Recessions and the costs of job loss.
Davis, Steven J. ; von Wachter, Till
ABSTRACT We develop new evidence on the cumulative earnings losses
associated with job displacement, drawing on longitudinal Social
Security records from 1974 to 2008. In present-value terms, men lose an
average of 1.4 years of predisplacement earnings if displaced in
mass-layoff events that occur when the national unemployment rate is
below 6 percent. They lose a staggering 2.8 years of predisplacement
earnings if displaced when the unemployment rate exceeds 8 percent.
These results reflect discounting at a 5 percent annual rate over 20
years after displacement. We also document large cyclical movements in
the incidence of job loss and job displacement and present evidence on
how worker anxieties about job loss, wage cuts, and job opportunities
respond to contemporaneous economic conditions. Finally, we confront
leading models of unemployment fluctuations with evidence on the
present-value earnings losses associated with job displacement. The 1994
model of Dale Mortensen and Christopher Pissarides, extended to include
search on the job, generates present-value losses that are only
one-fourth as large as observed losses. Moreover, present-value losses
in the model vary little with aggregate conditions at the time of
displacement, unlike the pattern in the data.
**********
Major economic downturns bring large increases in permanent layoffs
among workers with long tenure on the job. We refer to this type of job
loss event as a displacement. Previous research shows that job
displacements lead to large and persistent earnings losses for the
affected workers. (1) The available evidence also indicates that job
displacement leads to less stability in earnings and employment, worse
health outcomes, higher mortality, lower educational achievement by the
children of displaced workers, and other unwelcome consequences. (2)
We develop new evidence on the cumulative earnings losses
associated with job displacement and the role of labor market conditions
at the time of displacement. In present-value terms, men lose an average
of 1.4 years of predisplacement earnings if displaced in mass-layoff
events that occur when the national unemployment rate is below 6
percent. They lose a staggering 2.8 years of predisplacement earnings if
displaced when the unemployment rate exceeds 8 percent. These results
reflect discounting at a 5 percent annual rate over 20 years after
displacement. We also document large cyclical movements in the incidence
of job loss and job displacement, and we investigate how worker
anxieties about job loss, wage cuts, and other labor market prospects
respond to contemporaneous economic conditions. Finally, we confront
leading models of unemployment fluctuations in the tradition of work by
Peter Diamond, Dale Mortensen, and Christopher Pissarides with evidence
on the present-value earnings losses associated with job displacement.
Our study builds on three major areas of research: empirical work
on cyclical fluctuations in job destruction, job loss, and unemployment;
empirical work on earnings losses and other outcomes associated with job
displacement; and theoretical work on search-and-matching models of
unemployment fluctuations along the lines of Mortensen and Pissarides
(1994). In terms of a broad effort to bring together these areas of
research, the closest antecedent to our study is that by Robert Hall
(1995). In terms of its effort to confront equilibrium
search-and-matching models with evidence on the earnings losses
associated with job displacement, the closest prior work is that by
Wouter Den Haan, Garey Ramey, and Joel Watson (2000).
Our empirical investigation of the earnings losses associated with
job displacement draws heavily on recent research by von Wachter, Jae
Song, and Joyce Manchester (2011). They develop new evidence on the
short- and long-term earnings effects of job loss using longitudinal
Social Security records covering more than 30 years. Our first main
contribution is to characterize, drawing on their estimated empirical
models, how present-value earnings losses due to job displacement vary
with business cycle conditions at the time of displacement. For men with
3 or more years of job tenure who lose jobs in mass-layoff events at
larger firms, job displacement reduces the present value of future
earnings by 12 percent in an average year. The present-value losses are
high in all years, but they rise steeply with the unemployment rate in
the year of displacement. Present-value losses for displacements that
occur in recessions are nearly twice as large as for displacements in
expansions. The entire future path of earnings losses is much higher for
displacements that occur in recessions. In short, the present-value
earnings losses associated with job displacement are very large, and
they are highly sensitive to labor market conditions at the time of
displacement.
Drawing on data from the General Social Survey of the National
Opinion Research Center and from Gallup polling, we also examine the
relationship of anxieties about job loss, wage cuts, ease of job
finding, and other labor market prospects to actual labor market
conditions. The available evidence indicates that cyclical fluctuations
in worker perceptions and anxieties track actual labor market conditions
rather closely, and that they respond quickly to deteriorations in the
economic outlook. The Gallup data, in particular, show a tremendous
increase in worker anxieties about labor market prospects after the peak
of the financial crisis in 2008 and 2009. They also show a recent return
to the same high levels of anxiety. These data suggest that fears about
job loss and other negative labor market outcomes are themselves a
significant and costly aspect of economic downturns for a broad segment
of the population. These findings also imply that workers are well aware
of and concerned about the costly nature of job loss, especially in
recessions.
Our second main contribution is to analyze whether leading
theoretical models of unemployment fluctuations can account for our
evidence on the magnitude and cyclicality of present-value earnings
losses associated with job displacement. Following Hall and Paul Milgrom
(2008), we consider three variants of the basic Mortensen-Pissarides
model analyzed by Robert Shimer (2005) and many others. We also consider
a richer model by Simon Burgess and Helene Turon (2010) that introduces
search on the job and replacement hiring into the model of Mortensen and
Pissarides (1994). The richer model generates worker flows apart from
job flows, heterogeneity in productivity and match surplus values, and
recessionary spikes in job destruction, job loss, and unemployment
inflows of the sort we see in the data.
The search-and-matching models we consider do not account for our
evidence on the present-value earnings losses associated with job
displacement. The empirical losses are an order of magnitude larger than
those implied by basic versions of the Mortensen-Pissarides model. Wage
rigidity of the form considered by Hall and Milgrom (2008) greatly
improves the model's ability to explain aggregate unemployment
fluctuations, but it does not bring the model closer to evidence on the
earnings losses associated with displacement. The model of Burgess and
Turon (2010) generates larger present-value losses, because most
job-losing workers in the model do not immediately recover
predisplacement wage levels upon reemployment. Instead, unemployed
persons tend to flow into jobs on the lower rungs of the wage
distribution and move up the distribution over time. Yet when calibrated
for consistency with U.S. unemployment flows, the model of Burgess and
Turon yields present-value earnings losses due to job loss less than
one-fourth as large as the empirical losses. Moreover, present-value
losses in the model vary little with aggregate conditions at the time of
displacement, unlike the pattern in the data.
Present-value income (as opposed to earnings) losses associated
with job loss are even smaller in the search models we consider. Indeed,
a fundamental weakness of these models is their implication that job
loss is a rather inconsequential event from the perspective of
individual welfare. In this sense, and despite many virtues and
attractions, this class of models fails to address a central reason that
job loss, unemployment, and recessions attract so much attention and
concern from economists, policymakers, and others. For the same reason,
care should be taken in using this class of models to form conclusions
about the welfare effects of shocks and government policies.
The paper proceeds as follows. Section I presents evidence on the
incidence of job destruction, layoffs, unemployment inflows, and job
displacement over the business cycle. Section II first summarizes
previous research on the short- and long-term consequences of job
displacements for earnings. It then draws on work by von Wachter and
others (2011) to estimate near-term and present-value earnings losses
associated with job displacement, and to investigate how the losses vary
with business cycle conditions at displacement. Section III reviews
previous work on the nonmonetary costs of displacement and presents
evidence on cyclical fluctuations in perceptions and anxieties related
to labor market prospects. Section IV considers selected equilibrium
search-and-matching models of unemployment fluctuations and evaluates
their implications for the earnings and income losses associated with
job loss. Section V concludes.
I. The Incidence of Job Loss and Job Displacement over Time
Figure 1 displays four time series that draw on different sources
of data and pertain to different concepts of job loss. The job
destruction measure captures gross employment losses summed over
shrinking and closing establishments in the Business Employment Dynamics
(BED) database. (3) The layoff measure reflects data on
employer-initiated separations, as reported by employers in the Job
Openings and Labor Turnover Survey (JOLTS) and as aggregated and
extended back to 1990 by Davis, Jason Faberman, and John Haltiwanger
(2012). (4) We calculate unemployment inflow rates using monthly Current
Population Survey (CPS) data on the number of employed persons and the
number unemployed less than 5 weeks. Summing over months yields the
quarterly rates. The measure of initial unemployment insurance (UI)
claims is the quarterly sum of weekly new claims for UI benefits,
expressed as a percent of nonfarm payroll employment.
Figure 1 highlights two key points. First, the sheer volume of job
loss and unemployment incidence is enormous--in good economic times and
bad. For example, the JOLTS-based layoff rate averages 7 percent per
quarter from 1990 to 2011. Multiplying this figure by nonfarm payroll
employment in 2011 yields about 9 million layoffs per quarter. Quarterly
averages for job destruction and unemployment inflows are of similar
magnitude. Initial UI claims average about 5 million per quarter. In
short, the U.S. economy routinely accommodates huge numbers of lost jobs
and unemployment spells.
