispersion of ability and in the correlation between ability and
match quality. The former effect implies a larger job-hopping wage penalty when little is known
about a worker’s ability (meaning higher dispersion in the ability distribution), such as during
one’s early career. Thus, job hopping is more damaging to wages early in the career.
Third, the job-hopping wage penalty arising under asymmetric learning is mitigated dur-
ing business cycle downturns. The idea is that job hopping arising for publicly observable rea-
sons such as aggregate economic shocks provides no information to competing employers about
worker ability and, therefore, should not be penalized by lower competing wage offers. More
generally, anything that makes prospective employers sympathetic to a job-hopping event should
mitigate the job-hopping wage penalty under asymmetric learning (e.g., a quit that severs a match
formed during an economic downturn, since such a separation is relatively more likely to have

2
arisen from a poor match than from low ability).
The empirical work using data from the National Longitudinal Survey of Youth 1979 (NLSY79)
considers college graduates and less educated workers separately. Both education groups are fur-
ther partitioned by the amount of labor-market experience (i.e., “rookies” and “veterans”), given
that the variance of unobserved worker quality might differ by education group and as careers
evolve. For each of the four education-and-experience groups, the empirical analysis measures the
wage change associated with job hopping.5 Empirical discovery of a job-hopping wage penalty,
which is consistent with asymmetric learning, permits learning to have a symmetric component.
That is, an empirical finding of a negative and statistically significant job-hopping wage penalty
allows rejection of the null hypothesis that learning is completely symmetric but permits a blend
of both types of learning.6
The empirical results reveal little or no effect of job hopping on the current wage for less-
educated workers. In contrast, the current wages of college graduates are negatively correlated
with the number of previous quits, after controlling for measured ability, labor-market experience,
and current job tenure. One additional quit in a college graduate’s work history is associated with
a 2.9 percent decrease in the current wage. Furthermore, the job-hopping penalty for college grad-
uates is especially pronounced for job tenures less than one year, as foreshadowed in the Bullhorn
quote in the second footnote. Overall, the model’s first testable implication under symmetric
(asymmetric) learning is consistent with empirical evidence for high-school (college) graduates.
One interpretation for the differential results by education is that, for college graduates, the dis-
persion of unobserved ability is particularly large relative to that of match quality. Another is that,
compared to the less educated, college graduates have a higher correlation between unobserved
ability and match quality.
The job-hopping wage penalty is found to be larger for rookies than for veterans among college
graduates, which supports the model’s second testable implication under asymmetric learning.
Among the less-educated workers, there is no statistically significant job-hopping wage penalty
regardless of the career stage. Consistent with the model’s third testable implication, the job-
hopping wage penalty for college graduates is mitigated when workers enter the labor market
during an economic downturn or have experienced one subsequently (e.g., when there has been
a high average unemployment rate since the point of labor-market entry).
The theoretical literature on the signaling role of one’s current labor-market status, including
current turnover and job assignments, originated with Waldman (1984) and Greenwald (1986).
That literature assumes that potential employers draw inferences about a job applicant’s abil-
5 “Jobhopping” is defined as the number of previous voluntary separations in a worker’s job history. The definition
includes both job-to-job transitions and job-unemployment-job transitions. As shown in section 3, different types of job
hops have the same qualitative effect on current wages, though their magnitudes differ. If a worker changes occupations
but remains with the same employer, this is not defined as a job hop.
6 The present empirical analysis relates to a literature on mover-stayer models, and in particular to Munasinghe and

Sigman (2004), which adds a job mobility measure to the right-hand side of a wage regression estimated on NLSY data.
Unlike the present theoretical framework, the mover-stayer framework focuses only on whether the worker stays or
moves at the most recent opportunity and does not refer to the entire job history.

3
ity based on their current job. In the present study, employers consider the entire work history
rather than just the current job tenure or assignment. Due to complexity, most models that as-
sume asymmetric learning differ from the present study by considering only two periods, which
by assumption rules out the role of work history. That is, when prospective employers consider
hiring decisions at the start of period 2, workers have (at most) experience with only one prior
employer and have not had a chance to job hop.7
The theoretical model connects to the growing literature on employer learning in the labor
market.8 Although the symmetric9 and asymmetric10 learning perspectives are both supported
in the empirical literature, for higher-skilled, more highly-educated workers the evidence points
to the importance of asymmetric learning (e.g., Schönberg (2007) and Kim and Usui (2014) find
evidence of symmetric learning for high-school graduates and asymmetric learning for college
graduates, and DeVaro and Waldman (2012) and Cassidy et al. (2016) find evidence of a signaling
role of promotions only for workers with at least a college degree). Several related studies of asym-
metric employer learning examine the signaling role of job histories, though none of them yields
the present model’s testable implications or explains why college graduates face a job-hopping
wage penalty. Greenwald (1986) develops a three-period, asymmetric employer-learning model
but does not prove the existence of the equilibrium and does not derive (or empirically test) the
testable implications addressed here. Moreover, incorporating match quality permits analysis of
the potential welfare loss caused by inefficient turnover, whereas in Greenwald (1986) turnover
does not imply welfare losses.
7 The theoretical analysis of Golan (2009) studies a three-period model in which abilities are revealed in wages

through bargaining, and Zhang (2011) studies a three-period model in a partial-equilibrium analysis, ignoring work-
ers’ response to the signaling effect of work history. DeVaro and Waldman (2012) analyze a two-period model and
informally discuss an extension to three periods, but because of firm-specific human capital there is no equilibrium
turnover in that model. Similarly, Bernhardt (1995) analyzes a three-period model in which workers are differentiated
by education levels, and promotions signal worker ability, but because the equilibrium exhibits no turnover, there are
no results concerning job hopping.
8 There is a complementary literature on workers’ learning (about their abilities, or about worker-employer match

quality). For example, Papageorgiou (2013) studies how workers’ learning about their multi-dimensional skill type
affects their occupational choices and wages, and Pavan (2011) provides empirical evidence about how career and
firm choice (as in Neal 1999) influence wage growth. Gorry (2016) relates to the present study by using NLSY79 data
to analyze an environment in which workers (as opposed to employers) have incomplete information about match
quality and must rely on their own work histories to learn about new job opportunities and wage dynamics on new
jobs; whereas the present focus is on wage levels, Gorry (2016) focuses on wage volatility on new jobs.
9 Representative studies include Harris and Holmstrom (1982), Farber and Gibbons (1996), Gibbons and Waldman

(1999, 2006), Altonji and Pierret (2001), Lange (2007), Antonovics and Golan (2012), and DeVaro and Morita (2013).
10 In the asymmetric-learning literature, the incumbent employer typically observes the worker’s ability or produc-

tivity, perhaps with noise, whereas competing employers only observe signals, such as a worker’s voluntary or in-
voluntary turnover (Greenwald, 1986; Gibbons and Katz, 1991; Schönberg, 2007; Pinkston, 2009; Hu and Taber, 2011;
Zhang, 2011; Kahn, 2013), the firm’s job assignment or promotion decisions (Waldman, 1984, 1990, 1996; DeVaro and
Waldman, 2012), training levels (Acemoglu and Pischke, 1998), or wages (Golan, 2009).
Empirical support for asymmetric learning is found in Gibbons and Katz (1991), Pinkston (2009), Hu and Taber
(2011), and Kahn (2013), and also in Schönberg (2007) and Kim and Usui (2014) for college graduates. A subset of
the asymmetric-learning literature focuses on the role of job assignments (and particularly promotions) as signals of
worker ability, following Waldman (1984), Bernhardt (1995), Owan (2004), and the related analysis of Milgrom and
Oster (1987). Evidence from a recent empirical literature supports such asymmetric learning via promotions (e.g.,
DeVaro and Waldman 2012, Bognanno and Melero 2016, Cassidy et al. 2016, and DeVaro et al. 2018).

4
2 A Model of Employer Learning from Job Histories
There are three periods. The labor market has a measure one of workers. There is free entry of
firms and perfect competition for workers. A worker can only work for one employer at a time.11
Firms and workers are risk neutral and do not discount the future. Firms can only make spot wage
contracts for one period and cannot commit to future wage contracts, implying a zero-expected-
profit equilibrium condition for new hires.
When period 1 begins, potential employers enter the labor market and simultaneously post
wages. Attention is restricted to the symmetric equilibrium in which firms with the same infor-
mation make identical wage offers. All firms therefore simultaneously post identical period-1
offers that are high enough to ensure zero expected profit. Each worker then randomly chooses
an offer. The market clears instantaneously.
For a worker-firm pair, period-1 output, y1 , is y1 = θ + η1 . Ability, θ, and worker-firm-specific
match quality, η1 , follow a joint distribution f (θ, η ), with supports [θ L , θ H ] and [η L , η H ].12 Assume
that f (θ, η ) is log-concave and that the marginal distribution Fη (θ ) , ∀η is not proportional to ecx ,
with c > 0, and σθ + ρση > 0, where ρ is the correlation between θ and η.13 Further assume that
θ̄ − θ L ≤ η̄ − η L and θ̄ + η̄ ≤ (θ H + θ L + η H + η L ) /2.14
Match quality is an experience good, so the worker does not observe it before accepting the
wage offer. The next two subsections share the preceding features but differ according to what
information employers learn about θ, η, or y over time.

2.1 Symmetric Learning

Under symmetric learning, each worker’s θ is publicly revealed when period 1 ends, along
with η1 . Let V1 denote period-1 wage offers. When period 1 ends, all firms simultaneously post
two period-2 wage offers—an “incumbent offer”, V2 (θ, η1 ), for their own workers and a “market
offer”, V2o (θ ), for other workers—that are observed by all workers.15
Each worker accepts a period-2 offer. If the incumbent offer is accepted, period-2 output is
y2 = y1 = θ + η1 .16 If the market offer is accepted, a new match quality, η2 , which is independent
11 “Employer”, “firm”, and “job” are used interchangeably.
12 Any of the bounds can be infinite. Previous research generates equilibrium turnover by assuming exogenous
separations (Greenwald, 1986) or stochastic utility shocks (Acemoglu and Pischke, 1998; Schönberg, 2007). Generating
turnover via a match-quality component facilitates welfare analysis of the efficiency loss from asymmetric learning.
13 Assuming σ + ρσ > 0 guarantees that cov ( y, θ )=cov ( θ, θ )+ cov ( η, θ )=σ2 + ρσ σ > 0. Otherwise the productivity,
θ η θ θ η
y, becomes negatively correlated with θ, which is not of general interest.
14 The first says that the ability distribution’s range should be no larger than that of the match-specific productivity

distribution, which is reasonable as the latter could be unbounded. The second requires that the distribution of the
summation of ability and match quality cannot be severely negatively skewed. Examples of distributions satisfying
both assumptions include the uniform, left-truncated normal, and exponential distributions with θ̄ ≤ η̄ (Bagnoli and
Bergstrom, 2005).
15 The market offer is independent of the previous match quality (which becomes irrelevant upon separation). More-

over, the results are insensitive to allowing the incumbent firm to make a counteroffer after observing the highest
market offer, since in equilibrium the latter is correctly anticipated.
16 A productivity increase, which may be interpreted as firm-specific human-capital accumulation, is included in

Gibbons and Katz (1991) but omitted here for simplicity. Including it would not change the analysis unless it is large

5
of η1 , is drawn from the joint distribution f (θ, η ) and is publicly revealed when period 2 ends.
The period-2 output for the new matched worker-firm pair is y2 = θ + η2 .


When period 2 ends, two wage offers are made: the incumbent offer, V3 θ, η j , j ∈ {1, 2}, and
the market offer, V3o (θ ). Each worker chooses the highest offer.17 Assuming the worker switches
jobs after period 2, a new match quality, η3 , independent of η1 and η2 , is drawn from f (θ, η ) for
the newly-formed pair and is revealed to all parties when all workers retire from the labor market
at period 3’s end.
Proposition 1 describes the equilibrium.18 Denote the conditional expectation η̄θ = E {η |θ }
and the unconditional expectations η̄ = E {η }, θ̄ = E {θ }.
Proposition 1: In the symmetric-learning model, when period 1 ends, a worker of ability θ stays with
∗ , where
the incumbent firm if and only if η1 ≥ ηθs1

∗ 1
ηθs1 = η̄θ + E {(η2 − η̄θ ) · 1 {η2 ≥ η̄θ }} , (1)
2

The market wage offer is

V2o (θ ) = θ + η̄θ + E {(η2 − η̄θ ) · 1 {η2 ≥ η̄θ }} , (2)

and the incumbent wage offer is

V2 (θ, η1 ) = min {θ + η1 + (η1 − η̄θ ) · 1 {η1 ≥ η̄θ } , V2o (θ )} . (3)

When period 2 ends, a worker stays with the incumbent firm if and only if η j ≥ η̄θ , j ∈ {1, 2}. The
market wage offer is
V3o (θ ) = θ + η̄θ . (4)

The incumbent wage offer is

V3 θ, η j = min θ + η j , V3o (θ ) = θ + min η j , η̄θ .

 
(5)

In both periods 2 and 3, the accepted incumbent offer matches the market offer.
Given free entry to the labor market, the period-1 wage is determined by the zero-expected-profit condi-
tion,

∗
V1 = θ̄ + η̄ + E {[η1 − η̄θ − E {(η2 − η̄θ ) · 1 {η2 ≥ η̄θ }}] · 1 {η1 ≥ ηθs1 } + (η1 − η̄θ ) · 1 {η1 ≥ η̄θ }} (6)

Proposition 1 states that under symmetric learning the worker’s turnover decision depends on
enough to shut down turnover completely.
17 A firm retains no information about a previous worker once a match severs. Therefore, if a worker joins the same

employer in periods 1 and 3 (but a different one in period 2), then in period 3 the employer does not recall the worker’s
period-1 ability, and a fresh match quality is drawn. This point is irrelevant under symmetric learning but not under
asymmetric learning.
18 The appendix contains all proofs.

6
ability and the match quality realization. Workers with poor matches switch firms, and those with
good matches stay put. Furthermore, the worker’s period-3 wage is independent of the turnover
decision at period 1’s end. In equilibrium, a worker of ability θ receives a period-3 wage of θ + η̄θ
regardless of job-hopping history.19 There is no adverse selection in the labor market, and job
hopping has no signaling effect.
The sum of the market wages in periods 2 and 3 equals a worker’s expected output in a new
firm over both periods. Period-2 expected output is θ + η̄θ . Period-3 expected profit is the expected
output E {θ + η2 } net of the expected incumbent wage offer E {V3 (θ, η2 )} if the worker stays,
resulting in an option value to hiring in period 2, E {(η2 − η̄θ ) · 1 {η2 ≥ η̄θ }}. Hence, the zero-
expected-profit condition yields the period-2 market wage offer (2).
The period-2 incumbent offer never exceeds what competing firms would pay, namely V2o (θ ),
which is the second argument in the min function in equation (3). Moreover, the incumbent em-
ployer would never pay more over periods 2 and 3 than the worker’s productivity over both pe-
riods. Therefore, the first argument in the min function in equation (3) arises from a similar logic
to that in the preceding paragraph. The key difference is that the worker’s match quality with the
incumbent employer is observed when period 1 ends, which causes the second and third terms
on the right-hand side of equation (3) to differ from their counterparts in equation (2). Specifically,
the second term now is realized match quality rather than expected match quality, and the third
term no longer requires an expectation operator.
The worker stays put in period 3 if η j ≥ η̄θ , implying that the incumbent makes positive
period-3 profit. This has implications for the incumbent’s period-2 wage offer. Specifically, the
maximum period-2 amount that the incumbent firm would offer (which is the first term in the
min function in equation (3)) involves incurring a negative period-2 profit that exactly offsets the
anticipated positive period-3 profit.

2.2 Asymmetric Learning

Under asymmetric learning, the incumbent employer privately observes θ and η1 when period
1 ends. Competing firms do not observe any information about the employment relationship—
ability, match quality, output, or wage. Let W1 denote the period-1 wage offer. Compared to the
preceding analysis, the new information structure changes wage offers for periods 2 and 3.
When period 1 ends, the period-2 incumbent wage offer becomes W2 (y1 ) because the incum-
bent firm cares only about productivity and, therefore, offers a wage as a function of y1 only, rather
than θ or η1 . Let W2o denote the market wage offer. A worker quits if and only if W2 (y1 ) < W2o ,
and the worker stays if indifferent. Competing firms can only make inferences about ability from
observed quitting behavior. Workers are described as “stayers” and “quitters” according to their
decisions at period 1’s end, and they are indexed by s and q henceforth. Figure 1 illustrates the
timing.
19 “Job hopping history” (or “work history”) refers to the turnover history prior to the current job.

7
 The equilibrium is solved by backward induction. At period 3’s start, each worker’s work
history (i.e., sequence of prior employers) is observed by all firms. Since employers are identical,
the work history reveals whether or not the worker has switched employers after period 1.20
First, suppose that the worker switches jobs after period 1. Then a new match quality, η2 , is
revealed (along with the worker’s ability, θ) to the new matched pair when period 2 ends, whereas
competing firms observe neither. Alternatively, suppose that the worker stays beyond period 1.
Then both η1 and θ are already known to the matched pair since period 1’s end. When period 2
ends, each firm simultaneously posts three wage offers. A firm matched with a stayer in period
2 posts a wage offer of W3 (y1 , s) to its stayer, W3o (s) to other stayers, and W3o (q) to quitters. A
firm matched with a quitter in period 2 posts a wage offer of W3 (y2 , q) to its own quitter, W3o (s)
to stayers, and W3o (q) to other quitters.
Each worker then chooses the firm that makes the highest period-3 offer, remaining with the
current employer in the event of a tie. If a stayer continues to stay with the incumbent firm
(becoming a “stayer-stayer”), period-3 productivity coincides with period-2 productivity, i.e.,
y3 = y2 = θ + η1 . Similarly, if a quitter stays (becoming a “quitter-stayer”) then period-3 pro-
ductivity is y3 = y2 = θ + η2 . Period-3 productivity after a firm change is y3 = θ + η3 .
When period 1 ends, a worker of ability θ with output y1 quits if and only if

W2 (y1 ) + max {W3 (y1 , s) , W3o (s)} < W2o + Eη2 |θ [max {W3 (y2 , q) , W3o (q)}] [ IRθ ], (7)

where the left-hand side is the sum of wages from periods 2 and 3 if the worker stays, while the
right-hand side is the expected sum of wages if the worker quits. The expectation is with respect
to the distribution of η2 given θ. This condition is denoted IRθ and its complement IRθ .21
When period 2 ends, for stayers there exists a threshold value, ŷ2 (s), satisfying

W3o (s) = ŷ2 (s) =⇒ ŷ2 (s) = E θ + η3 |y1 < ŷ2 (s) , IRθ
 
(8)

such that the stayer quits if and only if y1 < ŷ2 (s). The incumbent firm offers

W3 (y1 , s) = min {y1 , ŷ2 (s)} , (9)

and the accepted incumbent offer, W̃3 (y1 , s), matches the market offer, i.e., W̃3 (y1 , s) = W3o (s) =
ŷ2 (s).
Similarly, for quitters there exists a threshold value, ŷ2 (q), satisfying

W3o (q) = ŷ2 (q) =⇒ ŷ2 (q) = E [θ + η3 |y2 < ŷ2 (q) , IRθ ] , (10)
20 Thus, the market interprets a move from firm A to B and from B to A identically.
21 The complement replaces < with ≥ in IRθ .

8
such that the quitter quits if and only if y2 < ŷ2 (q). The incumbent firm offers

W3 (y2 , q) = min {y2 , ŷ2 (q)} , (11)

and the accepted incumbent offer, W̃3 (y2 , q), matches the market offer, i.e., W̃3 (y2 , q) = W3o (q) =
ŷ2 (q).
The IRθ condition becomes

W2 (y1 ) < W2o + ŷ2 (q) − ŷ2 (s) [ IR]. (12)

which is independent of θ. When period 1 ends, the incumbent firm offers W2 (y1 ) ≤ y1 +
(y1 − ŷ2 (s)) · 1 {y1 ≥ ŷ2 (s)}. Therefore, there exists a threshold value, ŷ1 , satisfying

ŷ1 + (ŷ1 − ŷ2 (s)) · 1 {ŷ1 ≥ ŷ2 (s)} = W2o + ŷ2 (q) − ŷ2 (s) , (13)

such that the worker quits if and only if y1 < ŷ1 when period 1 ends. The incumbent firm offers

W2 (y1 ) = min {y1 + (y1 − ŷ2 (s)) · 1 {y1 ≥ ŷ2 (s)} , ŷ1 + (ŷ1 − ŷ2 (s)) · 1 {ŷ1 ≥ ŷ2 (s)}} (14)

and the accepted incumbent wage offer is

W̃2 (y1 ) = ŷ1 + (ŷ1 − ŷ2 (s)) · 1 {ŷ1 ≥ ŷ2 (s)} . (15)

This wage offer is determined by the marginal worker who has productivity ŷ1 and is indiffer-
ent between staying and quitting when period 1 ends. The first term is the period-2 contribution
of output ŷ1 , and the second term is the output ŷ1 net of the period-3 wage, ŷ2 (s), in the event that
the worker stays again in period 3.
When period 1 ends, the zero-profit condition determines the wage W2o , i.e.,

W2o = E [θ |y1 < ŷ1 ] + η̄ + E {[y2 − ŷ2 (q)] · 1 (y2 ≥ ŷ2 (q)) |y1 < ŷ1 } . (16)

The intuition for this expression parallels that for equation (2). The three terms on the right-
hand side are the contribution of ability, match quality, and option value, as explained at the end
of the preceding subsection. Now, however, ability is unobservable, so the first term becomes the
expectation of ability conditional on the worker quitting when period 1 ends. Similarly, the second
term becomes the unconditional expectation of match quality. As in the preceding subsection, the
third term’s conditioning event is that the worker (having switched firms in period 2) remains
with the new firm in period 3.

9
 Equations (8) and (10) become22

ŷ2 (s) = E [θ |ŷ1 ≤ y1 < ŷ2 (s)] + η̄, (17)

ŷ2 (q) = E [θ |y2 < ŷ2 (q) , y1 < ŷ1 ] + η̄. (18)

Proposition 2 summarizes the equilibrium characterized by equations (13), (17), and (18).
Proposition 2: There exists at least one group of threshold values, (ŷ1 , ŷ2 (s) , ŷ2 (q))∈ [θ L + η L , θ H + η H ] ×
+ η̄ × [θ L + η L , y∗ ], satisfying equations (13), (17), (18), and ŷ1 < ŷ2 (s), where y∗ is the unique
y∗ , θ̄
 

solution to
y∗ = E [θ |θ + η1 < y∗ ] + η̄. (19)

When period 1 ends, a worker stays if and only if y1 ≥ ŷ1 ; when period 2 ends, a stayer stays again if
and only if y1 ≥ ŷ2 (s), and a quitter stays if and only if y2 ≥ ŷ2 (q).
When period 1 ends, the market offer is characterized by equation (16), the incumbent offer by equation
(14), and the accepted incumbent offer by equation (15). When period 2 ends, the market offers are W3o (s) =
ŷ2 (s) for stayers and W3o (q) = ŷ2 (q) for quitters. The incumbent offers are characterized by equation (9)
for stayers and equation (11) for quitters. The accepted offers match the market offers.
The period-1 wage is determined by the zero-profit condition,

W1 = θ̄ + η̄ + E {[(y1 − ŷ1 ) + (y1 − ŷ2 (s)) · 1 (y1 ≥ ŷ2 (s))] · 1 (y1 ≥ ŷ1 )} . (20)

Competing firms observe two turnover signals from each worker. Focusing on stayers, the
subgame involving only periods 2 and 3 is identical to a two-period standard “lemons” model
where competing firms observe only one signal from each worker when period 2 ends, namely
the turnover decision (stay or quit).23 That is, each stayer sends one signal to competing firms,
and competing firms observe only one signal from each stayer. During period 3, stayers who quit
are paid their expected productivity, so the firm’s expected profit is zero. Incumbent firms match
the outside option up to the worker’s productivity level, so that stayers are paid the same during
period 3 regardless of whether they stay or quit.24
Similar reasoning applies to quitters. The difference between stayers and quitters is the distri-
bution of their abilities, conditional on their earlier decisions to stay or quit. Stayers are expected
to have higher abilities, on average, than quitters. Given that quitting sends competing firms a
signal that the worker is more likely to be a lemon, workers with relatively high productivity (i.e.,
ŷ1 ≤ y1 < ŷ2 (s)) choose to stay when period 1 ends and quit when period 2 ends, to differentiate
themselves from those with lower productivity (i.e., y1 < ŷ1 ) who would choose to quit when
period 1 ends. Thus, workers use staying (when period 1 ends) as a signaling device which allows
22 Since η3 is independent of η1 and η2 , the following expression holds:
E [η3 |y2 < ŷ2 (q) , y1 < ŷ1 ] = E [ E [ E [η3 |η2 < ŷ2 (q) − θ, η1 < ŷ1 − θ, η1 ] , η2 ]] = η̄
23 See
the upper branch of Figure 1.
24 Including
random utility shocks as in Acemoglu and Pischke (1998) and Schönberg (2007) would cause incumbent
offers to vary with productivity, without affecting the main results.

10
employers to distinguish between stayers and quitters. Therefore, the firm’s belief is consistent,
as equilibrium requires.
Furthermore, the first turnover decision cannot be perfectly inferred from the current labor
market experience and current tenure. For example, those who join new firms in period 3 all have
the same current experience (three periods) and current tenure (one period), but they potentially
differ in their work histories (i.e., some may be stayers and others quitters). Thus, the number of
prior employers plays an independent role in determining the worker’s period-3 wage.

2.3 Comparative Statics

Simulations illustrate the model’s comparative statics under both learning environments. Abil-
ity and match quality are assumed to be drawn from a joint normal distribution
     
θ 0 σ2 ρσ ρσ ρσ
     
 ∼ N  0 1 0 0 
 η1     ρσ 
 , 
 η   0   ρσ 0 1 0  
 2     
η3 0 ρσ 0 0 1
The match quality, η, is assumed to be independent across different periods, with a standard devi-
ation normalized to unity.25 There are two free parameters: the standard deviation of θ (denoted
σ) and the correlation between θ and ηi (denoted ρ).
Under symmetric learning the conditional probability density function of ηi given θ is
 
ρθ
ηi | θ ∼ N , 1 − ρ2
σ

Proposition 1’s threshold values are then:

r
∗ ρθ 1 1 − ρ2
ηθs1 = +
σ 2 2π
ρθ
η̄θ =
σ

The period-3 wage difference between quitters and stayers in the new job is always zero, ∆W3SL =
V3o (q)-V3o (s)=V3o (θ ) − V3o (θ ) = 0.
Under asymmetric learning, the equilibrium can only be solved numerically for different com-
bination of σ and ρ. The results of varying these two parameters are then shown. First, σ = 1
is fixed, and ρ ∈ (−1, 1) is varied.26 The two figures in the top panel of Figure 2 show how the
three equilibrium threshold values and the period-3 wage difference between quitters and stay-
 
25 This normalization is without loss of generality. Note that y = θ + η can be rearranged to obtain y = ση θ
+
η
=
ση ση
ση (θ 0 + η0 )
such that ση 0 = 1, given that ση is a known constant.
26 When
θ and η are perfectly correlated (i.e., ρ = ±1), this model is equivalent to one in which there is no η, no
turnover, and a standard deviation of θ of σ ± 1.

11
ers, ∆W3AL = W3o (q) − W3o (s), change with the correlation ρ. The wage difference between quitters
and stayers appears to be smallest (in level) when θ and η are independent. When ρ ≥ 0, ∆W3AL 
 

increases as ρ increases; when ρ ≤ 0, ∆W3AL  increases as ρ decreases. This is explained by the

 

signaling effect of job hopping. Intuitively, quitting offers the opportunity for a new η.
Observing the job-hopping history, potential employers estimate a quitter’s ability level, θ,
which suffers from adverse selection, and a new match draw, ηi . The correlation between match
quality and ability influences the inference of θ through the current match quality. It also impacts
the new match quality draw. As the correlation increases, the adverse selection in θ becomes
more severe, and the new match-quality draw becomes more dependent on θ, further increasing
the adverse selection of θ. On the other hand, when η and θ are independent, the η does not
exaggerate the adverse selection in θ. Thus, the wage difference between quitters and stayers is
smallest when ρ = 0, and as ρ gets further from 0 in either direction the wage difference increases.
Next, ρ = 0 is fixed and σ is varied, with results shown in the bottom panel of Figure 2. As σ
increases, the period-3 wage difference between quitters and stayers increases, because a higher σ
leads to a higher penalty for quitting; E (θ |θ < y − η ) decreases when σ increases for any given y or
η. Intuitively, learning about each worker’s ability from their work history is quicker if the ex ante
distribution of ability has less dispersion. In the extreme, if the ability distribution is degenerate,
then ability is learned immediately, which is equivalent to the symmetric learning model where
there is no job-hopping wage penalty.
In summary, under asymmetric learning the period-3 job-hopping wage penalty increases if
the variance of ability increases or the match quality becomes more correlated with ability.27
The preceding point about the variance of ability provides insight for interpreting one of the
forthcoming empirical results. Workers with poor period-1 matches are less likely to hop in search
of another match draw if they have high ability. There are, therefore, two effects on the observed
wage after a job hop: (i) lower expected ability, (ii) higher average match quality. Effect (i) pushes
the post-hop wage down relative to the pre-hop wage, whereas effect (ii) pushes it up. The net
effect is a job-hopping wage penalty if effect (i) is present and large, which happens when the vari-
ance of ability is large. Forthcoming empirical evidence documents a job-hopping wage penalty
for college graduates but not for the less educated. The absence of a penalty for the less-educated
group is consistent with symmetric learning, which is the extreme case of zero variance in ability.
Thus, an interpretation of the empirical evidence is that the variance of ability is sufficiently small
for the less-educated workers that a wage penalty is indiscernible in the data.
27 In the simulations, when period 1 ends, workers are sorted into stayers (s) and quitters (q), with expected ability
E (θ |q) < E (θ ) < E (θ |s), and the within-group variance decreases, σ (θ |s)< σ (θ ) and σ (θ |q) < σ (θ ). When period 2
ends, the stayer group is further sorted into stayer-stayers (ss) and stayer-quitters (sq), with expected ability E (θ |sq) <
E (θ |s) < E (θ |ss), and the further decreased within-group variation, σ (θ |ss)< σ (θ |s) and σ (θ |sq) < σ (θ |s). Similarly,
quitters are sorted into quitter-stayers and quitter-quitters, with E (θ |qq) < E (θ |q) < E (θ |qs), σ (θ |qs)< σ (θ |q) and
σ (θ |qq) < σ (θ |q).