Many, perhaps most, of these job loss events involve little
financial loss or other hardship for individuals and families. Indeed,
the high rates shown in figure 1 reflect an impressive capacity for
constant renewal and productivity-enhancing reallocation of jobs,
workers, and capital in the economy as a whole. (5) It is important to
keep this point in mind when interpreting the evidence on the costs
associated with job displacement. That evidence focuses, quite
deliberately, on the types of job loss events that often involve serious
consequences for workers and their families.
[FIGURE 1 OMITTED]
Second, all four series in figure 1 exhibit strongly
countercyclical movements, with clear spikes in the three recessions
covered by our sample period. (6) For example, the quarterly layoff rate
rises by 129 basis points from 1990Q2 to 1991Q1, 85 basis points from
2000Q2 to 2001Q4, and 208 basis points from 2007Q3 to 2009Q1.
Interestingly, each measure in figure 1 starts to rise before the onset
of a recession (as dated by the National Bureau of Economic Research)
and turns down before the resumption of an expansion. This pattern
confirms the well-known usefulness of initial UI claims as a leading
indicator for business cycles, and it suggests that other job loss
indicators behave similarly in this respect. (7)
Much of our study examines the earnings losses of long-tenure male
workers who lose jobs in large-scale layoff events. To quantify those
losses, we follow individual workers over time using annual earnings
records maintained by the Social Security Administration (SSA). Figure 2
plots an annual job displacement measure for men constructed from the
SSA data and compares it with annual measures of job destruction and
initial claims for unemployment insurance benefits. Here, we report
displacement rates in the population of male employees 50 years or
younger with at least 3 years of prior job tenure, excluding government
workers and certain services industries not covered by the Social
Security system throughout our full sample period. Also shown are annual
series for two measures of job destruction from the Census Bureau's
Business Dynamics Statistics (BDS) program and initial claims for UI
benefits. (8)
We regard a worker as displaced in year y if he separates from his
employer in y and the employer experiences a mass-layoff event in y. We
say a worker "separates" from an employer in year y when he
has earnings from the employer in y - 1 but not in y. To meet the prior
job tenure requirement, the worker must have positive earnings from the
employer in question in y - 3, y - 2, and y - 1. To qualify as a
mass-layoff event in year y, the employer must meet the following
criteria: 50 or more employees in y - 2; employment contracts by 30 to
99 percent from y - 2 to y; employment in y - 2 is no more than 130
percent of employment in y - 3; and employment in y + 1 is less than 90
percent of employment in y - 2. The 99 percent cutoff in the second
condition ensures that we do not capture spurious firm deaths due to
broken longitudinal links. The last two conditions exclude temporary
fluctuations in firm-level employment. Although these criteria miss some
displacements of long-tenure workers at larger employers, they help
ensure that the separations we identify as job displacement events are
indeed the result of permanent layoffs. (9) To qualify as a job
displacement event in y, we also require that the separation be from the
worker's main job, defined as the one that accounts for the largest
share of his earnings in y - 2. For additional details on the data,
sample, and measurement procedures, see von Wachter and others (2011).
[FIGURE 2 OMITTED]
To express job displacements in year y as a rate in figure 2, we
divide by the number of male workers 50 or younger in y - 2 with at
least 3 years of job tenure at firms with 50 or more employees in the
industries covered by Social Security throughout our sample period.
These workers make up 31 to 36 percent of all male workers 50 or younger
in industries continuously covered by the SSA from 1980 to 2008,
depending on the year, 40 to 48 percent when we also restrict attention
to those with 3 or more years of job tenure, and 70 to 74 percent when
we further narrow the focus to firms with 50 or more employees.
The annual frequency of the measures in figure 2 somewhat obscures
the timing of cyclical movements, but the broad patterns echo those in
figure 1: job loss rates move in a countercyclical manner, and
recessions involve notable jumps in job loss. The deep recession in the
early 1980s saw dramatic increases in rates of job destruction and job
displacement. For example, the annual job destruction rate at firms with
50 or more employees rose from 11.6 percent in 1979 to 18.3 percent in
1983. (To be clear, the latter figure reflects establishment-level
employment contractions that occur from March 1982 to March 1983.) Our
measure of the job displacement rate rose from 1.9 percent in 1980 to
5.0 percent in 1983. (10) More generally, the job displacement rate is
roughly 20 to 25 percent as large as annual job destruction rates,
although it is worth stressing that the two measures pertain to
different at-risk populations.
The incidence of job displacement might seem modest in any given
year, but it cumulates to a large number during severe downturns. For
example, summing the job displacement rates in figure 2 from 1980 to
1985 yields a cumulative displacement rate of more than 20 percent. (11)
This figure translates to about 2.7 million job displacement events over
the 6-year period among men 50 years or younger with 3 or more years of
job tenure and working in industries with continuous SSA coverage. This
figure is conservative, given our restrictive criteria for mass-layoff
events. According to the Displaced Worker Supplement to the CPS, 6.9
million persons with at least 3 years of prior tenure lost jobs due to
layoffs from 2007 to 2009 (Bureau of Labor Statistics 2010). This figure
includes women and does not impose our mass-layoff criteria. The Bureau
of Labor Statistics also reports that an additional 8.5 million persons
were displaced in 2007-09 from jobs held less than 3 years.
The top panel of figure 3 shows displacement rates for men with 3
to 5 years of job tenure and for men with 6 or more years. We impose the
same requirements for age, firm size, industry coverage, and mass-layoff
events as before. Displacement rates are considerably higher for workers
with 3 to 5 years of tenure and more cyclically sensitive in the
relatively shallow recessions and weak labor markets of the early 1990s
and 2000s. These patterns conform to the view that workers with lower
job tenure face greater exposure to negative firm-specific and aggregate
shocks. The bottom panel shows displacement rates for men in three broad
age groups. The basic pattern is clear: younger men tend to be more
exposed to negative firm-specific and aggregate shocks that lead to job
destruction.
Together, the two panels of figure 3 show that longer job tenure
and greater labor market experience afford some insulation from the
vicissitudes of firm-level employment fluctuations. However, it is well
worth noting that tenure and experience provide less insulation in the
deep aggregate downturn in the early 1980s. This aspect of figure 3
suggests that severe recessions bite especially deeply into the
distribution of valuable employment relationships. Evidence below on the
cyclical behavior of the earnings losses associated with job loss
supports this view as well.
[FIGURE 3 OMITTED]
II. The Long-Term Earnings Effects of Job Displacement
We turn now to evidence on the earnings losses associated with job
displacement.
II.A. Previous Research
A growing body of research finds that job displacements often lead
to large, persistent earnings losses. Most studies estimate the effect
as the change in earnings from before to after the job loss relative to
the contemporaneous earnings change of comparable workers who did not
lose jobs. Studies differ somewhat in how they measure job loss and how
they define the control group of nondisplaced workers.
Following earlier research, von Wachter and others (2011) define
job displacement as the separation of a "stable" worker from
his main employer during a period when the employer experiences a
lasting employment decline of at least 30 percent. A stable worker is
one with positive earnings at the firm in each of the three years
immediately preceding the displacement event. Their definition also
requires the employer to have at least 50 employees in the baseline
period before the mass lay-off. They exclude workers in two-digit
industries not covered by SSA in the early 1980s, chiefly the public
sector. Comparing the evolution of annual earnings for displaced workers
with that of a control group of similar workers who did not separate in
the displacement year or the next 2 years, von Wachter and others (2011)
find that displacements in the early 1980s led to average annual
earnings losses relative to the control group of more than 30 percent of
predisplacement annual earnings. Despite some recovery over time, even
after 20 years the earnings of displaced workers remain 15 to 20 percent
below the level implied by control group earnings.
The short- to medium-run effects of job displacement are larger in
depressed areas and sectors. For example, using information on earnings
and employers from UI records and a comparable definition of job
displacement, Louis Jacobson, Robert Lalonde, and Daniel Sullivan (1993)
find that job displacement in Pennsylvania in the early t980s led on
average to near-term earnings losses of more than 50 percent. Five years
after displacement, the losses average 30 percent of predisplacement
earnings, and they do not substantially fade even 10 years later
(Sullivan and von Wachter 2009). Robert Schoeni and Michael Dardia
(2003) and Yolanda Kodrzycki (2007) find similar results for job
displacement in manufacturing industries in the mild recession of the
early 1990s in California and Massachusetts, respectively.
Earnings losses are large and long lasting even in regions and
periods with stronger labor markets. For example, Kenneth Couch and Dana
Placzek (2010) examine job displacement using quarterly earnings data
from UI records in Connecticut in the 1990s. They find that long-tenure
workers suffer losses in earnings up to 5 years after a job
displacement. Similarly, Jacobson and others (1993) show that workers
displaced in Pennsylvania counties with below-average unemployment rates
and above-average employment growth fare significantly better than the
average displaced worker, but still suffer earnings losses. Von Wachter
and others (2011) find substantial earnings losses for job displacements
during the late-1980s expansion, losses that fade only after 15 years.