12
2.4 Testable Implications

Proposition 3 states the main result summarizing the signaling effect of work history on the
current wage.
Proposition 3: Under symmetric learning, period-3 wages are independent of work history. Under
asymmetric learning, quitters are paid less than stayers, even when they have identical labor market experi-
ence and current job tenure, i.e., W3o (q) − W3o (s) < 0.
To illustrate, consider two workers, A and B, with the same ability, but suppose that worker
A’s period-1 match quality exceeds worker B’s. Worker A is indifferent between staying and quit-
ting. Given the tie-breaking rule, he stays. Thus, worker B quits. So by period 2’s end, worker A
was continuously employed by one firm, whereas worker B worked for two firms. Suppose that
both workers’ period-2 match qualities are such that they switch firms when period 2 ends. Even
though both workers have the same ability, observed current job tenure, and labor-market expe-
rience, the market views them differently. Conditional on the observed job history, worker B is
expected to have lower ability than worker A, so worker A’s period-3 wage offer exceeds worker
B’s.
The intuition for this result is that prospective employers, when deciding what wage to offer,
cannot discern to what extent the worker’s history of job hopping is due to low match quality
(which would not justify a lower offer) versus low ability (which would justify a lower offer). If
a prospective employer knew for sure that a worker job hopped because of low match quality,
then since match qualities are independent across firms, the employer would not downgrade the
worker’s offer. Similarly, if a prospective employer knew that a worker job hopped because of
low ability, the employer would surely downgrade the offer. But since the prospective employer
cannot observe which of these two cases applies, some weight is assigned to the possibility that
the job hopping occurred because of low ability. The result is a job-hopping wage penalty.
There is a difference between a job-hopping wage penalty and the productivity effect of job
hopping. A worker with ability θ who quits when period 2 ends receives a new period-3 match
quality which is independent of prior match qualities, job-hopping history, or the learning envi-
ronment. In other words, job hopping neither helps nor hinders productivity for any particular
worker, regardless of the learning mechanism. If the worker’s ability is observable when period
2 ends (symmetric learning), then the period-3 wage at the new job should be independent of the
job-hopping history, after controlling for ability. However, in the asymmetric learning environ-
ment the entire job-hopping history serves as a signal that is effective because workers of different
abilities and match qualities self-select into different job-hopping histories in a labor-market equi-
librium where potential employers hold consistent beliefs.
The first testable implication follows from proposition 3 and is based on a difference between
the two types of learning on the marginal effect of job hopping on the period-3 wage, controlling
for labor-market experience, current job tenure, and observed ability. Under asymmetric learning,
that marginal effect is negative, implying a job-hopping wage penalty.28 Under symmetric learn-
28 Although a job-hopping wage penalty is predicted under asymmetric learning, other countervailing mechanisms

13
 ∗ > η̄ , there is no
ing, that marginal effect is not identified if the model is taken literally. Given ηθs1 θ
situation in which one stays when period 1 ends but quits when period 2 ends. Therefore, current
job tenure and number of prior employers are collinear, and their effects cannot be simultaneously
identified given labor market experience.29 Strictly speaking, this means that the marginal effects
of job hopping under asymmetric and symmetric learning cannot be compared.
Henceforth, however, a generalization of the model is assumed, which resolves the identi-
fication problem while leaving the results qualitatively unchanged. Suppose that with a small
positive probability the worker-firm match dissolves regardless of productivity (perhaps because
a worker’s spouse finds a good job in another location), as in Gibbons and Katz (1991), Greenwald
(1986), Waldman (2013), and Cassidy et al. (2016), or the worker draws a new match quality even
without switching jobs (e.g., the worker might get a new boss or be assigned to a new task), as
in Acemoglu and Pischke (1998) and Schönberg (2007). Then in period 3, a current tenure of one
could map to either one or two previous firms, and the marginal effect of job hopping is negative
under asymmetric learning and zero under symmetric learning.
The second testable implication derives from the observation that as a worker’s career ad-
vances, the market acquires more information about this worker and can make more precise infer-
ences about ability. This is equivalent to stating that the variance of unobservable ability decreases,
implying a smaller job-hopping wage penalty for later-career workers than for earlier-career ones.
Thus, if asymmetric learning is present, it should be more pronounced for “rookies” than for “vet-
erans”.
The third testable implication concerns employers’ tendency to be sympathetic when there is
reason to believe that a job hopper is likely to have separated because of a bad match (as opposed
to low ability). For example, even though prospective employers cannot observe whether job hop-
ping occurred because of low ability or because of a bad match, the relative likelihood of the latter
increases if the match was formed during an economic downturn. The wage penalty attached
to job hopping under asymmetric learning should, therefore, be lower when workers enter the
labor market during a bad year. This is not true under symmetric learning, because in that case
job hopping has no effect on the current wage anyway (regardless of economic conditions at the
point of labor market entry). An economic downturn serves as an observable signal revealing that
the reason for job hopping is economic conditions rather than low ability. This narrows down the
uncertainty, similar to a smaller dispersion in the ability distribution. The previous subsection
might be present in the data. For example, experience in multiple jobs spanning a number of employers might provide
career-enhancing skills that the labor market values. Such experience broadens information and skill sets and may
create versatile, jacks-of-all-trades who are well equipped for success, particularly in managerial and executive roles, as
in the entrepreneurship model of (Lazear, 2005). The following conclusion from a 2014 survey by CareerBuilder reflects
opposing perspectives on how employers view job hoppers: “43 percent of employers will not consider a candidate having
short tenures with several employers, while 53 percent of employers said job-hoppers tend to have a wide range of expertise and
can adapt quickly (51 percent).” Therefore, it is an empirical question whether a job-hopping wage penalty exists and, if
it does, how large it is and what determines its size.
29 When learning is symmetric, in period 3, current tenure of one maps to two previous firms; current tenure of two

maps to one previous firm; current tenure of three maps to zero previous firms. In the asymmetric-learning model,
however, current tenure of one in period 3 could map to one (for stayers) or two (for quitters) previous firms.

14
shows that the job-hopping wage penalty decreases with σ. Thus, it would be expected under
asymmetric learning that the economic downturn would mitigate the job-hopping wage penalty.
The preceding testable implication represents one application of a more general result from
the theoretical model. Under asymmetric learning, the extent to which the market punishes job
hopping should hinge on the relative dispersions of the match quality and ability distributions,
both of which are publicly observed. Consider the extreme cases. When the dispersion of match
quality is small relative to the dispersion of ability, employers know that job hopping was probably
due to low ability, and a large wage penalty results. Alternatively, when the dispersion of ability is
small relative to the dispersion of match quality, employers know that job hopping was probably
due to a bad match.

3 Empirical Analysis
The empirical analysis uses the 1979 cohort of the National Longitudinal Survey of Youth
(NLSY79). Consider the following regression:

yit = Xit0 β + W Hit0 γ + ε it , (21)

where yit is the natural logarithm of worker i’s current hourly wage at time t, the vector Xit de-
notes the set of both time-invariant individual characteristics and time-varying individual current
employment status (including labor-market experience and tenure with the current firm), ε it is an
i.i.d disturbance term with mean zero, and the vector W Hit contains measures of work history
such as job hopping. The number of prior voluntary quits (which is referred to as the “job-hopping
measure”) is a summary statistic to proxy for the full work history.30
The first testable implication’s null hypothesis is that there exists no asymmetric learning
(i.e., learning is entirely symmetric, and the coefficient of the job-hopping measure is zero) ver-
sus the alternative hypothesis that asymmetric learning is present (so that the coefficient of the
job-hopping measures is negative). The second and third testable implications concern how any
evidence of asymmetric learning that is found varies under certain conditions. Evidence of asym-
metric learning should be weaker (i.e., the negative coefficient of the job-hopping measure should
be smaller in magnitude) in workers’ late-career stages or when they enter the labor market during
bad economic times.
If the regression includes current job tenure, labor-market experience, and job-hopping mea-
sures on the right-hand side, omitting the job-hopping variables should affect the coefficients on
the other two variables under asymmetric learning (due to omitted variable bias) but not under
symmetric learning. Measures of labor-market experience and current job tenure are regularly
30 “Work
history” refers to one’s entire labor-market experience prior to joining the current employer. Including a
complete set of measures describing the entire work history and capturing the timing of each separation is impractical.
No currently available data set is large enough to accommodate even a 10-period analysis. A work history of N periods
of consecutive employment would generate 2 N −1 different turnover paths, given that a worker starts with a particular
employer in period 1 and can switch employers in any subsequent period.

15
included in wage regressions, with researchers’ interest focusing on the returns to both, but such
regressions typically do not include prior job hopping on the right-hand side. The theoretical anal-
ysis predicts that such omissions are not problematic under symmetric learning but lead to bias in
the coefficients of interest when learning is asymmetric.

3.1 Data and Measures

The NLSY79 is a nationally representative sample of men and women who were 14-22 years
old in 1979 and who were interviewed annually up to 1994, and biennially thereafter. Attention is
restricted to males in the private sector from 1979 to 2008. The data include each worker’s entire
(weekly) work history from the point of labor-market entry. The analysis is restricted to periods
after each individual is observed entering the labor market for the first time, which is defined
as “primarily” working for at least two consecutive years (Schönberg, 2007).31 One job is chosen
with the longest aggregated hours worked for each month for each individual, identifying the
employers for this job and linking them in continuous years to form the job turnover history. To
ensure a complete work history, all subsequent periods are eliminated after the first gap of three
or more years absent from the labor market for each individual.
The job-hopping measure is the number of previous voluntary separations, which is computed
for each year in which a worker is observed. Observations involving job switches due to layoffs
or firings are omitted because they are not modeled in the theoretical analysis. The regression also
include two work-history measures summarizing the worker’s current labor-market status: (1)
length of current job tenure, i.e., total weeks worked for the current employer; (2) experience, i.e.,
total weeks worked for all employers since first entering the labor market.
The wage measure is the hourly pay rate. Its natural logarithm is the dependent variable. Since
wage information is only collected annually, one observation is kept for each job in each year (from
the monthly employment-history data) and a work history is constructed by year. Hourly wages
below 1% ($2.69) or above 99% ($77.67) (in 2004 dollars, both inclusive) are dropped from the
regression, though tenure is still included in the work history construction.
The NLSY’s ability measure is the Armed Forces Qualification Test (AFQT) score, which is
used as a control, as in Neal and Johnson (1996), Altonji and Pierret (2001), Schönberg (2007) and
Pinkston (2009). It is standardized to a mean of zero and a standard deviation of one after filtering
out age-specific fixed effects among males. The time-invariant individual characteristics relevant
to wage determination include dummy variables for race and U.S. residency status at the age
of 14. The time-varying controls include years of schooling and its interaction with experience,
the interaction between AFQT score and experience (Altonji and Pierret, 2001), year dummies
and their interactions with years of schooling, a dummy variable for urban residency, a dummy
variable indicating the interview year (Schönberg, 2007), and dummy variables for occupation
31 An individual is said to be primarily working if they work at least 1000 hours per year without changing the highest

grade of education completed. Of 6403 males, 625 are excluded who are never observed to meet these criteria.

16
and industry codes at the one-digit level.32
Two education groups are considered separately (college graduates and those with a high-
school degree or less (as in e.g., Schönberg, 2007; Arcidiacono et al., 2010)) because the learning
mechanism might differ across education groups and because the prior literature on employer
learning has found asymmetric learning to be particularly relevant for more highly educated
workers.33 College graduates are defined to have sixteen or more years of schooling, and the
less educated have at most twelve years of schooling.34 The analysis sample consists of 665 col-
lege graduates with 13, 151 observations, and 3523 individuals with at most a high-school degree
with 67, 191 observations.35 Table 1 displays descriptive statistics for that sample.
Empirical analysis for the second testable implication is conducted separately for “rookies”
(with five or fewer years in the labor market) and “veterans” (with more than five years in the labor
market). The measure of labor-market conditions for addressing the third testable implication is
the unemployment rate in the year of labor-market entry, and an alternative measure is the average
unemployment rate weighted by the incidence of job quitting years.

3.2 Empirical Evidence

Table 2 reports the results from a regression of job hopping on the AFQT score, experience
and its square, and individual characteristics.36 For college graduates, workers with higher AFQT
scores hop jobs less often, leave fewer jobs voluntarily, and have experienced fewer layoffs and
firings. All associations are statistically significant at conventional levels. In contrast, for the less
educated the association between job hopping and the ability measure is negative but statistically
insignificant. Given that the AFQT score is a partial ability measure, these results are consistent
with job hopping being an ability signal for college graduates but not for the less educated.
To address the first testable implication, Table 3 displays the results of log-wage regressions
for college graduates (column 1) and the less educated (column 4), controlling for AFQT score,
current tenure, experience, and the product of AFQT and experience. Of primary interest is the
32 As in Schönberg (2007), seven occupations are distinguished: professional, technical, and kindred; managers and
administrators; sales, clerical, and kindred; craftsmen and kindred; operatives; laborers; and service. Twelve industries
are distinguished: mining; construction; manufacturing; transportation, communications, and other public utilities;
wholesale trade; retail trade; finance, insurance, and real estate; business and repair services; personnel services; enter-
tainment and recreation services; professional and related services; and public administration.
33 Prior empirical evidence suggests that asymmetric learning is particularly relevant for highly-educated workers

(e.g., Schönberg 2007, DeVaro and Waldman 2012, Kim and Usui 2014).
34 Unreported results for those with some college (defined as 13 to 15 years of education) are available upon request;

there are 874 individuals with 21, 372 observations in that group. Results for that group are similar to (but smaller in
magnitude than) those for college graduates.
35 The analysis sample excludes 625 individuals who never experienced the transition to the labor market according

to the definition in footnote 31, and 310 who entered the labor market before age 15 or after age 27. Individuals with
missing AFQT scores are dropped (357), as are those with missing U.S. residency status at age fourteen (11), those who
are farmers or workers in the agriculture, forestry and fisheries industry, or in the military, or with missing occupation
or industry codes (11), and those with missing urban residency status (28). For each individual, one observation is
kept per job per calendar year, in contrast to every interview year as in Schönberg (2007). This matters after 1994, when
NLSY interviews were conducted every other year. Using calendar years allows for the inclusion of jobs that ended in
a non-interview year.
36 Each worker’s last observation is retained.

17
marginal effect of the number of previous quits. Consistent with asymmetric learning, column
1 reveals that one additional quit in one’s work history is associated with a 2.9 percent decrease
in the current wage for workers with at least a college degree. In a competitive labor market,
this reflects the marginal effect of job hopping on the market’s perceived mean of the workers’
productivity distribution. When the market does not observe a worker’s ability directly, it can only
make inferences about the worker’s ability based on the observed job-hopping history. Comparing
two otherwise identical college graduates, the market will interpret a worker who has one more
voluntary job hop as having 2.9 percent lower expected ability than the other worker.
In contrast, for the less educated workers the corresponding effect in column 4 is statistically
insignificant and small in magnitude. These results suggest a job-hopping wage penalty for col-
lege graduates but not for the less educated, even after controlling for current job tenure, labor
market experience, and an ability measure. These contrasting results are consistent with those
from table 2, where the negative correlation between the number of previous quits and the AFQT
score is statistically significant for college graduates but not for the less educated.
Motivation for an exploration of whether the job-hopping wage penalty is particularly large
when the job tenure that preceded the job was extremely short can be found in the 2012 survey
by Bullhorn, cited in footnote 2, which suggests that leaving a job before one year of tenure sends
a particularly negative signal. To investigate whether there are “critical cutoffs” for job tenure
(such as one year) that a worker must clear to avoid a wage penalty in subsequent jobs, a dummy
variable is defined that equals one if job tenure is less than a given threshold value, and zero
otherwise. The log-wage regressions are then estimated, replacing the job-hopping measure with
a new variable, i.e., the number of previous job tenures that fall below the threshold. A sequence
of different threshold values (all in months) is considered, i.e., 3, 6, 9, 12, ..., 36.37
Figure 3 plots the results. The left panel (for college graduates), plots coefficients of the number
of previous tenures shorter than the threshold value indicated on the horizontal axis, along with
the 90% confidence interval. The right panel repeats the plots for the less educated. The left panel
shows that tenures shorter than one year (exclusive) are associated with considerably larger (in
magnitude) wage penalties for college graduates than tenures longer than a year. This result is
consistent with the quote from the 2012 Bullhorn survey in footnote 2. For the less educated the
effects of the numbers of previous tenures shorter than certain thresholds are all near zero and
statistically insignificant at the 10% level for all threshold values ranging from six months to three
years.
To statistically test for the increasing pattern (for college graduates) in the point estimates dis-
played in the figure, a regression is estimated on a stacked data set with 12 blocks corresponding
to the different threshold values (i.e., 3, 6, 9, 12, ..., 36).38 Letting γ j denote the coefficient of W Hit,j ,
37 In the sample, counting each previous job where the worker quits voluntarily, the median tenure length is 16

months for college graduates and 8 months for the less educated. Among college graduates (the less educated), 40%
(61%) of tenures are less than 12 months, and 5% (9%) of tenures are less than 3 months.
38 The reason for stacking the data is that the key independent variable, W H , is defined differently for each thresh-
it,j
old value, j, and the stacked regression allows its 12 coefficients, γ j , to be estimated in a single regression rather than in
12 separate regressions. This facilitates hypothesis testing on the γ j across different j.

18
i.e., the number of previous job tenures that fall below the threshold of j months, for each ed-
ucation group the null hypothesis H0 : γ j−3 ≥ γ j is tested against the one-sided alternative H1 :
γ j−3 < γ j , for j = 6, 9, 12, ..., 36. Rejection of the null supports the increasing pattern of point
estimates displayed in the figure for college graduates. For the less educated, the null hypothesis
cannot be rejected in most cases. For college graduates, however, the null can be rejected at the
10% level for all but three thresholds, especially at the 12-month threshold where the p-value is
near zero.39
The second testable implication is empirically supported given that the main result concern-
ing a job-hopping wage penalty for college graduates (column 1 of Table 3) holds more strongly
for rookies (column 2) than for veterans (column 3). The result from column 4 that there is no
job-hopping wage penalty for the less educated also holds for rookies (column 5) and veterans
(column 6).
Testing the third implication requires augmenting the regression with the interaction between
labor-market economic conditions and work history. Two different measures of economic condi-
tions are used, namely ut0 and ut0 ,t , where the former is a proxy for the unemployment rate in the
year of labor-market entry, t0 , and the latter is the average unemployment rate between the year
of labor-market entry and the present year t. The augmented regression is:

ln wit = Xit0 β + W Hit0 γ + ϕũ + ũ · W Hit0 ξ + eit , ũ ∈ {ut0 , ut0 ,t } (22)

For the first measure, column 1 of Table 4 displays results for rookie college graduates. The
estimate of ϕ is negative (i.e., −0.034), as first documented by Kahn (2010) using the same data but
without controlling for work history. Including the unemployment control leaves the main result
from Table 3 preserved. That is, the marginal effect of the number of previous jobs is negative and
statistically significant, though it is mitigated (given that the estimated ξ is positive and statisti-
cally significant) when macroeconomic conditions at the time of labor-market entry are poor. For a
college graduate who enters the labor market in a good year with a low unemployment rate of 2%,
the detrimental effect of one additional quit is −6.2%.40 When the unemployment rate increases
by one percentage point, to 3%, the detrimental effect of one additional quit shrinks in magnitude
to −1.3%. Such mitigation is concave, as shown in column 2 when the quadratic term of the unem-
ployment rate and its interaction with the number of previous quits is included.41 These results
are consistent with employers having some degree of sympathy for job hopping when workers
have the misfortune of entering the labor market during a bad economy.
Results for veteran college graduates (column 3) reveal a job-hopping wage penalty that is
smaller in magnitude than for rookies and that is also mitigated by poor macroeconomic con-
ditions at the time of labor-market entry, although the results are statistically insignificant. The
39 The three exceptions are j = 6 (where the p-value is 0.37), j = 15 (where the p-value is 0.43), and j = 36 (where the
p-value is 0.18).
40 −0.160 + 2(0.049) = −0.062.
41 This implies that it is unlikely that the total effect of job hopping will become significantly positive. The estimates

in Table 4 imply that the total effect of the number of previous quits is always negative.

19
difference in results for rookie versus veteran college graduates suggests that the effect of eco-
nomic conditions at the point of labor-market entry wears off with experience. For veterans, the
effect may fade out over time because the economy may have significantly improved.42 Another
interpretation is that bad economic conditions may be more damaging (in the sense of poor match
qualities) for new labor-market entrants than for those with a longer record of experience. That
would justify employers assigning lower wage penalties to the former group than to the latter
when matches are formed in an economic downturn.
All of the preceding results are weaker for the less educated than for college graduates, as seen
in columns 5-8. For rookies, the pattern of signs of the key coefficients matches that of column 1
for rookie college graduates, though all of the magnitudes are considerably smaller. Recall also
that the dependent variable is in logs, so the marginal effects in column 5 (versus those in column
1) represent smaller percentage changes starting from a smaller base. For less-educated veterans
(column 7), both of the aforementioned effects are even smaller. Therefore, the job-hopping stigma
for the less educated is, at most, minimal for both rookies and veterans, and is barely significant
for veterans, when controlling for economic conditions in the labor-market entry year.
Next consider the alternative proxy for economic conditions when the job-hopping event oc-
curred, namely the average unemployment rate weighted by the incidence of job quitting years.
The results, presented in Table 5, echo those in Table 4.
The main results from Table 3 are insensitive to using three alternative job-hopping measures.
The first is the length of the shortest prior job tenure. If a worker frequently hops jobs, the shortest
prior tenure is likely to be quite short. Therefore, a positive coefficient on this measure indicates
that current wages are higher when the shortest prior tenure is longer, which is consistent with
a job-hopping wage penalty. Panel A of Table 6 reveals that the coefficient of this alternative
measure is positive and statistically significant for college graduates but not for the less educated,
consistent with the results from Table 3. A one-year increase in the shortest prior job tenure for col-
lege graduates is associated with a 1% increase in the current wage, whereas for the less-educated
the corresponding effect is small and statistically insignificant.43 Although both the number of
previous jobs and the shortest previous tenure proxy job hopping, they do not capture the same
information. When both measures are included in the same model, results in appendix Table A1
show that the former measure has a stronger signaling effect than the latter for college graduates.
Intuitively, the number of previous jobs is a more comprehensive measure of job hopping.
The second measure uses different reasons workers report leaving their previous jobs and
counts each incidence up to the date at which the wage is measured.44 Table 6’s Panel B reports
42 Altonji et al. (2014) find that the effects of economic conditions at the point of labor-market entry fade out after the
first seven years.
43 Excluding workers from the sample who never had turnover leaves the results substantively unchanged. The

results are also insensitive to various specification changes following the literature on asymmetric employer learning.
In particular, including interactions between current tenure and schooling, as well as interactions between current
tenure and AFQT scores, as in Schönberg (2007), leaves the results substantively unchanged. To control for the effect
of unemployment (e.g., Ellwood, 1982), the continuous employment spell is included along with its interactions with
schooling and AFQT scores, as in Pinkston (2009). This leaves the results qualitatively unchanged.
44 They are categorized as either voluntary quits or involuntary separations (including layoffs and firings).

20
the results. The number of involuntary separations has a large scarring effect for college graduates,
especially for rookies for whom one additional involuntary separation lowers current wages by
about 7.1%. Its impact for the less-educated group is statistically insignificant. The effect of the
number of voluntary quits is similar to the results in Table 3. College graduates suffer more from
involuntary separations than from voluntary quits in their job history.
The third measure is the ratio of the number of previous quits to labor market experience,
as in Munasinghe and Sigman (2004). Panel C’s results are similar to those in Table 3, although
the magnitudes differ.45 Thus, even using the alternative measure of work history, job hopping
is associated with a wage penalty for college graduates but not for the less educated, and this
negative effect is stronger for early-career workers among college graduates.
To ensure a complete work history, all subsequent periods are eliminated after the first gap
of three or more years absent from the labor market for each individual. This excludes 8% of the
monthly sample. As an alternative cutoff, for each individual all subsequent periods are elimi-
nated after the first gap of six or more years absent from the labor market. Less than 2% of the
monthly sample is excluded in that case, and the results are essentially unchanged. Alternative
trimming criteria for hourly wages are also considered. When wages below the national minimum
wage ($5.15 in 2004) are excluded, or when top wages are kept, or when all wages are kept, results
are similar, and the negative signaling effects of job hopping are strengthened in the latter two
cases. Although the number of voluntary quits is the appropriate job-hopping measure given that
involuntary separations are not modeled in the theoretical analysis, results are similar when all
job hops are included regardless of the reasons. Regression (21) was also estimated on new jobs
only, that is, the first observation of each job for each individual. Table 7’s results show that job
hopping has a negative and statistically significant effect among college graduates. There is no
significant effect for the less educated. These results are similar to those in Table 3.
The prediction from Table 3 that the signaling role of job hopping is more harmful for early-
career workers is of interest given that it might have been expected, counter to these results, that
employers would be more tolerant of early-career hoppers who are experimenting with early jobs,
looking for a good fit. To explore this, the definition of early career is varied. Specifically, restrict-
ing the sample to periods after each individual is first observed working at least 1000 hours per
year for at least two consecutive years without changing schooling levels ensures that a labor-
market entry point is chosen after which each worker is fully engaged. But this rule might omit
early-career periods (e.g., the transition from school to work) during which workers often sam-
ple jobs. Employers might view job hopping during these periods differently and perhaps with
greater sympathy. To allow for this possibility, the selection criterion was relaxed by defining
labor-market entry as working at least 100 hours per year for at least one year. For this broadened
sample the result for both education groups is qualitatively unchanged.
Table 8 presents results from twelve additional robustness checks. The first includes a control
45 The average impact of the number of previous jobs (not the ratio) is −0.021 for college graduate rookies or veterans,

which is slightly smaller in absolute terms than Table 3’s results.

21
for a non-cognitive skill measure available in the NLSY (a summation of the Rotter Locus of Con-
trol Scale and the Rosenberg Self-Esteem Scale). The second controls for parental education levels
as well as parents’ one-digit occupation code and full-time working status in 1978. The results are
largely unchanged in either case, as shown in columns 1-4 of Panel A.46 The third includes the
initial log wage as a proxy for each individual’s unobservable ability or “fixed effect”. Panel A’s
columns 5 and 6 display the regression results with the initial log wage included and are quali-
tatively unchanged: the coefficient of interest is significantly negative for college graduates and
statistically insignificant for the less educated. The fourth includes dummies for the reasons for
leaving the last job (layoff, firing, program termination, or voluntary separation) and potential un-
employment spell (defined as potential experience net of actual experience) and yields results that
are largely unchanged (columns 1-2 in Panel B). The fifth, which also yields similar results, uses
two-digit, instead of one-digit, occupation and industry codes, as shown in Panel B’s columns 3
and 4.
To investigate whether unemployment spells have a scarring effect on future wages, the sixth
robustness check includes the number of previous Job-Unemployment-Job (JUJ) transitions as a
control and calculates the number of previous quits (the independent variable of interest) as the
number of previous Job-to-Job (JJ) transitions. Columns 5 and 6 of Panel B in Table 8 reveal that the
number of JJ hops still has a negative and significant effect on current wages for college graduates.
For those with at most a high-school degree, JJ hops actually have a positive and significant ef-
fect, although the magnitude is much smaller than that of college graduates.47 Job hops involving
changes in occupation or industry could destroy occupation-specific or industry-specific human
capital (e.g., Neal, 1995; Kambourov and Manovskii, 2009).48 To the extent that such destruction is
of a greater magnitude for the higher educated, it could explain the main result. That is, perhaps
it is not hopping, per se, that has adverse consequences for wages, but rather the loss in spe-
cific human capital that accompanies hops that span different occupations and/or industries. The
seventh robustness check therefore includes occupation-specific experience and industry-specific
experience in the model. Panel C’s columns 1 and 2 are largely insensitive to these inclusions. The
eighth robustness check further distinguishes between job hopping in previous occupations and
job hopping within the current occupation, focusing on the former measure, as it should be inde-
pendent of current wages in the hypothesis of destruction of occupation-specific human capital.
46 Many parents have missing occupation and industry codes if they are not working, so one dummy is added for

missing occupation code and one dummy for missing industry code. The results hardly change when only parental
education levels are included.
47 The coefficient of JUJ transitions is negative and statistically significant for college graduates (-0.044***) or for those

with at most a high-school degree (-0.015***), confirming the scarring effect of unemployment. The coefficient of JUJ
transitions includes both the scarring effect of unemployment and the signaling effect of job history. Separating these
two effects is difficult with the current data set. In addition, the scarring effect of unemployment could be partly
attributed to a signaling effect. Therefore, in this specification the focus is on JJ transitions, and JUJ transitions are
included as a control.
48 Among college graduates, 43% (50%) of job hops involve a switch in occupation (industry). The corresponding

figures are 55% (57%) for those with at most a high-school degree. Some workers change occupations or industries
within the same employers. For a switch in occupation (industry), 49% (66%) of the change comes with a switch in
employers for college graduates, or 62% (76%) for the less educated.

22
Panel C’s columns 3 and 4 show, again, a negative effect of job hopping for college graduates but
much less so for the less educated. The main differential results for these two groups also hold
when the occupation-year and industry-year fixed effects are included (Panel C’s columns 5 and
6), or the region-year fixed effects (Panel D’s columns 1 and 2), or the region-industry-year fixed
effects (Panel D’s columns 3 and 4).
As a worker stays longer in one firm, the firm would have more information about the worker’s
ability and the match, and thus would rely less on signals of ability and productivity. The twel-
veth robustness check adds an interaction term between the number of previous quits and current
tenure to explore the dynamics of the job-hopping penalty within one firm. Panel D’s columns
5 and 6 reveal that the negative effect of job hopping decreases with tenure for both education
groups, although the magnitude for the less-educated group is much smaller. This implies that
the signaling effect of job hopping is strongest when a worker just commences the job and the firm
does not have much information about the worker’s ability. The signaling effect gradually fades
away as the firm obtains more information about the worker’s ability and match quality.49
To summarize, college graduates suffer a job-hopping wage penalty that is larger for rookies
than for veterans and particularly pronounced for job tenures below one year. For the less edu-
cated, there is little or no job-hopping wage penalty. Given the main empirical finding in Table 3,
the symmetric-learning case is more applicable to the less-educated group, and the asymmetric-
learning case is more applicable to college graduates. Furthermore, the job-hopping wage penalty
incurred by rookie college graduates is mitigated when workers enter the labor market during an
economic downturn or when the previous job dissolved during poor economic conditions. These
results justify rejection of the null hypothesis that there is no asymmetric learning about worker’s
ability in the labor market for college graduates. The same null hypothesis cannot be rejected for
the less educated. The empirical evidence is consistent with the variance of match quality being
high (low) relative to the variance of ability for the less educated (college graduates).
The model under asymmetric learning offers an explanation for the persistent negative effect
of graduating from college in a recession. When the unemployment rate is high, job tenure is
generally shorter and is more likely to be the shortest job tenure in one’s work history.50 However,
even if employers recognize this business-cycle effect on job hopping and are sympathetic to it,
they cannot perfectly separate the aggregate shock from other factors specific to the worker. Thus,
if a worker’s employment history includes a bad start that occurred during a recession, it would
have a persistent effect on the worker’s wage through job-hopping stigma, even though match
49 The signaling effect of job hopping becomes insignificant after nine years for college graduates if the effects are all
assumed to be linear.
50 When the length of each terminated tenure is regressed on the unemployment rate of the year when this tenure

was terminated, the coefficient is −0.06 and is statistically significant at the 1% level, controlling for individual fixed
effects, experience and its square, and age and its square. That is, when a job terminates in one year, the length of its
tenure is shorter if economic conditions in that year are worse. A dummy variable is also constructed, with 1 indicating
that one job has the shortest tenure during one worker’s entire work history in the sample, and this dummy variable is
regressed on the unemployment rate, controlling for individual fixed effects. The coefficient is 0.036 and is significant
at the 1% level. This implies that when a job ends in one year, the probability of it being the shortest tenure increases
when labor-market conditions worsen.

23
qualities dissolve upon separation and are independent across employers.
Finally, the job-hopping wage penalty for college graduates does not imply that job hopping
causes wages to literally decrease on the new jobs. Previous research has found that job changes
are beneficial for wage growth, particularly for young workers (e.g., Topel and Ward, 1992). The
NLSY data reveal that, on average, a job change is associated with a 13.4% increase in the wage
level for “rookies” and a 6.8% increase for “veterans” among college graduates, suggesting that
job hopping causes slower wage growth upon job changes. For the less educated, the increase is
7.7% for rookies and 3.5% for veterans, independent of the job-hopping history.

4 Discussion
Section D of the appendix considers the following alternative explanations for the job-hopping
wage penalty for college graduates: (1) on-the-job search, (2) public ability with high-ability work-
ers infrequently changing jobs, (3) symmetric learning about workers’ tastes for staying, and (4)
human capital accumulation. None of these alternatives seems as plausible as the one proposed
here. Additionally, and in the spirit of the proposed model, section E of the appendix summa-
rizes results from an alternative (experimental) data set suggesting that job hopping reduces a job
applicant’s probability of receiving a callback in response to a job application.
The results for college graduates are consistent with prior work that has found empirical sup-
port for asymmetric learning for that group (Schönberg, 2007; Pinkston, 2009; Kahn, 2013; Kim
and Usui, 2014). Schönberg (2007) and Kim and Usui (2014) both find that learning is asymmet-
ric for college graduates but symmetric for workers with high-school degrees only. As discussed
in DeVaro and Waldman (2012), however, most jobs are probably better described by a blend of
asymmetric and symmetric learning than by either perspective by itself. The job-hopping stigma
for college graduates can be explained by asymmetric learning. On the other hand, as in Al-
tonji and Pierret (2001) and Arcidiacono et al. (2010), the present results show that for the less-
educated group the positive marginal effect of the AFQT score increases with experience (Tables
3 and 6), consistent with symmetric learning. Therefore, a reasonable interpretation of the dif-
ferential findings between the education groups is that the asymmetric-learning mechanism is
relatively stronger for college graduates, so that the predictions of the theoretical model (under
pure asymmetric learning) are clear in the data, whereas for the less-educated group asymmetric
learning is a less powerful force and is virtually indetectable in the data.51
To control for symmetric learning, observed ability is included using a proxy measure (AFQT
score) in the NLSY79, as in (e.g., Altonji and Pierret, 2001; Schönberg, 2007; Pinkston, 2009). As
the worker’s full information set is not observed, however, there may be omitted variable bias due
to unobservable worker, firm, or match-specific characteristics (e.g., the NLSY data lack detailed
51 Even for the less educated, however, the results are suggestive of asymmetric learning when the regression includes

economic conditions at the point of labor-market entry, or the average of economic conditions from the point of labor-
market entry up until the data are recorded.