Other studies (for example, Topel 1990, Ruhm 1991, and Stevens 1997) use
longitudinal survey data to compare earnings of job losers with those of
a control group. These studies typically do not focus on depressed areas
or periods, but they also find large and persistent losses in earnings
and wages.
The findings from administrative data pertain to annual or
quarterly earnings. Hence, the earnings losses potentially arise from
reductions in both employment and wages. However, the earnings loss for
the median worker in the sample is about as large as, and more
persistent than, the mean loss (von Wachter and others 2011, Schoeni and
Dardia 2003). This result and survey-based evidence that most job losers
return to employment (for example, Farber 1999) suggest that the bulk of
earnings losses after job displacement reflects reductions in wage rates
or hours worked.
One natural question about studies based on administrative data is
how the earnings loss results depend on the definition of job
displacement, the choice of control groups, and the specification of
mass-layoff events. Von Wachter and others (2011) find that their
results survive the use of alternative firm size thresholds, different
definitions of mass layoffs, alternative employment stability
requirements for control groups, and other robustness checks. Von
Wachter, Elizabeth Handwerker, and Andrew Hildreth (2008) obtain similar
results using control groups constructed from workers in similar firms
and industries. Studies based on panel survey data that do not impose
restrictions on firm size or firm events yield results for earnings
similar to results based on administrative data (for example, Topel
1990, Ruhm 1991, Stevens 1997).
Overall, a central finding in previous research is that job
displacement leads to large and long-lasting earnings losses, especially
under weak labor market conditions. This observation suggests that
workers who have experienced job displacement events since 2008 are
likely to suffer unusually severe and persistent earnings losses. Direct
evidence on the losses of recently displaced workers is limited,
however, in part because of lags in processing and analyzing
administrative data sources. The latest Displaced Worker Supplement
(DWS) to the CPS, conducted in January 2010, contains recall data for
workers displaced during 2007-09. Given the absence of a control group,
the inability to incorporate earnings losses due to employment
reductions, and the presence of measurement error in wages and job loss
events, the DWS data tend to show smaller earnings losses than studies
based on administrative data (von Wachter and others 2008). However,
even the DWS data imply substantial earnings losses for persons who lost
jobs during 2007-09. On the basis of the DWS data, the Bureau of Labor
Statistics (2010) reports that only 49 percent of workers with 3 or more
years of job tenure who were displaced during 2007-09 were employed as
of January 2010, and that among the reemployed, 36 percent reported
current earnings at least 20 percent lower than on the previous job.
The earnings losses associated with job displacement are large and
persistent for both women and men and in all major industries. Older
workers tend to have larger immediate losses than younger workers.
Relative to a control group of nondisplaced workers of similar age,
however, the losses of younger displaced workers are nonnegligible and
persist over 20 years (von Wachter and others 2011). Earnings losses
tend to rise with tenure on the job, industry, or occupation (for
example, Kletzer 1989, Neal 1995, Poletaev and Robinson 2008). Yet
losses for workers with 3 to 5 years of job tenure are substantial and
long lasting, and even workers with less than 3 years of job tenure
experience nonnegligible declines in annual earnings following a job
displacement event (von Wachter and others 2011).
II.B. Estimated Earnings Losses Associated with Job Displacement
We now follow von Wachter and others (2011) in estimating the
earnings effects of job displacement and their sensitivity to economic
conditions at the time of displacement. We define job displacement as in
section I as the separation of long-tenure men, 50 years or younger, in
mass-layoff events at firms with at least 50 employees at baseline. We
also provide some results for women and for older men. To estimate the
effects of job displacement, we compare the earnings path of workers who
experience job displacement with the earnings path of similar workers
who did not separate during the same time period, while controlling for
individual fixed effects and differential earnings trends.
We implement this comparison by estimating the following
distributed-lag model separately for each displacement year y from 1980
onward:
(1) [e.sup.y.sub.it] = [[alpha].sup.y.sub.i] +
[[gamma].sup.y.sub.t] + [[bar.e].sup.y.sub.i][[lambda].sup.y.sub.t] +
[[beta].sup.y][X.sub.it] + [20.summation over (k = -6)]
[[delta].sup.y.sub.k][D.sup.k.sub.it] + [u.sup.y.sub.it],
where the outcome variable [e.sup.y.sub.it] is real annual earnings
of individual i in year t in 2000 dollars (deflated using the consumer
price index), [[alpha].sup.y.sub.i] are coefficients on worker fixed
effects, [[gamma].sup.y.sub.t] are coefficients on calendar-year fixed
effects, [X.sub.it] is a quartic polynomial in the age of worker i at
year t, and the error [u.sup.y.sub.it] represents random factors. To
allow further differences in annual earnings increments by a
worker's initial level of earnings, the specification includes
differential year effects that vary proportionally to the worker's
predisplacement average earnings, [[bar.e].sup.y.sub.i], calculated
using the years y - 5 to y - 1. The [D.sup.k.sub.it], are dummy
variables equal to 1 in the worker's kth year before or after his
displacement, and zero otherwise, where k = 1 denotes the displacement
year and k = 0 denotes the final year of earnings with the
predisplacement employer. In the 1985 displacement-year regression, for
example, [D.sup.5.sub.it] = 1 for t = 1989 and zero otherwise for a
worker i who experiences displacement in 1985 by our criteria.
We estimate equation 1 by displacement year using annual,
individual-level observations in the SSA data from 1974 to 2008. To
construct our regression sample for displacement year y, we start with a
1 percent sample of men with a valid Social Security number in y. We
then keep those that had positive Social Security earnings in y and
impose the same restrictions with respect to firm size, industry, worker
age, and job tenure as in figure 2. We then select data on workers
displaced in y, y + 1, and y + 2 plus data on workers in a control group
described below. (12) For the control group workers in a given
displacement-year sample, we set [D.sup.k.sub.it] = 0 for all t.
Although we consider displacement events through age 50, we use earnings
data through age 55. We follow the same approach for women in all
respects but analyze their earnings outcomes separately.
The earnings data for the control group help identify the year
effects [[gamma].sup.y.sub.i] and [[lambda].sup.y.sub.t]. Given the
presence of the year effects and worker fixed effects in equation 1, the
coefficients [[delta].sup.y.sub.k] on the dummies [D.sup.k.sub.it]
measure the time path of earnings changes for job separators from 6
years before and up to 20 years after a displacement, relative to the
baseline and relative to the change in earnings of the control group.
(13) The baseline consists of years 7 and 8 before displacement. (14) To
interpret the estimated [[delta].sup.y.sub.k] coefficients as the
earnings effect of job displacement requires that, conditional on worker
fixed effects and the other control variables, the control group
earnings capture the counterfactual earnings of displaced workers in the
absence of job displacement. Mechanically, to obtain the counterfactual
earnings path of a displaced worker i absent displacement, we evaluate
equation 1 at [D.sup.k.sub.it] = 0 for all k.
For the displacement-year y regression sample, the control group
consists of workers not separating in y, y + 1, and y + 2
("nonseparators"). Hence, as is typical in the literature on
job displacement based on administrative data, we exclude so-called
non-mass-layoff separators from y to y + 2 from the control group.
Non-mass-layoff separators are workers who quit their jobs or were laid
off by firms with an employment drop of less than 30 percent. We impose
the same restrictions with respect to firm size, industry, worker age,
job tenure, and sex as for displaced workers. We discuss the impact of
alternative control groups and concerns related to potential selection
bias in the earnings loss estimates in section II.D.
Figure 4 reports results for men 50 or younger with at least 3
years of job tenure as of the displacement year. The top panel shows the
average time paths of mean raw earnings before and after displacement
for workers displaced in recessions and expansions. If a peak or a
trough falls within a given calendar year, we weight the year according
to the number of its months in expansion or recession when computing the
averages. The middle panel shows the average earnings loss profiles for
workers displaced in recessions and in expansions, relative to the
control group, and normalized to reflect changes relative to mean
earnings in years t - 4 to t - 1 before displacement. To obtain average
earnings losses for job displacements in expansions and recessions, we
average over estimated values of [[delta].sup.y.sub.k] in recession and
expansion years, respectively. The bottom panel shows these losses as a
fraction of predisplacement mean earnings.
The bottom panel of figure 4 shows that the earnings losses of
displaced workers relative to the control group are very large
initially: 39 percent of predisplacement earnings in the first year for
displacements that occur in recessions and 25 percent for displacements
that occur in expansions. They are also long lasting, ranging from 15 to
20 percent from 10 to 20 years out for displacements that occur in
recessions and about 10 percent for those that occur in expansions.
These estimates are robust to many alternative specifications, as
discussed below and in von Wachter and others (2011). For example, the
earnings losses are similar if one defines a mass-layoff event as a
firm-level employment decline of at least 80 percent rather than 30
percent. They are slightly larger for workers with 6 years or more of
job tenure, the main comparison group of Jacobson and others (1993), and
slightly smaller for workers with 3 to 5 years of job tenure.