24
information on firm characteristics or worker-firm specific matches).52 On the other hand, such
endogeneity should be less threatening to two of the present specifications. First, the preceding
eighth robustness check (Columns 3 and 4 of Table 8’s Panel C) distinguishes between across-
occupation and within-occupation job hopping. The former is less likely to be correlated with
the most recent job change because it is more distant in time and because it crosses occupations.
Second, the NLSY79 results are echoed in those from the experimental data used in Bertrand and
Mullainathan (2004) (section E of the appendix).53 Nevertheless, the co-existence of symmetric
learning (of AFQT) and asymmetric learning (signaled by the job-hopping history) demonstrates
the presence of other unobservable quality measures that are negatively correlated with, and thus
signaled by, job hopping because of selection.54
The conclusion that asymmetric learning is a particularly powerful force for college graduates
is consistent with prior literature, but why does employer learning differs in this way between
the two worker groups? One interpretation is that the variance of unobserved ability among less-
educated workers is smaller than the variance among college graduates, relative to the match-
quality distribution. Consequently, the signaling effect or job-hopping wage penalty is smaller for
the former group, as simulated in subsection 2.3. For example, the less-educated workers are more
likely to stay in their local communities, and the dense network of social connections developed
over a long tenure in the same location facilitates learning more precise public information about
the worker’s ability, approximating a symmetric-learning environment. In contrast, college grad-
uates are relatively more likely to enter the labor market in a geographic location different from
where they grew up (and move around more throughout their careers), and the lack of a dense
local social network might make the asymmetric-learning perspective more relevant. Another
possibility is that the difference between the educational groups is driven by the type of work
they perform. High-school dropouts might perform very simple tasks, such as those performed in
fast food restaurants. Such a worker’s tenure in a fast food restaurant might provide all relevant
information to a competing employer, i.e., the worker exceeds some minimum proficiency level,
and ability does not affect productivity much, which implies a smaller variance in the relevant
ability distribution among this group. In contrast, higher-educated workers tend to engage in
more complex tasks or there exists a greater correlation between ability and match quality. In such
cases, there is more adverse selection in the labor market, as illustrated in subsection 2.3.
Among college graduates, the results indicate that job hopping is more harmful for rookies
than for veterans. There are two possible reasons. First, the variance of the unknown component
52 See Buchinsky et al. (2010) for an overview of these issues. Detailed worker-firm matched data may offer op-
portunities to account for such unobserved heterogeneity, though such data sets typically lack ability measures (both
cognitive and non-cognitive) that are present in the NLSY.
53 In the experiment the number of prior jobs listed on an applicant’s résumé is an exogenous random variable that

has a negative effect on callbacks for college graduates but not high-school graduates, which evokes the present study’s
results on wages (see section E of the appendix).
54 Table A2 of the appendix’s section A shows that omitting work history (asymmetric learning of this unobserved

quality) from the right-hand side leads to underestimates of the effects of experience and overestimates of the effects of
current tenure, especially for rookie college graduates. This is because the number of job hops is positively correlated
with experience and negatively correlated with current tenure, and such correlations are stronger among rookies.

25
of ability decreases over the worker’s career as the labor market gains more precise information
about the worker’s ability. As the variance decreases, the job-hopping wage penalty decreases.
Second, columns 5 and 6 in Table 8’s Panel D show that the negative effect of job hopping decreases
with tenure, indicating that the current firm gradually learns more about the worker’s ability and
thus relies less on the signaling effect of job hopping. For college graduates, the average tenure is
1.5 years among rookies and 5.1 years among veterans. As college graduates change jobs sooner
and more frequently during their early career, before the current firms gain enough information
about their abilities, it implies a higher wage penalty for early-career job hoppers.
The analysis suggests that match quality, despite being independent across employers, has
persistent effects on careers in an asymmetric-learning setting, even when a worker switches firms.
Bad luck with a particular firm (via a poor match quality) sticks with a worker even after leaving
that firm, because prospective employers cannot distinguish bad luck from low ability. In contrast,
in the symmetric-learning environment, all employers have the same information about ability at
all times, so separations are seen as arising only from bad luck and not from low ability, and
previous match quality has no effect on current or future wages.
These arguments have welfare implications. If learning is symmetric, poor matches are tem-
porary because workers have no incentive to preserve them. A poorly matched worker seeks a
new job, and the match is replaced by a new draw from the distribution, which is independent of
the worker’s ability. But when learning is asymmetric, workers who switch firms are adversely
selected, and turnover is inefficient. When period 1 ends, workers quit if and only if y1 < ŷ1 .
High-ability workers with poor match qualities choose to stay, so as to differentiate themselves
from low-productivity lemons; and low-ability workers with high match qualities choose to quit,
as long as competing firms pay more. Both scenarios result in inefficiency and welfare loss.55
A further result is that the welfare loss increases with the number of periods, based on the
following logic. From an efficiency standpoint, workers with poor matches in period 1 should
separate. But doing so sends a negative signal to the market, which has negative consequences
for future earnings. The longer the future horizon, the greater the magnitude of those losses in
future earnings, so the more tempted a poorly-matched worker will be to stay with the original
employer in an effort to signal high ability to the market.56 An implication is that the inefficiency
should be greatest for early-career workers. A late-career worker who falls into a poor match has
less to lose from ending it than would a new entrant facing an entire future career.
55 Not all asymmetric employer learning implies welfare loss. In Greenwald (1986), where the secondary market
is sustained by an exogenous separation of worker-firm matches, turnover does not have any welfare consequences.
Asymmetric employer learning in this context only shifts the utility allocation between workers and firms. The model
in Pinkston (2009) has similar implications. By assuming that turnover creates a loss of productivity or some transaction
costs, the model in Gibbons and Katz (1991) implies that turnover always causes welfare loss. The idiosyncratic utility
shock in Schönberg (2007) also implies some welfare loss caused by asymmetric employer learning.
56 Thus, the inefficiency in the period-1 turnover rate under asymmetric learning (i.e., the extent to which it falls short

of the efficient period-1 turnover rate in the benchmark case of symmetric learning) is larger than the corresponding
inefficiency in the period-2 turnover rate, and so on.

26
5 Conclusion
The job-hopping wage penalty for college graduates (but not the less educated) is statistically
and economically significant, larger for new labor-market entrants than for veterans, particularly
large when jobs are held for less than one year, and mitigated when workers enter the labor market
during an economic downturn. The result can be interpreted as a test for the presence of asym-
metric learning. The differential results by education level echo those of Schönberg (2007) and
Kim and Usui (2014) based on different arguments.
The analysis introduces the idea that work history is a signal of ability, which workers consider
when making decisions to stay with or leave an employer. Because employers cannot discern
whether a separation arose from a bad match or from low ability, they assign some weight to the
latter possibility, which negatively affects job hoppers’ wages. Thus, even though match quality
is independent across firms, workers carry the ill effects of a bad match with them even after
separating. This causes poorly matched workers to think twice before ending a match. It also
means that early-career workers have the most to lose by prematurely ending a bad match, since
they have an entire career ahead of them in which to suffer wage penalties imposed by the market.
The main insights have application beyond the labor market. For example, consider two re-
cently divorced men re-entering the marriage market. Both have been married for nine years, but
one was married to the same spouse for all nine years, whereas the other was married to three
different spouses, each for three years. For both men, the prior divorces may have been caused
by bad luck (i.e., poor match quality) or because they were bad husbands. But since prospective
spouses cannot distinguish between those cases, the man with three divorces in his past will be
viewed more skeptically than the once-divorced man, in the eyes of prospective spouses. Since
the thrice-divorced man has lower market value than the once-divorced man, the expected qual-
ity of his subsequent spouse is reduced. A further implication is that even if it is apparent to
both parties to a new marriage that the match is horrible, it may be wise to hold off on a divorce
“for a respectable amount of time”, to mitigate the negative signal resulting from divorce, thereby
improving the prospects of a better marriage in the future.57
There are some unaddressed elements that, if incorporated into the theoretical model in future
research, could generate further insights. One is workers’ search behavior. Another is school-
ing choices, or observable worker characteristics (other than work history) that relate to ability.
In such an extension, workers could signal ability to prospective employers both through their
schooling choices and through their career histories. Another extension would incorporate invol-
untary turnover, which would allow for unemployment spells as in Pinkston (2009). However,
the empirical results hold in specifications that include continuous employment spells. The anal-
ysis also does not allow for promotions, as studied in Waldman (1984), Bernhardt (1995), Owan
57 The
model also predicts that, under asymmetric learning, employers should be more forgiving when there is good
reason to believe that job hoppers faced bad matches in their employment histories. Similarly, in the marriage market,
prospective spouses should be more forgiving of divorcees who were previously in arranged (rather than voluntary)
marriages.

27
(2004), DeVaro and Waldman (2012), Prasad and Tran (2013), Cassidy et al. (2016), and DeVaro et
al. (2018), which empirically are an important source of wage growth and internal (i.e., within-
firm) job mobility. Empirically testing that extension would be infeasible in the NLSY79 data set,
which lacks information on internal promotions. It would, however, be feasible in a large-scale,
linked, employer-employee Finnish data set (Kauhanen and Napari, 2012).

28
References
Acemoglu, Daron and Jorn-Steffen Pischke, “Why Do Firms Train? Theory and Evidence,” Quar-
terly Journal of Economics, Feburary 1998, 113 (1), 79–119.

Akerlof, George A., “The Market for “Lemons”: Quality Uncertainty and the Market Mecha-
nism,” Quarterly Journal of Economics, August 1970, 84 (3), 488–500.

Altonji, Joseph G. and Charles R. Pierret, “Employer Learning and Statistical Discrimination,”
Quarterly Journal of Economics, February 2001, 116 (1), 313–350.

, Lisa B. Kahn, and Jamin D. Speer, “Cashier or Consultant? Entry Labor Market Conditions,
Field of Study, and Career Success,” Working Paper 20531, National Bureau of Economic Re-
search September 2014.

Antonovics, Kate and Limor Golan, “Experimentation and job choice,” Journal of Labor Economics,
2012, 30 (2), 333–366.

Arcidiacono, Peter, Patrick Bayer, and Aurel Hizmo, “Beyond Signaling and Human Capital:
Education and the Revelation of Abilit,” American Economic Journal: Applied Economics, 2010, 2
(4), 76–104.

Bagnoli, Mark and Ted Bergstrom, “Log-concave probability and its applications,” Economic The-
ory, 2005, 26, 445–469.

Beaudry, Paul and John DiNardo, “The Effect of Implicit Contracts on the Movement of Wages
Over the Business Cycle: Evidence from Micro Data,” Journal of Political Economy, August 1991,
99 (4), 665–688.

Ben-Porath, Yoram, “The Production of Human Capital and the Life Cycle of Earnings,” Journal
of Political Economy, August 1967, 75 (4), 352–365.

Bernhardt, Dan, “Strategic Promotion and Compensation,” Review of Economic Studies, April 1995,
62 (2), 315–339.

Bertrand, Marianne and Sendhil Mullainathan, “Are Emily and Greg More Employable Than
Lakisha and Jamal? A Field Experiment on Labor Market Discrimination,” American Economic
Review, 2004, 94 (4), 991–1013.

Bills, David B, “Employers’ Use of Job History Data for Making Hiring Decisions: A Fuller Spec-
ification of Job Assignment and Status Attainment,” Sociological Quarterly, Spring 1990, 31 (1),
23–35.

, “Labor market information and selection in a local restaurant industry: The tenuous balance
between rewards, commitments, and costs,” Sociological Forum, December 1999, 14 (4), 583–607.

29
Bognanno, Michael and Eduardo Melero, “Promotion Signals, Experience and Education,” Jour-
nal of Economics and Management Strategy, 2016, 25 (1), 111–132.

Buchinsky, Moshe, Denis Fougere, Francis Kramarz, and Rusty Tchernis, “Interfirm mobility,
wages and the returns to seniority and experience in the United States,” Review of Economic
Studies, 2010, 77 (3), 972–1001.

Burdett, Kenneth, “A Theory of Employee Job Search and Quit Rates,” American Economic Review,
1978, 68 (1), pp. 212–220.

and Dale T. Mortensen, “Wage Differentials, Employer Size, and Unemployment,” International
Economic Review, May 1998, 39 (2), 257–273.

Cassidy, Hugh, Jed DeVaro, and Antti Kauhanen, “Promotion Signaling, Gender, and Turnover:
New Theory and Evidence,” Journal of Economic Behavior and Organization, 2016, 126, Part A,
140–166.

DeVaro, Jed and Hodaka Morita, “Internal promotion and external recruitment: a theoretical and
empirical analysis,” Journal of Labor Economics, 2013, 31 (2), 227–269.

and Michael Waldman, “The Signaling Role of Promotions: Further Theory and Empirical
Evidence,” Journal of Labor Economics, January 2012, 30 (1), 91–147.

, Suman Ghosh, and Cindy Zoghi, “Job characteristics and labor market discrimination in pro-
motions,” Industrial Relations, 2018, 57 (3), 389–434.

Ellwood, David T, “Teenage unemployment: Permanent scars or temporary blemishes?,” in R.B.

Freeman and D.A. Wise, eds., The youth labor market problem: Its nature, causes, and consequences,
Chicago: University of Chicago Press, 1982, pp. 349–390.

Farber, Henry S, “Employment Insecurity: the Decline in Worker-firm Attachment in the United
States,” Working Paper 1068, Princeton University 2008.

Farber, Henry S. and Robert Gibbons, “Learning and Wage Dynamics,” Quarterly Journal of Eco-
nomics, November 1996, 111 (4), 1007–1047.

Gibbons, Robert and Lawrence F. Katz, “Layoffs and Lemons,” Journal of Labor Economics, Octo-
ber 1991, 9 (4), 351–380.

and Michael Waldman, “A Theory of Wage and Promotion Dynamics inside Firms,” Quarterly
Journal of Economics, 1999, 114 (4), pp. 1321–1358.

and , “Enriching a Theory of Wage and Promotion Dynamics inside Firms,” Journal of Labor
Economics, 2006, 24 (1), 59–108.

Golan, Limor, “Wage Signaling: A Dynamic Model of Intrafirm Bargaining and Asymmetric
Learning,” International Economic Review, August 2009, 50 (3), 831–854.

30
Gorry, Aspen, “Experience and worker flows,” Quantitative Economics, 2016, 7 (1), 225–255.

Greenwald, Bruce C., “Adverse Selection in the Labour Market,” Review of Economic Studies, July
1986, 53 (3), 325–347.

Harris, Milton and Bengt Holmstrom, “A Theory of Wage Dynamics,” Review of Economic Studies,
July 1982, 49 (3), 315–333.

Heckman, James J. and Bo E. Honore, “The Empirical Content of the Roy Model,” Econometrica,
September 1990, 58 (5), 1121–1149.

Hu, Luojia and Christopher Taber, “Displacement, Asymmetric Information, and Heterogeneous
Human Capital,” Journal of Labor Economics, January 2011, 29 (1), 113–152.

Imai, Susumu and Michael P. Keane, “Intertemporal Labor Supply and Human Capital Accum-
lation,” International Economic Review, May 2004, 45 (2), 601–641.

Kahn, Lisa B., “The long-term labor market consequences of graduating from college in a bad
economy,” Labour Economics, 2010, 17, 303–316.

, “Asymmetric Information between Employers,” American Economic Journal: Applied Economics,

October 2013, 5 (4), 165–205.

Kambourov, Gueorgui and Iourii Manovskii, “Occupational specificity of human capital,” Inter-
national Economic Review, 2009, 50 (1), 63–115.

Kauhanen, Antti and Sami Napari, “Career and wage dynamics: Evidence from linked employer-
employee data,” Research in Labor Economics, 2012, 36, 35–76.

Kim, Seik and Emiko Usui, “Employer Learning, Job Changes, and Wage Dynamics,” Working
Paper UWEC-2012-01, University of Washington 2014.

Lange, Fabian, “The speed of employer learning,” Journal of Labor Economics, 2007, 25 (1), 1–35.

Lazear, Edward P., “Entrepreneurship,” Journal of Labor Economics, October 2005, 23 (4), 649–680.

Milgrom, Paul and Sharon Oster, “Job discrimination, market forces, and the invisibility hypoth-
esis,” Quarterly Journal of Economics, 1987, pp. 453–476.

Munasinghe, Lalith and Karl Sigman, “A hobo syndrome? Mobility, wages, and job turnover,”
Labour Economics, 2004, 11 (2), 191–218.

Neal, Derek A, “Industry-specific human capital: Evidence from displaced workers,” Journal of
labor Economics, 1995, 13 (4), 653–677.

, “The Complexity of Job Mobility among Young Men,” Journal of Labor Economics, April 1999,
17 (2), 237–261.

31
 and William R Johnson, “The Role of Premarket Factors in Black-White Wage Differences,”
Journal of Political Economy, 1996, 104 (5), 869–895.

Oreopoulos, Philip, Till von Wachter, and Andrew Heisz, “The Short- and Long-Term Career
Effects of Graduating in a Recession,” American Economic Journal: Applied Economics, January
2012, 4 (1), 1–29.

Owan, Hideo, “Promotion, turnover, earnings, and firm-sponsored training,” Journal of Labor Eco-
nomics, 2004, 22 (4), 955–978.

Papageorgiou, Theodore, “Learning your comparative advantages,” Review of Economic Studies,

2013, p. rdt048.

Pavan, Ronni, “Career choice and wage growth,” Journal of Labor Economics, 2011, 29 (3), 549–587.

Pinkston, Joshua C., “A Model of Asymmetric Employer Learning with Testable Implications,”
Review of Economic Studies, January 2009, 76 (1), 367–394.

Postel-Vinay, Fabien and Jean-Marc Robin, “Equilibrium wage dispersion with worker and em-
ployer heterogeneity,” Econometrica, November 2002, 70 (6), 2295–2350.

Prasad, Suraj and Hien Tran, “Work practices, incentives for skills, and training,” Labour Eco-
nomics, 2013, 23, 66–76.

Schönberg, Uta, “Testing for Asymmetric Employer Learning,” Journal of Labor Economics, 2007,
25 (4), 651–691.

Topel, Robert H. and Michael P. Ward, “Job Mobility and the Careers of Young Men,” Quarterly
Journal of Economics, May 1992, 107 (2), 439–479.

Waldman, Michael, “Job Assignments, Signalling, and Efficiency,” RAND Journal of Economics,
Summer 1984, 15 (2), 255–267.

, “Up-or-Out Contracts: A Signaling Perspective,” Journal of Labor Economics, April 1990, 8 (2),
230–250.

, “Asymmetric Learning and the Wage/productivity Relationship,” Journal of Economic Behavior

& Organization, December 1996, 31 (3), 419–429.

, “Classic promotion tournaments versus market-based tournaments,” International Journal of

Industrial Organization, 2013, 31 (3), 198–210.

Zhang, Ye, “Employer Learning Under Asymmetric Information: The Role of Job Mobility,” Work-
ing Paper, Indiana University-Purdue University Indianapolis 2011.

32
 Table 1: Summary Statistics of NLSY79 Sample (1979-2008)

College High School

Graduates Graduates or less
Mean S.D. Mean S.D.
No. individuals 665 3523
No. observations 13151 67191
Real hourly wage (2004$) 25.561 13.949 14.009 7.427
Number of all jobsa 5.845 3.701 7.361 5.568
Number of previous jobsa 4.845 3.701 6.361 5.568
Number of previous quitsa,c 1.074 1.495 1.775 2.278
Number of previous involuntary separationsa,d 0.444 0.852 1.166 1.782
Experienceb 9.754 6.668 9.028 7.094
Current tenureb 4.029 4.486 3.056 4.199
AFQTa 1.185 0.737 -0.321 0.846
Whitea 0.770 0.421 0.569 0.495
Hispanica 0.077 0.266 0.174 0.379
Blacka 0.153 0.361 0.257 0.437
Years of schoolinga 17.174 1.460 11.225 1.341
Age 33.522 7.278 29.754 8.193
Urban area 0.843 0.363 0.759 0.428
USA residency at age 14 0.989 0.102 0.976 0.152
Unemployment rate at entrya 2.465 0.466 9.160 3.027
Average unemployment rate 2.376 0.238 9.276 2.654
"High School Graduates or less" are those with 12 or fewer years of schooling; "College
Graduates" are those with 16 or more years of schooling.
a Variables reported at the individual level.
b Current tenure and experience are divided by 52 (i.e., measured in years).
c Quits include separations due to reasons other than layoff or firing.
d Involuntary separations refers to the number of previous involuntary job separations,

including layoffs and firings.

33
 Table 2: AFQT and Job Hopping

1 2 3 4 5 6
College graduates High school graduates or less
All Quits Involuntary All Quits Involuntary
separations separations
AFQT -0.750*** -0.260*** -0.128** -0.043 -0.038 -0.049
(0.229) (0.100) (0.056) (0.107) (0.048) (0.038)
Experience 0.427*** 0.118*** 0.042** 0.986*** 0.305*** 0.208***
(0.074) (0.033) (0.018) (0.035) (0.016) (0.013)
Experience2 -0.009*** -0.004*** -0.001* -0.025*** -0.009*** -0.006***
(0.003) (0.001) (0.001) (0.001) (0.001) (0.0004)
R-squared 0.16 0.04 0.04 0.29 0.11 0.09
Observations 665 665 665 3,523 3,523 3,523
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust Huber/White standard errors are in parenthe-
ses. The sample includes each individua’s last observation.
"All" refers to the number of previous jobs regardless of reaons left; "Quits" refers to the
number of previous jobs one left not due to layoffs or firings; "Involuntary separations"
refers to the number of previous involuntary job separations, including layoffs and firings.
All regressions include years of schooling, Hispanic, Black, urban residency, USA resi-
dency at age 14, a dummy variable for the interview year, and (for high school graduates
or less) a dummy for finishing high school.

Table 3: Current Log-wages and Job Hopping

1 2 3 4 5 6
College graduates High school graduates or less
All Rookies Veterans All Rookies Veterans
No. of previous quits -0.029*** -0.041*** -0.030*** -0.002 0.003 -0.003
(0.011) (0.015) (0.011) (0.003) (0.005) (0.003)
AFQT score 0.082*** 0.063** 0.095*** 0.026*** 0.016** 0.048***
(0.020) (0.024) (0.029) (0.007) (0.008) (0.013)
AFQT×Experience 0.002 0.010 0.001 0.005*** 0.008*** 0.003***
(0.002) (0.007) (0.002) (0.001) (0.003) (0.001)
R-squared 0.37 0.31 0.29 0.33 0.25 0.30
Observations 13,151 3,942 9,209 67,191 22,352 44,839
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust Huber/White standard errors are in parenthe-
ses, clustered at the individual level.
"Rookies" refer to the first five years in the labor market and "Veterans" refer to periods
after the first five years. All regressions include experience and its square, current tenure
and its square, years of schooling, the interactions of years of schooling with experience,
Hispanic, Black, urban residency, USA residency at age 14, a dummy variable for the in-
terview year, dummy variables for years, the interactions of year dummies with years of
schooling, dummy variables for one-digit industry and occupation codes, (for the whole
sample) a dummy for being college graduate, and (for the whole sample or the high school
graduates or less) a dummy for finishing high school.

34
 Table 4: Current Log-wages and Job Hopping (the Role of Economic Conditions at Labor-market Entry)

1 2 3 4 5 6 7 8
College graduates High school graduates or less
Rookies Veterans Rookies Veterans
Number of previous quits -0.160** -0.487 -0.071 -0.429 -0.039*** -0.020 -0.015* -0.030
(0.062) (0.346) (0.056) (0.305) (0.014) (0.042) (0.008) (0.025)
Number of previous quits 0.049** 0.309 0.0170 0.291 0.005*** 0.0004 0.001 0.004
×Unemp rate at LM entry (0.024) (0.264) (0.021) (0.229) (0.001) (0.008) (0.001) (0.005)
Number of previous quits -0.049 -0.050 0.0004 -0.0004
×Unemp rate at LM entry2 (0.048) (0.041) (0.0004) (0.0004)
AFQT score 0.057** 0.057** 0.097*** 0.098*** 0.021** 0.022*** 0.052*** 0.052***
(0.024) (0.024) (0.029) (0.029) (0.008) (0.008) (0.013) (0.013)
AFQT×Experience 0.012* 0.012* 0.001 0.001 0.007** 0.007** 0.003*** 0.003***

35
(0.007) (0.007) (0.002) (0.002) (0.003) (0.003) (0.001) (0.001)
Unemp rate at LM entry -0.034 0.108 -0.070* -0.138 -0.021*** -0.064*** -0.006** -0.011
(0.027) (0.285) (0.039) (0.386) (0.002) (0.010) (0.003) (0.015)
Unemp rate at LM entry2 -0.026 0.012 0.002*** 0.0002
(0.052) (0.071) (0.0002) (0.001)
R-squared 0.30 0.30 0.28 0.28 0.24 0.24 0.29 0.29
Observations 3,942 3,942 9,209 9,209 22,352 22,352 44,839 44,839
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust standard errors in parentheses are Huber/White, clustered at the in-
dividual level.
"Rookies" refer to the first five years in the labor market, and "Veterans" refer to periods after the first five years.
All regressions include experience and its square, current tenure and its square, years of schooling, the interac-
tions of years of schooling with experience, Hispanic, Black, urban residency, USA residency at age 14, a dummy
variable for the interview year, dummy variables for one-digit industry and occupation codes, and (for high
school graduates or less) a dummy for finishing high school.
 Table 5: Current Log-wages and Job Hopping (the Role of Average Economic Conditions)

1 2 3 4 5 6 7 8
College graduates High school graduates or less
Rookies Veterans Rookies Veterans
Number of previous quits -0.232** -0.472 -0.106 -0.193 -0.029* -0.032 -0.012 -0.029
(0.087) (0.453) (0.208) (3.006) (0.016) (0.053) (0.010) (0.055)
Number of previous quits 0.079** 0.279 0.033 0.104 0.003** 0.003 0.001 0.004
×Average unemp rate (0.035) (0.357) (0.088) (2.483) (0.002) (0.011) (0.001) (0.012)
Number of previous quits -0.041 -0.014 -0.0002 -0.0002
×Average unemp rate2 (0.069) (0.512) (0.001) (0.001)
AFQT score 0.057** 0.057** 0.097*** 0.097*** 0.021*** 0.022*** 0.050*** 0.049***
(0.024) (0.024) (0.029) (0.029) (0.008) (0.008) (0.013) (0.013)
AFQT×Experience 0.012* 0.012* 0.001 0.001 0.007** 0.007** 0.003*** 0.003***

36
(0.007) (0.007) (0.002) (0.002) (0.003) (0.003) (0.001) (0.001)
Average unemp rate -0.075** 0.118 -0.024 -1.848 -0.026*** -0.096*** 0.014** 0.038
(0.029) (0.273) (0.106) (2.306) (0.002) (0.011) (0.006) (0.036)
Average unemp rate2 -0.037 0.381 0.003*** -0.001
(0.052) (0.478) (0.001) (0.002)
R-squared 0.30 0.30 0.28 0.28 0.24 0.24 0.29 0.29
Observations 3,942 3,942 9,209 9,209 22,352 22,352 44,839 44,839
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust standard errors in parentheses are Huber/White, clustered at the
individual level.
"Rookies" refer to the first five years in the labor market, and "Veterans" refer to periods after the first five years.
All regressions include experience and its square, current tenure and its square, years of schooling, the inter-
actions of years of schooling with experience, Hispanic, Black, urban residency, USA residency at age 14, a
dummy variable for the interview year, dummy variables for one-digit industry and occupation codes, and (for
high school graduates or less) a dummy for finishing high school.
 Table 6: Current Log-wages and Alternative Measures of Job Hopping

1 2 3 4 5 6
College graduates High school graduates or less
All Rookies Veterans All Rookies Veterans
Panel A: Shortest previous tenure
Shortest previous tenure 0.010** 0.045*** 0.009** 0.003 0.003 0.003
(0.004) (0.015) (0.004) (0.003) (0.006) (0.003)
AFQT score 0.083*** 0.065*** 0.097*** 0.026*** 0.016** 0.048***
(0.020) (0.024) (0.029) (0.007) (0.008) (0.013)
AFQT×Experience 0.002 0.009 0.001 0.005*** 0.008*** 0.003***
(0.002) (0.007) (0.002) (0.001) (0.003) (0.001)
R-squared 0.37 0.31 0.28 0.33 0.25 0.30
Observations 13,151 3,942 9,209 67,191 22,352 44,839
Panel B: Reasons left previous jobs
No. of previous quits -0.023** -0.043*** -0.023** -0.002 0.003 -0.003
(0.010) (0.015) (0.011) (0.003) (0.005) (0.003)
No. of previous -0.049** -0.071** -0.047** -0.001 -0.010 0.0003
involuntary separations (0.020) (0.031) (0.020) (0.004) (0.007) (0.004)
AFQT score 0.080*** 0.061** 0.091*** 0.027*** 0.016* 0.048***
(0.020) (0.024) (0.029) (0.007) (0.008) (0.013)
AFQT×Experience 0.002 0.010 0.001 0.005*** 0.008*** 0.003***
(0.002) (0.007) (0.002) (0.001) (0.003) (0.001)
R-squared 0.38 0.31 0.29 0.33 0.25 0.30
Observations 13,151 3,942 9,209 67,191 22,352 44,839
Panel C: Ratio of the number of previous quits to labor market experience
No. of previous quits/exp -0.085*** -0.043** -0.238*** -0.006 0.010 -0.026
(0.031) (0.022) (0.091) (0.008) (0.007) (0.022)
AFQT score 0.081*** 0.062** 0.091*** 0.026*** 0.017** 0.049***
(0.020) (0.024) (0.029) (0.007) (0.008) (0.013)
AFQT×Experience 0.002 0.011 0.001 0.005*** 0.009*** 0.003***
(0.002) (0.007) (0.002) (0.001) (0.003) (0.001)
R-squared 0.37 0.30 0.29 0.33 0.25 0.30
Observations 13,151 3,942 9,209 67,191 22,352 44,839
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust Huber/White standard errors are in parentheses,
clustered at the individual level.
All regressions include experience and its square, current tenure and its square, years of school-
ing, the interactions of years of schooling with experience, Hispanic, Black, urban residency,
USA residency at age 14, a dummy variable for the interview year, dummy variables for years,
the interactions of year dummies with years of schooling, dummy variables for one-digit indus-
try and occupation codes, and (for high school graduates or less) a dummy for finishing high
school.

37
 Table 7: Current Log-wages and Job Hopping (New Jobs)

1 2 3 4 5 6
College graduates High school graduates or less
All Rookies Veterans All Rookies Veterans
No. of previous quits -0.032*** -0.039* -0.036*** -0.003 0.003 -0.004
(0.011) (0.017) (0.011) (0.003) (0.005) (0.003)
AFQT score 0.051** 0.059** 0.040 0.024*** 0.030*** 0.054***
(0.023) (0.028) (0.046) (0.007) (0.009) (0.017)
AFQT×Experience 0.004* 0.007 0.004 0.004*** -0.002 0.002
(0.002) (0.013) (0.003) (0.001) (0.005) (0.001)
R-squared 0.37 0.31 0.32 0.26 0.22 0.25
Observations 3,516 1,765 1,751 24,586 12,151 12,435
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust Huber/White standard errors are in parenthe-
ses, clustered at the individual level.
All regressions include experience and its square, current tenure and its square, years of
schooling, the interactions of years of schooling with experience, Hispanic, Black, urban
residency, USA residency at age 14, a dummy variable for the interview year, dummy vari-
ables for years, the interactions of year dummies with years of schooling, dummy vari-
ables for one-digit industry and occupation codes, and (for high school graduates or less)
a dummy for finishing high school.