Figure 5 plots estimated short-term earnings losses against the
national unemployment rate in the year of displacement. We define the
short-term earnings loss as the loss in year t + 2 for a job
displacement in t, as estimated from equation 1, divided by
predisplacement mean earnings in years t - 4 to t - 1. The figure
displays a clear inverse relationship. Regressing the earnings loss on
the unemployment rate at displacement yields an [R.sup.2] of 0.22 and a
slope coefficient of -0.022 (with a standard error of 0.008). That is, a
rise in the unemployment rate from 5 percent to 9 percent at the time of
displacement implies that the earnings loss in the third year of
displacement increases from 18 percent to 26 percent of average annual
predisplacement earnings. Since the earnings recovery pattern in the
bottom panel of figure 4 is approximately parallel in expansions and
recessions, figure 5 suggests that the state of the labor market at
displacement sets the initial level of losses, from which a gradual
recovery ensues. We will use this result when calculating present-value
earnings losses in the next subsection.
II.C. Present-Value Earnings Losses Associated with Job
Displacement
Figures 4 and 5 point to large short-term and long-term earnings
losses associated with job displacement and large earnings loss
differences between displacements that occur in expansions and those
that occur in recessions. To estimate the present discounted value (PDV)
of the annual earnings losses summarized in figure 4, we proceed as
follows. Using a real interest rate of 5 percent, we sum the discounted
losses over a 20-year period starting with the year of displacement.
Since we do not observe the full 20 years of earnings after a job
displacement for workers displaced in later years, we impose a common
rate of decay past the 10th year. Hence, the estimated mean PDV earnings
losses for displacements that occur in, say, a recession are
(2) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]
where [[bar.[delta]].sup.R.sub.s] is the average estimated earnings
loss in year s after displacement (derived by averaging equation 1
estimates over displacement-year regressions), and
[[bar.[delta]].sup.R.sub.10] [(1 - [bar.[lambda]]).sup.s-10] is an
extrapolated earnings loss using the common decay rate [bar.[lambda]].
The evolution of earnings losses is roughly parallel for displacements
in expansions and recessions, so we use the average decay rate of
earnings losses from years 11 to 20 after displacement, estimated using
data for all available workers and years. (15)
[FIGURE 4 OMITTED]
[FIGURE 5 OMITTED]
Other approaches are possible. Rather than a common decay rate, we
could use estimated earnings losses for the largest available sample of
years and workers for each value of s up to s = 20. That approach,
however, involves a different mix of years for each value of s, and for
large values of s the sample would be dominated by displacement events
in the 1980s. Moreover, as the sample of workers displaced in a given
year ages and their labor force participation declines, the estimates
for long after the displacement year may be affected by changes in
composition and greater sampling error in the increasingly smaller
samples. Similarly, using actual estimates for the long-run follow-up
period may put weight on cohorts that experience particularly
long-lasting effects. Given our aim to approximate the average PDV loss
for a typical worker in boom years and in recession years, we choose a
common decay rate for all displacement cohorts. To smooth out sampling
variability in the recovery pattern and to maximize the number of
available cohorts, we calculate the decay rate as the average of
annualized log differences in earnings losses from years 6 to 10 to
years 11 to 15 after displacement. This approach balances the influence
of displacements in the early 1990s, which reflect a strong recovery in
the high-pressure labor market of the mid- to late 1990s, with the
influence of displacements in other periods.
Since earnings levels change over time and may differ between
displacements that occur in expansions and those that occur in
recessions, we consider two ways of normalizing the absolute earnings
losses. First, we scale the PDV earnings loss by displaced workers'
mean annual earnings in years t - 4 through t - 1 before displacement.
This approach expresses the loss as the number of earnings years lost at
the previous level of earnings. Second, we express the PDV earnings loss
as a percentage of PDV earnings along a counterfactual earnings path in
the absence of displacement. To do so, we first construct the
counterfactual by adding the absolute value of the estimated earnings
loss (middle panel of figure 4) back to the actual level of average
earnings (top panel of figure 4). In the notation of equation I, for
workers displaced in year y, we thereby effectively obtain [MATHEMATICAL
EXPRESSION NOT REPRODUCIBLE IN ASCII]. Using the mean earnings of
displaced workers as a benchmark ensures that we average over the fight
worker fixed effects and obtain the right earnings levels. We then take
the average of the counterfactual in years belonging to recessions and
the average in years belonging to expansions. (16) Using these averages,
we divide the PDV earnings loss by the resulting PDV of counterfactual
earnings in booms and recession, respectively.
Table 1 reports these alternative measures of the PDV earnings loss
after a job displacement, again for men 50 years or younger with at
least 3 years of positive earnings at an employer with at least 50
workers. The definition of displacement is the same as in figure 4. The
first row shows estimated PDV earnings losses, averaged over all
displacement years, of $77,557 in dollars of 2000. This amounts to 1.71
years of average predisplacement earnings and 11.9 percent of the PDV of
counterfactual earnings. The next two rows show our measures of PDV
earnings losses separately for expansions and recessions. As anticipated
from figure 4, job displacements lead to very large declines in PDV
earnings, and the losses are much larger for displacements occurring in
recessions. The average worker displaced in a recession experiences PDV
losses of $109,567, equivalent to 2.50 years of average predisplacement
earnings, and an 18.6 percent loss relative to counterfactual earnings.
In contrast, the PDV earnings loss experienced by workers displaced in
an expansion averages $72,487, which amounts to 1.59 years of
predisplacement earnings and an 11.0 percent shortfall relative to the
counterfactual.
Recall from figure 1 that the incidence of job displacement is also
much greater in recessions. Given that displacements have more severe
consequences in recessions, the unweighted averages over years in the
first row of table 1 effectively give less weight to persons displaced
in recessions, and thus understate average PDV earnings losses taken
over all displaced workers. Similarly, because we weight all recession
years equally, and recessions with higher displacement rates also
involve higher earnings losses, table 1 understates the average PDV
earnings losses for job displacements that occur in recessions.
The last five rows of table 1 show how estimated PDV earnings
losses vary with the unemployment rate in the year of displacement. The
unemployment rate reflects contemporaneous labor market conditions in a
different way than business cycle dating. As before, to calculate the
table entries, we first estimate PDV earnings losses by year of
displacement. We then average over all years falling into an indicated
unemployment range, assigning fractional weights to years that fall
partly into a given range. The results show that PDV earnings losses
rise steeply with the unemployment rate in the year of job displacement.
This important finding strongly reinforces and extends the evidence in
figure 5.
To take this result one step further, we repeat our procedure for
calculating PDV earnings losses by year of displacement. We now depart
from working with averages over multiple displacement years and consider
a separate earnings loss path for each displacement year. When we have
more than 10 years of postdisplacement information, we use the first 10
years and extrapolate from year 11 to year 20 using the same average
rate of decay as before. When we have less than 10 years of
postdisplacement information (that is, starting in 1999), we also use
the available information for other years to construct decay rates in
the earlier postdisplacement years. For displacement years with less
than 10 but more than 5 years of postdisplacement data, we set the decay
rate to the annualized log difference of losses between the 6th and the
10th year after displacement, taken from displacement years for which
this information is available. For those years with less than 6
displacement years, we use the annualized log differences of losses
between the 2nd and the 5th displacement year. For years closer to the
end of our sample period, we necessarily rely more heavily on
extrapolation.
Figure 6 plots the resulting PDV earnings losses (expressed as
multiples of average annual predisplacement earnings) against the
unemployment rate in the year of displacement. The figure again shows an
approximately linear relationship, which is not surprising given the
roughly linear relationship in figure 5 and our use of a common decay
rate beyond the 10th year after displacement. Even allowing for
different postdisplacement recovery patterns, the figure suggests that
PDV earnings losses increase approximately linearly with the
unemployment rate in the year of displacement. A linear regression of
the PDV loss measure on the unemployment rate at displacement yields an
[R.sup.2] of 0.27 with a slope coefficient of --0.23 (standard error of
0.08). Thus, an increase in the unemployment rate at displacement from 5
percent to 9 percent implies that PDV earnings losses rise from 1.6
years to 2.5 years of predisplacement earnings. When we add an indicator
for recession years to this descriptive regression model, it is not
statistically significant.
Table 2 shows PDV earnings losses for displaced women and for
various age and tenure subgroups of displaced men. (17) The PDV earnings
losses due to job displacement are large for all these groups. They are
smaller for women than for men, but not dramatically so in the last two
columns, which effectively control for differences in average earnings
levels between men and women. For example, the average losses for women
amount to 1.5 years of predisplacement earnings (table 2), compared with
1.7 years for the corresponding group of men (table 1). Comparison of
tables 1 and 2 also shows that the losses are larger for men with longer
job tenure before displacement. The panels reporting results for male
age subgroups show that, except for men displaced near the end of their
working lives, PDV earnings losses are much larger for displacements
that occur in recessions.