38
 Table 8: Current Log-wages and Job Hopping, Robustness Checks

1 2 3 4 5 6
College HS or less College HS or less College HS or less
Panel A:
No. of previous quits -0.027*** -0.003 -0.033*** -0.002 -0.022* -0.001
(0.010) (0.003) (0.010) (0.003) (0.012) (0.003)
Additional controls (Non-cognitive) (Parental characteristics) (Initial log-wage)
R-squared 0.38 0.33 0.39 0.33 0.39 0.34
Observations 12,998 64,243 12,483 53,898 10,510 58,214
Panel B:
No. of previous quits -0.043*** -0.001 -0.030*** -0.003
(0.013) (0.003) (0.010) (0.003)
No. of previous JJ -0.015** 0.004**
transitions (Reasons leaving & (0.006) (0.002)
Additional controls potential unemp spell) (Two-digit Occup/Indus) (Previous JUJ transitions)
R-squared 0.40 0.32 0.46 0.34 0.38 0.33
Observations 3,233 24,223 13,151 67,191 13,151 67,191
Panel C:
No. of previous quits -0.029*** -0.001 -0.038*** -0.005* -0.029*** -0.003
(0.011) (0.003) (0.011) (0.003) (0.010) (0.003)
Additional controls (Occup/Indus specific exp) (Occupation switches) (Occup/Indus-Year FE)
R-squared 0.37 0.33 0.38 0.33 0.40 0.34
Observations 13,151 67,191 13,151 67,191 13,151 67,191
Panel D:
No. of previous quits -0.043*** -0.001 -0.036*** -0.001 -0.041*** -0.005*
(0.012) (0.003) (0.010) (0.003) (0.013) (0.003)
No. of previous quits 0.004** 0.001
×tenure (0.002) (0.001)
Additional controls Region-Year FE Region-Industry-Year FE N/A
R-squared 0.30 0.27 0.41 0.35 0.37 0.33
Observations 13,149 67,165 13,149 67,165 13,151 67,191
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust Huber/White standard errors are in parentheses, clustered at the
individual level.
"HS or less" denotes "High School graduates or less". All regressions include experience and its square, AFQT
score, the interaction of AFQT score with experience, current tenure and its square, years of schooling, the in-
teractions of years of schooling with experience, Hispanic, Black, urban residency, USA residency at age 14,
a dummy variable for the interview year, dummy variables for years, the interactions of year dummies with
years of schooling, (except Columns 3-4 in Panel B) dummy variables for one-digit industry and occupation
codes, (for high school graduates or less) a dummy for finishing high school. In Panel A, "Parental character-
istics" in Columns 3-4 include mother’s and father’s highest education levels, their full-time working status
and dummies for one-digit occupation codes in 1978. In Panel B, "Reasons leaving & potential unemp spell"
in Columns 1-2 include dummies for reasons leaving the previous job and potential unemployment spell (de-
fined as potential experience net of actual experience). In Panel C, "Occupation switches" in Columns 3-4 in-
clude the number of previous quits within the current occupation, and the "No. of previous quits" refers to the
job hopping in previous occupations (exclusive of the current occupation). Two observations in college grad-
uates and thirty-two observations in the high school graduates or less have missing variables in region.

39
 Figure 1: The Asymmetric-Learning Model

, s)
W3 ( y 1
stay
) [ŷ2 (s)] quit
(y 1
W 2 , stay
W o(s
≥ ŷ
1
3 )
y1
[ŷ1 ]
y1
< ŷ , q)
1, q
uit W3 ( y 2
Wo stay
2
[ŷ2 (q)] quit
W o(q
3 )
period 1 period 2 period 3 retire

In the symmetric equilibrium, at the end of period 1 the incumbent wage offer for a worker with
productivity y1 is W2 (y1 ), and the market wage offer is W2o . Workers with productivity less than
ŷ1 quit (and are called “quitters”) and otherwise stay (and are called “stayers”). At the end of
period 2, for a stayer with productivity y1 the incumbent wage offer is W3 (y1 , s), and the market
wage offer is W3o (s); for a quitter with productivity y2 the incumbent wage offer is W3 (y2 , q), and
the market wage offer is W3o (q). Among stayers, those with productivity less than ŷ2 (s) quit, and
otherwise stay. Among quitters, those with productivity less than ŷ2 (q) quit, and otherwise stay.
All workers retire at the end of period 3.

40
 Figure 2: Simulation: θ and ηi follow a joint normal distribution

Threshold values ∆W3AL = W3o (q) − W3o (s)

-1
0
ŷ1
ŷ2(s)
ŷ2(q)
-.5

∗
y

-1.2
W3 (q)-W3 (s)
-1

-1.4
-1.5

-1.6
-2

-1.8
-2.5

-.5 0 .5 1 -.5 0 .5 1
ρ (σ=1) ρ (σ=1)

-.5
0
-.5
-1

W3 (q)-W3 (s)
o

-1
-1.5

o
-2

ŷ1
ŷ2(s)
ŷ2(q)
-1.5
-2.5

∗
y

.6 .8 1 1.2 .6 .8 1 1.2
σ (ρ=0) σ (ρ=0)

Figure 3: Current Log-wages and Job Hopping Prior to Various Tenure Cutoffs

College Graduates High School Graduates or less

0

0
−.02

−.02
Coefficients

Coefficients
−.04

−.04
−.06

−.06
−.08

−.08

0 6 12 18 24 30 36 0 6 12 18 24 30 36
Tenure cutoffs (in months) Tenure cutoffs (in months)

Dots represent regression coefficients, and vertical lines represent 90% confidence intervals. For
example, consider the second vertical line in the left panel (for college graduates), corresponding
to a tenure cutoff of six months. In a log-wage regression, the coefficient of “number of previous
jobs with tenure less than 6 months” is -0.045. The underlying regressions are the same as those
in Table 3, apart from the measures of work history (i.e., in the regressions underlying Figure 3,
the work-history measure is the “number of previous jobs with tenure less than X months”,
where X is indicated on the horizontal axis).

41
Appendix

A Proof of Proposition 1: Symmetric Learning Model

When period 2 ends, the zero-profit condition yields the market offer,

V3o (θ ) = θ + η̄θ ,

where η̄θ = E (η |θ ). The incumbent offer is

V3 θ, η j = min θ + η j , V3o (θ ) = θ + min η j , η̄θ , j ∈ {1, 2} .

 

Therefore, a worker quits if and only if η j < η̄θ . The accepted incumbent offer, Ṽ3 (θ ), is the same
as the market offer, i.e., Ṽ3 (θ ) = V3o (θ ).
When period 1 ends, the market offer, V2o (θ ), is determined by the zero-expected-profit condi-
tion,
V2o (θ ) = θ + η̄θ + E {(η2 − η̄θ ) · 1 {η2 ≥ η̄θ }} ,

and the incumbent offer is

V2 (θ, η1 ) = min {θ + η1 + (η1 − η̄θ ) · 1 {η1 ≥ η̄θ } , V2o (θ )} ,

where the first argument in the min is the maximum wage that the incumbent employer is willing
to offer. The worker will receive the same period-3 wage regardless of the turnover decision when
period 1 ended. Therefore, only the period-2 wage affects turnover.
The preceding implies that a worker quits if and only if

η1 + (η1 − η̄θ ) · 1 {η1 ≥ η̄θ } < η̄θ + E {(η2 − η̄θ ) · 1 {η2 ≥ η̄θ }} .

Since given θ the right-hand side is a constant, and the left-hand side is a strictly increasing func-
∗ , which solves
tion of η1 , there exists a unique threshold value, ηθs1

∗ ∗ ∗
ηθs1 + (ηθs1 − η̄θ ) · 1 {ηθs1 ≥ η̄θ } = η̄θ + E {(η2 − η̄θ ) · 1 {η2 ≥ η̄θ }} , (A.1)

∗ . Note that η ∗ > η̄ , since

such that when period 1 ends, a worker quits if and only if η1 < ηθs1 θs1 θ
the second term on the right-hand side of equation (A.1) is positive.58 Therefore, equation (A.1)
becomes
∗ 1
ηθs1 = η̄θ + E {(η2 − η̄θ ) · 1 {η2 ≥ η̄θ }} .
2
Again, the accepted incumbent offer matches the market offer, i.e., Ṽ2 (θ ) = V2o (θ ).
The period-1 wage is determined by the zero-expected-profit condition given free entry to the
58 Intuitively, this is due to convexity in the expected match quality, given the option of quitting again after period 2.

42
labor market,

∗
V1 = θ̄ + η̄ + E {[η1 − η̄θ − E {(η2 − η̄θ ) · 1 {η2 ≥ η̄θ }}] · 1 {η1 ≥ ηθs1 }
+ (η1 − η̄θ ) · 1 {η1 ≥ η̄θ }} . 

B Proof of Proposition 2: Asymmetric Learning Model

The proof has two steps. First, the existence and uniqueness of y∗ is shown. Second, the
existence of (ŷ1 , ŷ2 (s) , ŷ2 (q)) is shown.
In the first step, the intermediate value theorem is applied to prove the existence of y∗ .
Define the right-hand side of Equation (19) as

R ( x ) = E [θ |θ + η1 < x ] + η̄
ˆ ηH
= f η1 (η1 ) E [θ |θ < x − η1 , η1 ] dη1 + η̄
ηL
ˆ ´ x − η1
ηH
θL θdFθ |η1 (θ |η1 )
= f η1 ( η 1 ) dη1 + η̄
ηL Fθ |η1 ( x − η1 )
ˆ ( ´ x − η1 )
ηH
θL Fθ |η1 (θ |η1 ) dθ
= f η1 ( η 1 ) ( x − η1 ) − dη1 + η̄.
ηL Fθ |η1 ( x − η1 )

When x = θ L + η L , R ( x ) = θ L + η̄ > θ L + η L = x.
When x = θ̄ + η̄, R ( x ) = E[θ |θ < θ̄ + η̄ − η1 ] + η̄ < θ̄ + η̄ = x.
By the intermediate value theorem, there exists a solution to Equation (19), y∗ ∈ θ L + η L , θ̄ + η̄ .

Since the slope of the left-hand side in Equation (19) is 1, if the slope of the right-hand side is
always less than 1 in θ L + η L , θ̄ + η̄ , then y∗ is unique.

ˆ ( ´ x − η1 )
∂R ( x ) ηH f θ | η1 ( x − η 1 ) θL Fθ |η1 (θ |η1 ) dθ Fθ |η1 ( x − η1 )
= f η1 ( η 1 ) 1+ − dη1
∂x ηL Fθ |η1 ( x − η1 )2 Fθ |η1 ( x − η1 )
ˆ ( ´ x − η1 )
ηH f ( x − η1 ) θL Fθ |η1 (θ |η1 ) dθ
= f η1 ( η 1 ) dη1
ηL Fθ |η1 ( x − η1 )2

Following Heckman and Honore (1990), define

ˆ z
Fj+1 (z) = Fj (θ )dθ, where F0 = F (z) ≡ Pr (θ ≤ z) .
θL

Then
ˆ ( )
∂R ( x ) ηH
F100 ( x − η1 ) F1 ( x − η1 )
= f η1 ( η 1 ) dη1 .
∂x ηL F10 ( x − η1 )2
F100 ( x −η1 ) F1 ( x −η1 ) F100 ( x −η1 ) F1 ( x −η1 )
From log-concavity, ∈ [0, 1]. A necessary condition for = 1 is
F10 ( x −η1 )2 F10 ( x −η1 )2

43
F ( x ) ∝ ecx , with c > 0, which is ruled out by the assumption in Subsection 2.1. Therefore,
ˆ ηH
∂R ( x )
< f η1 (η1 ) dη1 = 1.
∂x ηL

Thus, y∗ is unique.59
In the second step, define ŷ = (ŷ1 , ŷ2 (s) , ŷ2 (q)), and D3 = [θ L + η L , θ H + η H ] × y∗ , θ̄ + η̄
 

× [θ L + η L , y∗ ], which is a compact, convex set. Given ŷ1 < ŷ2 (s), define mappings

T (ŷ) = ( T1 (ŷ) , T2 (ŷ) , T3 (ŷ)) ,

where

T1 (ŷ) = E [θ |y1 < ŷ1 ] + η̄ + E {[y2 − ŷ2 (q)] · 1 (y2 ≥ ŷ2 (q)) |y1 < ŷ1 }
+ E [θ |y2 < ŷ2 (q) , y1 < ŷ1 ] − E [θ |ŷ1 ≤ y1 < ŷ2 (s)] ,
T2 (ŷ) = E [θ |ŷ1 ≤ y1 < ŷ2 (s)] + η̄,
T3 (ŷ) = E [θ |y2 < ŷ2 (q) , y1 < ŷ1 ] + η̄.

The Brouwer fixed-point theorem is applied to prove there exists a fixed point of ŷ = T (ŷ) in
D3 .
It is straightforward to verify the following:

T2 (ŷ) ≥ E [θ |θ L + η L ≤ y1 < y∗ ] + η̄ = E [θ |y1 < y∗ ] + η̄ = y∗ ,

   
T2 (ŷ) ≤ E θ |y1 = θ̄ + η̄ + η̄ = θ̄ + η̄ − η̄ + η̄ = θ̄ + η̄.

T3 (ŷ) ≥ E [θ |y2 < θ L + η L , y1 < θ L + η L ] + η̄ = θ L + η̄ > θ L + η L ,

T3 (ŷ) ≤ E [θ |y2 < y∗ , y1 < θ H + η H ] + η̄ = E [θ |y2 < y∗ ] + η̄ = y∗ .

T1 (ŷ) ≥ E [θ |y1 < θ L + η L ] + η̄ + E [θ |y2 < θ L + η L , y1 < θ L + η L ]

 
− E θ |y1 = θ̄ + η̄


= θ L + η̄ + θ L − θ̄ + η̄ − η̄ = 2θ L + η̄ − θ̄
≥ θ L + ηL .
59 This y∗ is actually the equilibrium threshold value in a two-period asymmetric learning model, where a worker

stays with the incumbent firm if and only if y1 ≥ y∗ when period 1 ends.

44
 T1 (ŷ) ≤ E [θ |y1 < θ H + η H ] + η̄ + E [y2 − (θ L + η L ) |y1 < θ H + η H ]
+ E [ θ | y2 < y ∗ , y1 < θ H + η H ] − E [ θ | θ L + η L ≤ y1 < y ∗ ]
≤ θ̄ + η̄ + θ̄ + η̄ − (θ L + η L ) + y∗ − y∗
 

= 2θ̄ + 2η̄ − (θ L + η L )
≤ θH + ηH .

So T (ŷ) is a continuous mapping from D3 to D3 . According to the Brouwer fixed-point theo-

rem, there exists a fixed point which is the solution to equations (13), (17), and (18).
Equilibrium wage setting is proved in the text. 

C Proof of Proposition 3: Asymmetric versus Symmetric Learning

First, notice that in the symmetric-learning model, the worker’s turnover decisions in all peri-
ods are independent of ability. Whether the worker quits the incumbent employer or not depends
entirely on the current match quality. During period 3, a worker with ability θ is paid θ + η̄ regard-
less of work history. In contrast, if learning is asymmetric, then W3o (s)= ŷ2 (s) and W3o (q)= ŷ2 (q).
From Equations 17 and 18,

W3o (s) = ŷ2 (s) = E [θ |ŷ1 ≤ y1 < ŷ2 (s)] + η̄

> E [θ |y1 < ŷ1 ] + η̄
≥ E [θ |y2 < ŷ2 (q) , y1 < ŷ1 ] + η̄
= ŷ2 (q) = W3o (q)

Therefore, W3o (s) > W3o (q). 

D Alternative Explanations
This section addresses four potential alternative explanations for the empirical results that a)
there is a job-hopping wage penalty for college graduates but not for the less educated, and b) that
wage penalty is mitigated when workers enter the labor market during economic downturns.

D.1 On-the-job Search

In a standard on-the-job search model, as in Burdett (1978) and Burdett and Mortensen (1998),
workers search for and switch jobs in pursuit of higher wages, moving only when such offers
appear. In general, such models predict a zero correlation between the current log-wage and the
number of previous jobs, after controlling for labor market experience. Specifically, in a standard

45
on-the-job search model with observable worker ability,60 the wage in the current period, t, is the

maximum wage offer received so far, i.e., wt = θ + max η j , j = t0 , ..., t , where t0 is the first period
since labor market entry (Burdett, 1978) or since the end of the last unemployment spell (Burdett
and Mortensen, 1998). The match quality in period j, η j , is a random draw from the common
distribution, G (η ). Therefore, the wage, wt , is a random variable with cumulative distribution
function G (wt − θ )t−t0 +1 . After controlling for ability, θ, and labor-market experience or the con-
tinuous employment spell, t − t0 , the conditional mean of wt is independent of all other measures,
including the number of previous jobs.
In a variant of the search model with heterogeneous firm productivity, as in Postel-Vinay and
Robin (2002), the wage is a function of the worker’s (observable) ability, the incumbent firm’s
productivity, and the distribution of productivity of all firms in the market. It is independent of
the number of previous jobs. Thus, an alternative model based on on-the-job search is inconsistent
with the result that college graduates suffer a wage penalty from job hopping.

D.2 Public Ability with High-Ability Workers Infrequently Changing Jobs

Suppose that ability is publicly observed but that higher-ability workers are less likely to
change jobs. Then job changers are, on average, lower-ability (and, hence, get paid less) even
though ability is perfectly observed. This argument, which has been suggested as a potential al-
ternative explanation for the result that college graduates (but not the less educated) suffer a wage
penalty from job hopping, is unappealing for at least three reasons. First, it assumes that (for what-
ever reason) high-ability workers are less likely to be job changers, whereas in the present model
this is a result rather than an assumption. Second, in contrast to the asymmetric-learning argu-
ment, the alternative does not offer an obvious explanation for why the job-hopping wage penalty
for college graduates is mitigated when workers enter the labor market during economic down-
turns. Third, the argument fails when ability is held constant. Controlling for ability, under the
alternative explanation current wages should not depend on prior work history, conditional on
work experience and current job tenure. Recall that AFQT is a control variable in the regressions.
One might argue that AFQT is an imperfect, error-laden measure of ability, so that the alternative
explanation is still somewhat viable. But in that case, when AFQT is dropped as a control the main
result should strengthen, which in fact does not happen, as seen in Table A3.

D.3 Symmetric Learning About Workers’ Tastes for Staying

Prior work history might serve as a signal even in a symmetric-learning model. Suppose that
employers (including, perhaps, the incumbent employer) do not know how likely a worker is
to invest in a long-term employment relationship until they observe the worker’s behavior. The
current employer might learn about that characteristic (i.e., the worker’s “tastes for staying”) at
the same time as the rest of the market, by observing the worker’s decisions to quit or stay. Thus,
60 The case in which ability is unobservable is described by the main model.

46
prior work history is used as a signal even though there is no asymmetric employer learning. Prior
job tenures are correlated with current wages because long tenures are seen by all employers as
suggesting stability, so they are all willing to pay extra for that.
This argument assumes that there is an individual characteristic—tastes for staying or for job
hopping—in the error term of regression model (21) that affects both work history and current
wages. Employers might not know the characteristic immediately, but can learn about it, or they
might observe it from the beginning. In either case, if econometricians do not observe it, there is an
endogeneity problem. Finding an instrumental variable for work history is challenging. Economic
conditions (either contemporaneously or at the point of labor-market entry) offer one potential
candidate. However, prior research (and, indeed, Table 4) has found that economic conditions
such as unemployment rates have direct effects on wages through various channels—human cap-
ital accumulation, search (Kahn, 2010; Oreopoulos et al., 2012), or implicit contracts (Beaudry and
DiNardo, 1991).
There are two reasons why a symmetric-learning explanation for the results for college grad-
uates is unappealing. First, if the “tastes for staying” argument applies to college graduates, it
is unclear why it would not also apply to less-educated workers. That is, even if the symmetric-
learning argument is accepted as the explanation for the results for college graduates, it leaves
the results for the less-educated group unexplained. Second, the asymmetric-learning model can
explain not only the main result from Table 3 but also the fact that this result (for rookie college
graduates) is mitigated when workers enter the labor market during an economic downturn. The
symmetric-learning argument does not offer an explanation for the latter result. The reason is that,
under the symmetric-learning argument, employers would need to believe that workers’ “tastes
for staying” are stronger when they happen to enter the labor market during an economic down-
turn. But that is implausible; it is expected that workers’ “tastes for staying” should be unrelated
to the state of the economy when they finish college and enter the labor market.

D.4 Human Capital Accumulation

An alternative to a learning model is one in which workers accumulate firm-specific or general

human capital, either by investing (Ben-Porath, 1967) or via learning-by-doing (Imai and Keane,
2004). For example, being a good fit for a job might allow a worker to improve her own type on
the job and earn a higher wage. Workers without long stints of previous employment will also not
have improved. This mechanism suggests that the current wage should be positively correlated
with the long stints of previous jobs. To investigate this possibility, an additional variable that
measures the longest previous tenure is included in the baseline regression model.
Table A4 shows that the correlation between the current wage and the longest previous tenure
is not statistically significant, while there still exists a significant penalty for job hopping.61 On the
61 The job-hopping wage penalty is similar for rookies and veterans among college graduates when the regression

includes the longest previous tenure. This is because this variable is more correlated with the job hopping variable for
rookies who have shorter labor market experience than for veterans.

47
other hand, if human capital is entirely firm specific, each job change makes prior human capital
irrelevant, and workers begin their next job with a clean slate. Therefore, a worker’s human-
capital level (and wage) would be affected by current job tenure but not by prior job hopping
history, which is not consistent with the data.

E Further Evidence From Employer Callbacks

Job hopping can potentially affect career-relevant labor-market outcomes other than wages.
Even before an employment relationship begins, a job applicant’s probability of receiving a call-
back from a prospective employer in response to a submitted job application may be reduced if
the applicant’s résumé exhibits a history of job hopping.
Bertrand and Mullainathan (2004) conduct an experiment by sending out fictitious résumés
(i.e., job applications) in response to help-wanted ads posted in newspapers. The experiment’s
purpose is to determine whether (pseudo) applicants whose names are commonly associated with
non-whites are called back by prospective employers at a lower rate than those whose names are
commonly associated with whites. Each résumé includes the number of prior jobs held. Those
data are used here to investigate the correlation between the probability of receiving an employer
callback and the number of prior jobs listed on the résumé. The results suggest job-hopping penal-
ties for employer callbacks—and in particular the difference between college graduates and the
less educated—evoking the present results for current wages.
The following logit model is estimated:

Prob (callback i = 1) = Logit Xi0 β + njobsi · γ ,

(E.1)

where callback i is a dummy variable equaling one if résumé i receives an employer callback. The
vector Xi includes résumé and job ad characteristics. The independent variable njobsi is the num-
ber of prior jobs listed on the résumé.62
Table A5 reports the logit results for various groups of résumés. For the whole sample, column
1 reveals that the effect of the number of jobs listed on the résumé on the likelihood of an employer
callback is negative but statistically insignificant. However, the effect is negative and statistically
significant for those with six or fewer years of experience listed on the résumé (column 2), while it
becomes positive but insignificant for those with more than six years of experience (column 3).63
One additional job listed on the résumé is associated with a 1.3-percentage-point decrease in the
probability of a callback for those rookie applicants.64
The same logit models are then estimated separately for each education subsample, namely
college graduates (columns 4-6), those with some college (columns 7-9), and high-school graduates
62 On the 4870 résumés the number of jobs ranges from 1 to 7, with median 4, mean 3.661, and standard deviation
1.219. The correlation between the number of jobs and being black is essentially zero (-0.0025).
63 The results are similar if the threshold is five years. Both are statistically insignificant at conventional levels if the

threshold is seven years.

64 The mean of the callback dummy is 0.08 for the whole sample.

48
or dropouts (column 10).65 The results for college graduates almost replicate those for the whole
sample in the first three columns. The results for those with some college are somewhat different—
the number of jobs listed on the résumé is negatively associated with the likelihood of receiving an
employer callback, regardless of the experience group, with a larger effect for rookie applicants.
However, it is only statistically significant for the whole subsample. For the subsample of high-
school graduates or dropouts, the effect is positive but statistically insignificant at conventional
levels.66
In summary, for those résumés with at least some college education, the number of prior jobs
listed on the résumé is negatively related to the likelihood of an employer callback; that is espe-
cially so within the first six years of experience, but there is no such negative effect for those with
a high-school degree or less. This pattern of results, particularly the differing results by education
level, suggests the pattern of evidence that has been documented for current wages may extend
to other career-relevant labor market outcomes.

65 There are only 287 observations in the subsample of high-school graduates or dropouts, so the statistical model is
estimated on both groups combined.
66 The results from these callback data, which are based on a different dependent variable, are weaker than the main

results from the wage analysis. One possible explanation is that callbacks are only the first screening step during the
recruitment process. Employers have many other steps during the recruitment process to screen job candidates. For
example, some employers may want to interview a candidate with many previous jobs to see if the candidate has a
good explanation for the job hopping.

49
Table A1: Current Log-wages and Job Hopping: Number of Previous Quits and Shortest Previous
Tenure

1 2 3 4 5 6
College graduates High school graduates or less
All Rookies Veterans All Rookies Veterans
No. of previous quits -0.027** -0.033** -0.028** -0.002 0.004 -0.003
(0.011) (0.016) (0.012) (0.003) (0.005) (0.003)
Shortest previous tenure 0.005 0.031* 0.004 0.003 0.005 0.002
(0.004) (0.016) (0.004) (0.003) (0.007) (0.003)
AFQT score 0.083*** 0.065*** 0.095*** 0.026*** 0.016** 0.048***
(0.020) (0.024) (0.029) (0.007) (0.008) (0.013)
AFQT×Experience 0.002 0.009 0.001 0.005*** 0.008*** 0.003***
(0.002) (0.007) (0.002) (0.001) (0.003) (0.001)
R-squared 0.37 0.31 0.29 0.33 0.25 0.30
Observations 13,151 3,942 9,209 67,191 22,352 44,839
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust Huber/White standard errors are in parentheses,
clustered at the individual level.
"Rookies" refer to the first five years in the labor market and "Veterans" refer to periods af-
ter the first five years. All regressions include experience and its square, current tenure and
its square, years of schooling, the interactions of years of schooling with experience, His-
panic, Black, urban residency, USA residency at age 14, a dummy variable for the interview
year, dummy variables for years, the interactions of year dummies with years of schooling,
dummy variables for one-digit industry and occupation codes, and (for high school gradu-
ates or less) a dummy for finishing high school.

50
 Table A2: Comparison of Returns to Experience and Tenure (Including or Excluding Work-history Regressor)

1 2 3 4 5 6 7 8 9 10 11 12
College graduates High school graduates or less
All Rookies Veterans All Rookies Veterans
Include work history Yes No Yes No Yes No Yes No Yes No Yes No
Experience 0.069*** 0.066*** 0.156*** 0.135*** 0.051*** 0.051*** 0.056*** 0.055*** 0.088*** 0.088*** 0.046*** 0.045***
(0.007) (0.007) (0.024) (0.022) (0.011) (0.011) (0.003) (0.003) (0.005) (0.005) (0.004) (0.004)
Experience2 -0.002*** -0.002*** -0.013*** -0.011*** -0.001*** -0.001*** -0.001*** -0.001*** -0.005*** -0.005*** -0.001*** -0.001***
(0.0003) (0.0003) (0.004) (0.004) (0.0003) (0.0003) (0.0003) (0.0003) (0.001) (0.001) (0.0002) (0.0002)
Current tenure 0.009*** 0.012*** 0.018* 0.030*** 0.008*** 0.011*** 0.016*** 0.016*** 0.044*** 0.044*** 0.015*** 0.015***

51
(0.003) (0.003) (0.010) (0.010) (0.003) (0.003) (0.001) (0.001) (0.004) (0.004) (0.001) (0.001)
AFQT 0.104*** 0.107*** 0.088*** 0.087*** 0.108*** 0.112*** 0.068*** 0.068*** 0.033*** 0.033*** 0.088*** 0.088***
(0.019) (0.019) (0.018) (0.018) (0.022) (0.022) (0.007) (0.007) (0.006) (0.006) (0.010) (0.010)
R-squared 0.37 0.36 0.30 0.30 0.28 0.28 0.32 0.32 0.25 0.25 0.29 0.29
Observations 13,151 13,151 3,942 3,942 9,209 9,209 67,191 67,191 22,352 22,352 44,839 44,839
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust Huber/White standard errors are in parentheses, clustered at the individual level.
"Rookies" refer to the first five years in the labor market, and "Veterans" refer to periods after the first five years. Odd-numbered columns (1, 3,...11) include
the number of previous quits, while even-numbered columns (2, 4, ...12) do not. All regressions include years of schooling, Hispanic, Black, urban residency,
USA residency at age 14, a dummy variable for the interview year, dummy variables for years, the interactions of year dummies with years of schooling,
dummy variables for one-digit industry and occupation codes, and (for high school graduates or less) a dummy for finishing high school.
 Table A3: Current Log-wages and Job Hopping (Excluding AFQT)

1 2 3 4 5 6
College graduates High school graduates or less
All Rookies Veterans All Rookies Veterans
Number of previous quits -0.032*** -0.040*** -0.033*** -0.003 0.004 -0.004
(0.011) (0.015) (0.011) (0.003) (0.005) (0.003)
R-squared 0.36 0.29 0.27 0.32 0.25 0.28
Observations 13,151 3,942 9,209 67,191 22,352 44,839
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust Huber/White standard errors are in parentheses,
clustered at the individual level.
"Rookies" refer to the first five years, and "Veterans" refer to periods after the first five years.
Regressions are identical to those in Table 3 but dropping the AFQT score and the interaction
of the AFQT score with experience.

Table A4: Current Log-wages and Job Hopping, with Longest Previous Tenure

1 2 3 4 5 6
College graduates High school graduates or less
All Rookies Veterans All Rookies Veterans
No. of previous quits -0.033*** -0.033*** -0.034*** -0.001 0.003 -0.002
(0.011) (0.017) (0.011) (0.003) (0.006) (0.003)
Longest previous tenure -0.007 0.038 -0.008 0.002 0.003 0.002
(0.006) (0.033) (0.006) (0.003) (0.011) (0.003)
AFQT score 0.081*** 0.064** 0.093*** 0.026*** 0.016** 0.048***
(0.020) (0.024) (0.029) (0.007) (0.008) (0.013)
AFQT×Experience 0.002 0.010 0.001 0.005*** 0.008*** 0.003***
(0.002) (0.007) (0.002) (0.001) (0.003) (0.001)
R-squared 0.37 0.31 0.29 0.33 0.25 0.30
Observations 13,151 3,942 9,209 67,191 22,352 44,839
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust Huber/White standard errors are in parentheses,
clustered at the individual level.
"Rookies" refer to the first five years in the labor market and "Veterans" refer to periods after
the first five years. All regressions include experience and its square, current tenure and its
square, years of schooling, the interactions of years of schooling with experience, Hispanic,
Black, urban residency, USA residency at age 14, a dummy variable for the interview year,
dummy variables for years, the interactions of year dummies with years of schooling, dummy
variables for one-digit industry and occupation codes, and (for high school graduates or less)
a dummy for finishing high school.

52
 Table A5: Employer Callback Probability and Job Hopping

1 2 3 4 5 6 7 8 9 10
The Whole Sample a College Graduates Some College High School
all exp<=6c exp>6d all exp<=6 exp>6 all exp<=6 exp>6 or lessb
Number of prior jobs -0.007 -0.013* 0.003 -0.005 -0.014* 0.002 -0.025** -0.014 -0.007 0.031
(0.004) (0.007) (0.007) (0.004) (0.008) (0.008) (0.012) (0.020) (0.019) (0.025)
Has employment gaps 0.022** 0.038** -0.005 0.020 0.041* -0.005 0.007 0.029 0.047 -0.077
(0.011) (0.017) (0.018) (0.013) (0.023) (0.020) (0.026) (0.060) (0.035) (0.077)
Black -0.033*** -0.025*** -0.038*** -0.028*** -0.018* -0.031** -0.054*** -0.044** -0.059** -0.049
(0.007) (0.008) (0.011) (0.008) (0.010) (0.012) (0.014) (0.019) (0.026) (0.037)
Female 0.008 -0.017* 0.029 0.016 -0.008 0.037** -0.067 -0.053 0.021 -0.089
(0.011) (0.010) (0.018) (0.010) (0.011) (0.018) (0.052) (0.070) (0.062) (0.054)
Experience 0.008*** 0.037 0.019** 0.006* 0.011 0.016* 0.007 0.071 -0.017 0.033*
(0.003) (0.029) (0.008) (0.004) (0.049) (0.009) (0.008) (0.045) (0.023) (0.018)
Experience2 -0.000** -0.004 -0.001** -0.000 -0.001 -0.000 -0.000 -0.010* 0.001 -0.001

53
(0.000) (0.003) (0.000) (0.000) (0.005) (0.000) (0.000) (0.006) (0.001) (0.001)
Observations 4,784 2,404 2,380 3,436 1,636 1,788 990 527 463 287
*** p < 0.01, ** p < 0.05, * p < 0.1. Robust standard errors in parentheses are clustered at the employment-ad level.
Each column gives the results of a logit regression where the dependent variable is the callback dummy equaling one if the applicant
receives an employer callback and zero otherwise. Reported in the table are estimated marginal changes in probability for the contin-
uous variables and estimated discrete changes for the dummy variables. All regressions include a city dummy, (for Columns 1-3 only)
education dummies or (for Column 10 only) a dummy for graduating from high school, a vector of dummy variables for résumé’s char-
acteristics (volunteering, military, having an email address, working while at school, honors, computer skills, and special skills), a vector
of dummy variables for job requirements as listed in the employment ad (experience, computer skills, communication skills, organiza-
tion skills, and education), six occupation dummies, a dummy if the ad mentions it is an equal opportunity employer (EOE), and some
demographic variables in the applicant’s zip code (fraction of blacks, log median household income, fraction of high school dropouts,
fraction of college graduates, and log per capita income).
a The entire set of sent résumés.
b The high school graduates and dropouts.
c Those with 6 or fewer years of experience.
d Those with more than 6 years of experience.