[FIGURE 6 OMITTED]
II.D. On Selection Bias and Sensitivity to Control Group Choice
We now discuss two potential concerns about the earnings loss
estimates that underlie our results in figures 4 to 6 and tables 1 and
2, namely, selection bias and the sensitivity of our results to the
choice of control group. Relative to nonseparators (our control group),
non-mass-layoff separators experience earnings losses that are smaller
and less persistent than the losses experienced by mass-layoff
separators. Thus, if we include non-mass-layoff separators in the
control group, the estimated earnings losses due to job displacement
become smaller. Von Wachter and others (2011) estimate a version of
equation 1 with non-mass-layoff separators as part of the control group.
This change in the composition of the control group reduces the
estimated earnings losses by about one-quarter. Von Wachter and others
also consider instrumental variables estimates that are not affected by
the presence of voluntary separators, which we discuss below, and obtain
results very similar to those reported here. After considering various
estimators, they confirm the conclusion from previous research that the
"true" loss at displacement is closer to the estimates that
exclude non-mass-layoff separators from the control group.
Estimates based on equation 1 may overstate earnings losses at
displacement because displaced workers are negatively selected on
observable and unobservable characteristics with respect to the control
group: employers may lay off workers who are less productive and have
less future earning potential. Von Wachter and others (2011) conduct an
in-depth investigation of this question and conclude that earnings
losses based on equation 1 are robust to a range of important
sensitivity checks. The presence of worker fixed effects in equation 1
implies that selection based on fixed worker attributes with a
time-invariant effect on earnings poses no problem. However, different
trends in counterfactual earnings between displaced workers and the
control group may introduce a bias. For example, it is well known that
different parts of the earnings distribution experience different
earnings growth rates (see, for example, Autor and Katz 1999). Since
displaced workers have lower average earnings before displacement than
nondisplaced workers, our regression models include interactions between
average earnings in the 5 years before displacement and fixed effects
for calendar years. Von Wachter and others also present estimates that
include differential trends by two-digit industry and by other
observable characteristics of workers and firms before displacement. The
estimates are reasonably robust to these modifications and decline only
somewhat with the inclusion of industry-specific trends.
However, ex ante differences in unobservable characteristics
between treatment and control groups can still lead to different
counterfactual earnings trends. In this context, von Wachter and others
(2011) address two types of selection: that within and that between
employers. To address the concern that displaced workers are negatively
selected on potential unobserved earnings trends within firms, they
replicate equation 1 using the mass-layoff event at the firm level as an
instrumental variable for displacement. That is, they use a dummy for
the year of the mass layoff at the firm, [D.sup.k.sub.f(i)t], where f(i)
is the worker's employer, to instrument for the dummy of the
individual layoff ([D.sup.k.sub.it]). Hence, the comparison is now
between the earnings of all workers at firms undergoing mass layoffs and
the earnings of all workers at non-mass-layoff firms. Using this type of
firm-level indicator to instrument for displacement, and controlling for
differential trends by pre-mass-layoff characteristics at the firm
level, von Wachter and others obtain results very similar to those
reported here based on equation 1. This instrumental variables estimator
is also robust to the presence of non-mass-layoff separators, since the
instrument should be orthogonal to the rate of retirement or voluntary
mobility.
To address the possible concern that workers with lower potential
earnings trends sort into firms more likely to experience mass layoffs,
von Wachter and others (2011) follow previous work and consider a
version of equation 1 that includes an interaction between year effects
and firm fixed effects. This specification yields somewhat smaller
estimated earnings losses, because the losses of workers remaining at
firms with mass layoffs are now subtracted from the losses of the
displaced workers. It is not clear whether the decline in earnings for
those remaining at mass-layoff firms should be subtracted or treated as
part of the outcome. In any event, the estimated losses for the
displaced workers remain substantial and very persistent. Von Wachter
and others conclude that estimates based on equation 1, on which we
rely, are robust to a range of important sensitivity checks. Hence,
despite some variation depending on the exact specification, we believe
our calculations based on estimated versions of equation 1 provide a
reasonable characterization of the magnitude and persistence of the
individual earnings losses caused by job displacement.
III. Other Costs of Job Displacement and Unemployment
Section II focused on earnings losses associated with displacement
events. We turn now to the effects of job displacement on other outcomes
such as consumption, health, mortality, and children's educational
achievement. We also present new evidence on cyclical movements in
worker anxieties and perceptions about the risk of job loss and the ease
or difficulty of job finding.
III.A. Effects on Income, Consumption, and Employment Stability
It is not easy to estimate the effects of job displacement on
consumption and income. Few, if any, data sets that track large numbers
of workers over time contain high-quality information about consumption
outcomes. Likewise, very few data sets that track large numbers of
workers include the data on earnings, asset incomes, and public and
private transfer payments needed to identify income responses to job
displacement events. Moreover, transfer payments are understated greatly
in many household surveys that include such information (Meyer, Mok, and
Sullivan 2010).
The few studies that estimate the effects of job loss or
unemployment on consumption typically find sizable near-term declines in
consumption expenditure but lack evidence on long-term consumption
responses. See Gruber (1997) and Stephens (2004), for example. The
consumption responses tend to be concentrated at the lower end of the
income distribution (Browning and Crossley 2001, Congressional Budget
Office 2004). Although transfer programs often mitigate the earnings
loss due to job displacement, the replacement amounts are quite modest
compared with our estimates of present-value earnings losses. Even the
generous, long-lasting benefits available under the German unemployment
insurance system replace only a modest share of the earnings loss
associated with job displacement (Schmieder, von Wachter, and Bender
2009).
Previous research also finds that job displacement leads to other
adverse consequences. Lasting postdisplacement earnings shortfalls occur
alongside lower job stability, greater earnings instability, recurring
spells of joblessness, and multiple switches of industry or occupation
(Stevens 1997, von Wachter and others 2011). Much of the increased
mobility between jobs, between industries, and between occupations
probably reflects privately and socially beneficial adjustments. On
average, however, displaced workers who immediately find a stable job in
their predisplacement industry obtain significantly higher earnings.
Lower job stability and higher earnings volatility persist up to 10
years after displacement. Thus, there is no indication that laid-off
workers trade a lower earnings level for a more stable path of
employment and earnings.
III.B. Effects on Health, Mortality, Emotional Well-Being, and
Family
There is also evidence that displaced workers suffer short- and
long-term declines in health. Survey-based research in epidemiology
finds that layoffs and unemployment spells involve a higher incidence of
stress-related health problems such as strokes and heart attacks (see,
for example, Burgard, Brand, and House 2007).
Whereas studies of self-reported health and job loss outcomes face
significant challenges related to measurement error and to recall and
selection bias, the analysis of mortality outcomes can exploit large
administrative data sources that are less subject to these problems.
Sullivan and von Wachter (2009) study the effects of job displacement on
mortality outcomes over the 20 years following displacement, using
administrative data on earnings and employers from the Pennsylvania UI
system and mortality data from the SSA. Their results show that mature
men who lost stable jobs in Pennsylvania during the early 1980s
experienced near-term increases in mortality rates of up to 100 percent.
The initial impact on mortality falls over time, but it remains
significantly higher for job losers than for comparable workers
throughout the 20-year postdisplacement period. If sustained until the
end of life, the higher mortality rates for displaced workers imply a
reduction in life expectancy of 1 to 1.5 years.
Because the 1980s recession was especially deep in Pennsylvania and
involved unusually large earnings losses for displaced workers, the
mortality effects estimated by Sullivan and von Wachter (2009) reflect a
very bad case scenario. It is reasonable to expect smaller mortality
effects of job displacements in most other years and places.
Unfortunately, labor market conditions nationwide in the past 3 years
have also been dismal, with persistently high unemployment rates. Thus,
the mortality estimates in Sullivan and von Wachter may well provide a
suitable guide to mortality effects for recently displaced American
workers. The available evidence indicates that job displacement also
raises mortality rates in countries with universal public health
insurance systems and generous social welfare systems, such as Sweden
(Eliason and Storrie 2009) and Norway (Rege, Telle, and Votruba 2009).
These studies find higher mortality rates in the years following job
displacement, but they contain little information about long-term
effects.
Several studies point to short- and long-term effects of layoffs on
the children and families of job losers and unemployed workers. In the
short run, parental job loss reduces the schooling achievement of
children (Stevens and Schaller 2011). In the long run, it appears that a
lasting reduction in the earnings of fathers reduces the earnings
prospects of their sons (Oreopoulos, Page, and Stevens 2008). Patrick
Wightman (2009) also finds that parental job loss is harmful for the
educational attainment and cognitive development of children. Other
studies find that layoffs raise the incidence of divorce, reduce
fertility, reduce home ownership, and increase the rate of application
to and entry into disability insurance programs (Charles and Stephens
2004, von Wachter and Handwerker 2009, Rupp and Stapleton 1995). Last
but not least, and perhaps not surprisingly given the magnitude and
range of adverse consequences discussed above, job loss and unemployment
also lead to a reduction in happiness and life satisfaction (see Frey
and Stutzer 2002).
Clearly, care should be taken in drawing welfare conclusions and
policy prescriptions from the range of adverse consequences associated
with job displacement. However, this brief review makes clear that job
displacement entails a variety of significant short- and long-run costs
for affected workers and their families. Neither the large present-value
earnings losses we estimate nor the estimated consumption responses
capture the full measure of costs associated with job displacement.
III.C Cyclical Movements in Worker Anxieties and Perceptions
Given the severity of job displacement effects on earnings and
other outcome measures, it is natural to ask how worker anxieties and
perceptions about labor market conditions track actual conditions.
Evidence on this issue is potentially informative in several respects.
First, if recessions or high unemployment rates cause employed workers
to become more fearful about layoffs and wage cuts, they involve
psychological costs beyond the direct effects on job-losing workers and
their families. Second, perceptions about labor market conditions are
likely to influence search behavior by employed and unemployed workers,
including those who experience a displacement event. Third, high worker
anxiety about labor market conditions is likely to undermine consumer
confidence and depress consumption expenditure. (18) Fourth, perceptions
about labor market conditions have important influences on policymaking,
politics, and electoral outcomes. Because they potentially influence so
many voters, anxieties about labor market conditions may have more
important political consequences than actual conditions.
A long-running source of data on perceptions about labor market
conditions is the General Social Survey (GSS), a repeated
cross-sectional household survey conducted since 1972. The GSS includes
two categorical response questions that are useful for gauging cyclical
movements in perceptions about labor market conditions. One question
asks the respondent about the perceived likelihood that he or she will
lose a job or be laid off in the next 12 months. The other asks about
the perceived difficulty of finding a job with the same income and
fringe benefits as the respondent's current job.
The top panel of figure 7 shows, for each available year in the
GSS, the percentage of prime-age workers who consider it "very
likely" or "fairly likely" that they will lose a job or
be laid off in the next 12 months. The figure plots these values against
the average CPS unemployment rate in the 5-month window that brackets
the corresponding GSS interview months. There is a strong, positive
relationship: an increase in the prime-age unemployment rate from 4
percent to 8 percent raises from 10 percent to 15 percent the share of
prime-age workers who perceive job loss as fairly or very likely. The
online appendix shows a very similar pattern for all employed workers 18
to 64 years of age.
The bottom panel of figure 7 shows the percent of prime-age workers
who perceive it to be "not easy" to find a job with income and
fringe benefits similar to those in their current job. Plotting these
values against contemporaneous unemployment rates, we again find a
strong relationship: an increase in the prime-age unemployment rate from
4 percent to 8 percent raises from 35 percent to 52 percent the share of
prime-age workers who regard it as hard to find another job with a
comparable compensation package. In this context it is also worth noting
that quit rates are highly procyclical (see, for example, Davis and
others 2012). Quit rates plummeted in the most recent recession and
remain extraordinarily low, another indication that workers perceive
good jobs as hard to find.
Gallup polls provide another long-running, consistent source of
data on perceived labor market conditions. The Gallup data cover a
shorter time period than the GSS data, but they pertain to a highly
eventful period in terms of economic developments. In addition, one of
the Gallup measures is available at a (roughly) monthly frequency, which
is useful for assessing the shorter-term relationship between perceived
and actual conditions. Figure 8 draws on the Gallup data to plot over
time the percent of adult interviewees who respond yes to the following
question: "Thinking about the job situation in America today, would
you say that it is now a good time or a bad time to find a quality
job?" The responses are highly cyclically sensitive. As the labor
market tightened, the share of yes responses rose from about 20 percent
in early 2003 to nearly 50 percent in the first half of 2007. It then
dropped to about 10 percent over the next 2 years and has remained at
very low levels ever since. This evidence suggests that perceptions
about labor market conditions respond rapidly to actual conditions.
[FIGURE 7 OMITTED]
[FIGURE 8 OMITTED]
Table 3 reports data from Gallup polls conducted during the month
of August in 1997 and every year from 2003 to 2011. The table shows a
tremendous increase in worker anxiety levels following the peak of the
financial crisis in the latter part of 2008 and early 2009. The
percentages of employed adults expressing worries that they personally
would experience a cutback in hours, a wage cut, a benefit cut, or a
layoff in the near future jumped dramatically. After some lessening
between August 2009 and August 2010, the most recent data for August
2011 show worker anxiety returning to peak or near-peak levels.
In summary, the evidence presented in figures 7 and 8 and table 3
indicates that worker perceptions about labor market conditions are
closely attuned to actual conditions. The Gallup polling data, in
particular, point to a dramatic deterioration in perceptions about labor
market conditions and prospects after the financial crisis--a
deterioration that persists to the present day and that involves
widespread concerns about layoff risks, wage and benefit cuts, shorter
hours, and the difficulty of finding a good job. Whether or not these
fears show up in realized earnings outcomes, they involve psychological
costs in the form of heightened anxiety for much of the population.
IV. The Effects of Job Loss in Leading Theoretical Models of
Unemployment and Labor Market Dynamics
Mortensen and Pissarides (1994) present an equilibrium
search-and-matching model that, in various formulations, has become the
leading framework for analyzing aggregate unemployment fluctuations. We
now evaluate how well certain versions of the Mortensen-Pissarides (MP)
model account for our evidence on the magnitude and cyclicality of the
earnings losses associated with job displacement. (19)
IV.A. MP Models of Unemployment Fluctuations
Shimer (2005) considers a basic version of the MP model with
risk-neutral workers and firms, uniform match quality, Nash bargaining,
and a constant rate of job destruction and job loss. Aggregate shocks
drive employer decisions about vacancy posting and fluctuations in job
creation, job finding, and unemployment. Shimer shows that the basic MP
model delivers too little volatility in unemployment for reasonable
specifications of the aggregate shock process (see also Costain and
Reiter 2008). Under Nash bargaining, the equilibrium wage largely
absorbs shocks to labor productivity in the basic model. As a result,
realistic shocks have little impact on employer incentives to post
vacancies, and the model generates small equilibrium responses in job
finding rates, hiring, and unemployment. This unemployment volatility
puzzle has motivated a great deal of research in recent years.
One prominent strand of this research stresses the consequences of
wage rigidities. (20) Hall and Milgrom (2008), for example, step away
from Nash bargaining while retaining privately efficient compensation
and separation outcomes. They replace Nash bargaining with the
alternating-offer bargaining protocol proposed by Ken Binmore, Ariel
Rubinstein, and Asher Wolinsky (1986). Whereas the standard Nash wage
bargain treats termination of the match opportunity as the threat point,
the threat point in Hall and Milgrom's "credible
bargaining" setup is a short delay followed, with high probability,
by a resumption of bargaining. This change in bargaining regime goes a
long way to insulate the equilibrium wage bargain from aggregate shocks
and outside labor market conditions.
A key point is that the cost of a small delay during the bargaining
process is less cyclical than the value of outside opportunities. Hence,
closing the basic MP model in the manner of Hall and Milgrom leads to
greater sensitivity of the employer surplus value to aggregate shocks
and bigger responses in vacancies, job finding rates, and unemployment.
Hall and Milgrom show that their specification of the bargaining
environment resolves the unemployment volatility puzzle in a reasonably
calibrated version of the basic MP model.
In our analysis below, we adopt Hall and Milgrom's credible
bargaining version of the basic MP model and two versions with Nash
bargaining. We follow this approach for two reasons. First, Hall and
Milgrom offer perhaps the most successful version of the basic MP model
in terms of explaining the cyclical behavior of job finding rates,
vacancies, and unemployment. Second, by comparing the credible
bargaining and Nash versions of the model, we can determine whether a
particular form of wage rigidity improves the model's ability to
account for the facts about earnings losses associated with job loss.
Despite much attention to the basic MP model in recent work, the
model misses some first-order features of labor market fluctuations. The
basic MP model cannot reproduce the recessionary spikes in job
destruction, job loss, and unemployment inflows depicted in figures 1
and 2. Moreover, the model has no role for hires and separations apart
from job flows. There is no search by employed workers, no job-to-job
movement, and no replacement hiring. As a related point, the basic model
entails no heterogeneity of productivity, match surplus values, or
wages. This sort of heterogeneity seems important for generating large
earnings losses due to job loss. Given these limitations, we also
consider a model of Burgess and Turon (2010) that extends Mortensen and
Pissarides (1994) by incorporating search on the job and other changes.
Burgess and Turon's model produces hires and separations apart from
job flows and recessionary spikes in job destruction, job loss, and
unemployment inflows.
There are also good reasons to anticipate that the model of Burgess
and Turon will generate larger earnings losses associated with job loss
than the basic MP model. Like models by Kenneth Burdett and Mortensen
(1998) and by Fabien Postel-Vinay and Jean-Marc Robin (2002) and other
models that include search on the job, their model generates persistent
heterogeneity in match surplus values and wages for workers of a given
quality. It also delivers a job "ladder" whereby newly
reemployed workers tend to obtain jobs on the lower rungs of the wage
distribution initially and to move up the wage distribution over time
through search on the job. This job ladder feature prolongs the period
of earnings recovery after displacement. Finally, Andreas Hornstein, Per
Krusell, and Giovanni Violante (2010) show that plausibly parametrized
versions of basic search models yield very modest levels of frictional
wage dispersion, which implies little scope for earnings losses due to
job loss when unemployment spells are short. Hornstein and others also
consider several extensions to basic search models, and among those they
consider, the only ones that offer much scope for cross-sectional wage
dispersion are models with search on the job.
IV.B. Income and Earnings Losses in the Basic MP Model
Table 4 reports statistics for three versions of the basic MP
model: the credible bargaining version of Hall and Milgrom (2008) and
two versions with Nash bargaining--a standard calibration similar to
that of Shimer (2005) and another calibration similar to that of Marcus
Hagedorn and Iourii Manovskii (2008). These two calibrations differ
chiefly in the level of income imputed to the unemployed, which we
interpret as the sum of UI benefits, the value of additional leisure and
home production activity, and any savings on work-related costs.
Hagedorn and Manovskii set this value to a level nearly as large as the
productivity of the employed, thereby amplifying the equilibrium
response of unemployment to aggregate shocks. The standard calibration
involves a much larger gap between productivity and the imputed income
value of unemployment, yielding much smaller equilibrium responses to
shocks of a given size. Our calibrations follow Hall and Milgrom (2008)
in their choice of parameter values for each version of the basic MP
model. See the online appendix for a detailed discussion of the model
simulations and our calculations for the present-value losses associated
with job loss.
The first row of table 4 highlights an important message: job loss
and unemployment are a rather inconsequential event for persons living
in the basic MP world. With a 5 percent annual discount rate, job loss
reduces the present value of income by about 0.2 percent in the MP-CB
and standard MP-Nash versions of the model and by less than 0.05 percent
in the Hagedorn-Manovskii calibration. We compute these present-value
income losses directly from value functions. That is, for each of five
aggregate states we calculate the difference between the asset value of
employment and the asset value of unemployment, expressing the
difference relative to the former. Performing this calculation for all
five aggregate states yields the reported ranges. If these results
capture the real-world costs of job loss, one might well wonder why all
the fuss--why are job loss and unemployment perceived as important
economic phenomena and potent political issues?
The rest of the table reports statistics on unemployment, job
finding, and the distribution of present-value income and earnings
losses for the different models. To compute these statistics, we
simulate aggregate and individual paths. Starting in the middle
aggregate state, we simulate 1,000 aggregate paths for each version of
the model, letting each simulation run for 20 years (5,000 days at 250
working days per year). Along each aggregate path, we simulate paths for
large numbers of workers who either lose jobs or remain employed on day
1. Flow income equals the annuity value of the wage bargain when
employed and the imputed flow value of unemployment otherwise. The PDV
of income includes the discounted asset value of the individual's
realized terminal state. To compute the realized income loss for a
day--1 job loser, we compare the PDV of that individual's realized
income path with the mean realized PDV of income for persons who remain
employed on day 1 on the same aggregate path. In this way, by comparing
day--1 job losers with persons who remain employed along the same
aggregate path, we obtain a comparison between the treatment group
(day--1 job losers) and the controls (day--1 employed).
To compute the realized earnings loss for a day--1 job loser, we
compare the PDV of that individual's realized earnings path over
the 20-year horizon with the mean PDV of realized earnings for
individuals living on the same aggregate path who remain employed on day
1. Earnings equal the wage when employed and zero when unemployed. We
set the terminal value to zero to match the 20-year horizon in our
empirical estimates of PDV earnings losses. Thus, the earnings losses in
table 4 are larger than the corresponding income losses for two reasons:
earnings exclude the imputed income value of unemployment, and we set
terminal values to zero in the earnings comparisons.
Consider the results for the MP-CB model in the first panel of
simulations in table 4. Averaging over all day--1 job losers on all
aggregate paths yields an average realized PDV income loss of 0.23
percent. This figure essentially replicates the income loss result for
the MP-CB model in the top row, as it should. However, the simulation
approach enables us to compute the full distribution of outcomes: the
90th-percentile income loss in the MP-CB version is only 1.04 percent,
still a rather modest value, and job losers at the 10th percentile of
the distribution actually experience a gain of 0.51 percent in PDV
income.
Turning to earnings losses, we report results for the MP-CB version
only, because the other two versions yield very similar results. Mean
PDV earnings losses are 1.28 percent in the basic MP model--an order of
magnitude smaller than the 11.9 percent figure in the last column and
first row of table 1. One potential concern about this earnings loss
comparison is that table 1 considers losses associated with job
displacement events, which by design exclude many job loss events that
involve little or no loss of earnings or income. So there is a sense in
which we have compared average job loss outcomes in the basic MP model
with bad-case outcomes in the data. Although this argument has some
force, we do not find it persuasive. The estimated earnings losses
reported in section Il pertain to an ex ante identifiable group of
workers (men 50 or younger with 3 or more years of job tenure at firms
with 50 or more employees), and this group accounts for a large share of
U.S. employment. We would like to have a theoretical model that explains
the magnitude and cyclicality of the PDV earnings losses associated with
job loss for this large group.
The remaining panels in table 4 consider selected aggregate paths
defined by the mean realized PDV income or earnings losses.
"Boom" paths are those near the 10th percentile of average
losses for day--1 job losers, and "bust" paths are those near
the 90th percentile. Mean PDV income losses remain small along both boom
and bust paths. Even when we isolate the worst 1 percent of individual
outcomes along the bust paths, the PDV income losses amount to only 2.2
percent in the CB and standard Nash versions of the model and only 0.7
percent in the Hagedorn-Manovskii calibration. In short, the basic MP
model cannot produce large welfare losses for job losers, even at the
extremes of aggregate and individual outcomes. The model can produce
large PDV earnings losses at the extremes of the distribution of
individual outcomes. For example, the worst 1 percent of individual
outcomes yield earnings losses comparable to the mean loss reported in
table 1. (21) This result, however, hardly amounts to a success for the
model.
Why are the consequences of job loss so modest in the basic MP
model? Two aspects of the model deliver the result almost immediately.
First, wages are uniform in the cross section, so that unemployment
spells are the only source of earnings loss upon job loss. Second, when
calibrated to job finding rates typical of the postwar U.S. experience,
expected unemployment durations are short, about 2 or 3 months. Short
unemployment spells coupled with uniform wages in the cross section
imply small earnings losses associated with job loss.
The basic MP model also implies a close relationship between the
cost of job loss to the worker and the vacancy supply condition (as has
been stressed to us by Robert Hall). Given free entry, the zero-profit
condition for job-creating employers says that the daily vacancy filling
rate times the asset value of a filled job equals the daily flow cost of
maintaining a vacancy. The JOLTS data imply a vacancy filling rate of
about 5 percent per day. Drawing on work by Jose Silva and Manuel Toledo
(2009) and Hagedorn and Manovskii (2008), Hall and Milgrom (2008)
conclude that the daily flow cost of a vacancy is about one-half of a
worker's daily output. Thus, the employer's asset value of a
newly filled job is equivalent to about 10 days of the output generated
by a (newly hired) worker. If employer and worker share equally in the
surplus generated by a new match, then the worker's value of
transitioning from unemployment to employment is also about 10 days of
output. In other words, not much value is at stake in the creation and
destruction of employment relationships in the basic MP model. Richer
models in the MP class need not imply such a tight relationship between
the cost of filling a new job and the surplus value of the average
existing job.
In summary, we draw three conclusions from table 4 and the related
discussion. First, job loss is a rather inconsequential event for
individual welfare in the basic MP model, even at the extremes of
individual and aggregate outcomes. Second, the basic MP model cannot
rationalize the empirical evidence on PDV earnings losses associated
with job displacement. Third, although wage rigidity of the form
considered by Hall and Milgrom (2008) greatly improves the ability of
the basic MP model to explain aggregate unemployment fluctuations, it
does not bring the model closer to the evidence on the magnitude and
cyclicality of earnings losses associated with job displacement.
IV.C Losses in an MP Model with Job Destruction Spikes and Search
on the Job
Burgess and Turon (2010) depart from Mortensen and Pissarides
(1994) by introducing search on the job, at a cost, and by adopting a
different vacancy creation process that gives meaning to the concept of
a job apart from an employer-worker match. Specifically, they assume a
finite supply elasticity of potential new job creation each period, so
that firms find it optimal to refill certain jobs left open by departing
workers. Like Mortensen and Pissarides (1994), their model also differs
from the basic MP model in capturing cross-sectional heterogeneity in
match products and surplus values. These extensions lead to
cross-sectional wage dispersion, a distinction between job flows and
worker flows, and endogenous job destruction spikes in the wake of
negative aggregate shocks. The model also gives rise to a job ladder
that prolongs the recovery of predisplacement earnings for job-losing
workers.
The model is set in continuous time. Idiosyncratic productivity
shocks arrive according to independent Poisson processes, and aggregate
productivity, p, follows a three-state Markov chain. When hit by an
idiosyncratic shock, a job draws a new idiosyncratic productivity value
in the interval [-[sigma], [sigma]], possibly higher or lower than the
previous value. Optimizing behavior yields three idiosyncratic
productivity thresholds, as shown in figure 9. If idiosyncratic
productivity exceeds S(p) in a filled job, the worker's net
expected gains to search are negative. For productivity less than S(p)
in a filled job, the worker's net expected gains to search are
positive. If the worker finds a vacant job, he quits and the firm
decides whether to search for a replacement. It does so if idiosyncratic
productivity exceeds T(p); otherwise, it lets the job lapse. If a filled
job draws a new idiosyncratic productivity value below R(p), the job is
destroyed and the worker experiences job loss. As the figure indicates,
the productivity thresholds are functions of the aggregate state. A
negative shock to p shifts R(p) to the right, triggering a burst of job
destruction. An important implication is that job losses due to
idiosyncratic shocks occur throughout the distribution of
productivities, whereas job losses due to aggregate shocks occur only
for low-value jobs.
[FIGURE 9 OMITTED]
Table 5 reports PDV income and earnings losses for the model of
Burgess and Turon. We modify their calibration to generate job finding
rates and unemployment spell durations comparable to postwar U.S.
experience. (22) The top panel reports results for a period of time
corresponding to 3 months with no change in the aggregate state. The
remaining two panels involve transitions between states and focus on
outcomes for workers who lose jobs in the early part of a downturn,
roughly corresponding to the recessionary spikes in job destruction and
job loss seen in figures 1 and 2. All loss calculations pertain to
workers who separate from their employer in job destruction events and
exclude separations that result from search on the job.
The first two rows of table 5 report PDV income and earnings losses
for job losers in the good, middle, and bad aggregate states. We compute
the income losses using differences in value functions at each level of
idiosyncratic productivity, e, and then integrate over the distribution
of e that prevails in the indicated aggregate state to obtain the mean
PDV income losses. These losses are larger than in the basic MP model,
but they remain quite modest: about 0.3 to 0.4 percent.
For earnings losses we adopt a simulation approach similar to the
one used for table 5. However, we now compare the realized PDV earnings
of workers who lose jobs characterized by a given [epsilon] with the
mean realized PDV earnings among workers who remain employed (in the
displacement period) at the same value of [epsilon]. Once we obtain the
comparison for each E, we integrate with respect to the appropriate
distribution to obtain the mean realized PDV earnings loss. As before,
we use a 20-year horizon for the earnings calculations. The online
appendix describes the model simulations and PDV calculations in detail.
The remaining panels consider job loss events that occur in the
quarter when the economy gets hit by a negative aggregate shock. Job
loss events now arise for two reasons. As before, a flow of negative
idiosyncratic shocks produces a stream of job loss events. In addition,
the negative aggregate shock erases the surplus value of marginal jobs,
producing a burst of job destruction and job loss. All workers at jobs
below the new, higher destruction threshold R become unemployed in the
wake of a negative aggregate shock. That is, for treatment-control
comparisons conditional on the idiosyncratic productivity value e, all
workers below the new destruction threshold are in the same position.
(Hence, losses are zero in the row in each panel that reports the
"mean loss due to aggregate shock.") (23) For control group
comparisons, job loss produces PDV income losses of about 0.3 percent in
these "recession" periods. The disproportionate loss of
marginal jobs in the wake of a negative aggregate shock pulls down the
average present-value income loss. So the model of Burgess and Turon
does not shed much light on why job loss events in recessions are more
consequential.
With respect to earnings, our calibrated version of the Burgess and
Turon model produces nontrivial PDV losses. For a given aggregate state,
the losses reported in the top panel of table 5 range from 2.4 to 2.7
percent of PDV earnings, about one-quarter of the empirical PDV earnings
losses reported in tables 1 and 2. Thus, search on the job and
heterogeneity in match surplus values clearly help move the model closer
to the evidence on the PDV earnings losses associated with job loss.
In this respect, the job ladder feature of the model plays an
important role. The online appendix displays the cross-sectional wage
function, the density of all filled jobs, and the density of first jobs
for newly reemployed workers who leave unemployment. For our calibrated
version of the model, the maximum wage in the good aggregate state
exceeds the minimum wage by 49 percent. The density of first jobs is
much more concentrated at the low end of the wage distribution than the
density of all jobs. The average difference between the predisplacement
wage and the wage on the first postdisplacement job is 10 percent in the
good aggregate state, 8.4 percent in the middle state, and 6.7 percent
in the bad state. These observations and statistics are different ways
of saying that the model incorporates a significant job ladder.
A few additional remarks are in order. First, in generating the
results in table 5, we do not impose a job tenure requirement on either
displaced workers or control group workers. Doing so may increase the
earnings losses. Second, search intensity is a binary decision variable
in the model of Burgess and Turon. Variable search intensity for
employed workers, as in work by Matthias Hertweck (2010), may generate
an elongated climb up the job ladder after displacement and, as a
result, produce larger PDV earnings losses. (24) We conclude that job
ladder models can produce nontrivial earnings losses due to job
displacement but are unlikely to account for the bulk of observed
losses. For one thing, they do not explain why the earnings of displaced
workers remain well below those of control group workers 10 or more
years after displacement. Moreover, it does not appear that a pure job
ladder model can rationalize the striking cyclical pattern in PDV
earnings losses that we documented in section II.
V. Concluding Remarks
Long-tenure workers who lose jobs in mass-layoff events experience
large and persistent earnings losses compared with otherwise similar
workers who retain their jobs. That is the central message of a
now-sizable literature on the earnings losses associated with job
displacement. We focus on displacements from 1980 to 2005 among men 50
or younger with 3 or more years of prior job tenure. For this group, job
loss in a mass-layoff event reduces the present value of earnings by an
estimated $77,557 (in 2000 dollars) over 20 years at a 5 percent annual
discount rate, equivalent to 1.7 years of predisplacement earnings.
Losses are larger for men with greater job tenure, They are smaller for
women, even as a multiple of predisplacement earnings.
Present-value losses rise steeply with the unemployment rate at the
time of displacement. The average loss equals 1.4 years of
predisplacement earnings if unemployment at displacement is less than 6
percent, and 2.8 years if unemployment exceeds 8 percent. More
generally, the evidence in tables 1 and 2 and figures 4 to 6 says that
tight labor market conditions at displacement strongly improve the
medium- and long-term future earnings prospects of displaced workers.
The highly procyclical behavior of job finding rates among the
unemployed implies that tight labor market conditions strengthen
near-term reemployment and earnings prospects as well. Seen in this
light, economic policies that set the stage for strong growth and low
unemployment are highly beneficial to displaced workers. Indeed,
pro-growth policies may be the most efficient and cost-effective means
available to policymakers to alleviate the hardships experienced by
displaced workers.
Previous work shows that job displacement also has negative
consequences for employment and earnings stability, household
consumption expenditure, health and mortality outcomes, children's
educational achievement, and subjective well-being. We present evidence
that worker perceptions about layoff risks, job finding prospects, and
the likelihood of wage cuts closely track cyclical fluctuations in
actual labor market conditions. Perception measures point to a
tremendous increase in worker anxieties about labor market prospects
after the financial crisis of 2008, an increase that persists through
August 2011. It seems likely that these high anxiety levels produce
important stresses and psychological costs for a large segment of the
population.
We also consider whether models of unemployment fluctuations along
the lines of the canonical contribution by Mortensen and Pissarides
(1994) can account for the earnings losses associated with job
displacement. Basic versions of the MP model featured in much recent
research imply theoretical earnings losses an order of magnitude smaller
than empirical losses. The explanation is straightforward. The basic
model has uniform wages in the cross section and, when calibrated to
U.S. job finding rates, short unemployment spells. Thus, job loss has
little impact on present-value earnings. Because so little is at stake
in the destruction of employment relationships in the basic MP model, it
cannot rationalize the earnings losses associated with job displacement.
Lastly, we evaluate an MP model of Burgess and Turon (2010) with
search on the job and replacement hiring. Unlike the basic MP model,
Burgess and Turon's model is at least qualitatively consistent with
several first-order features of the data: cross-sectional wage
dispersion, worker flows in excess of job flows, and recessionary spikes
in job destruction and unemployment inflows. The model also exhibits a
job ladder that prolongs the earnings recovery path after displacement.
When calibrated to match U.S. job finding rates, job loss in the model
produces present-value earnings losses that, on average, are about
one-quarter of the mean empirical losses due to job displacement. This
is a sizable improvement over the basic MP model, but it leaves a very
large gap between theory and evidence. Moreover, the model cannot
explain the larger losses for displacements that occur in recessions,
because negative aggregate shocks trigger the destruction of lower-value
jobs in the model.
In our view, a major shortcoming of existing MP models of
unemployment fluctuations is their implication that job loss is a rather
inconsequential event for the affected workers. The consequences of job
displacement, and fears of displacement, are among the main reasons that
recessions and high unemployment create so much concern in the general
population. The negative consequences of job displacement are why
unemployment is such a potent political issue. We also think the serious
consequences of job displacement are a major reason that unemployment
and unemployment fluctuations attract so much attention from economists.
It is important to put our criticism of MP models in proper
context. We see MP models, in particular, and the larger class of
Diamond-Mortensen-Pissarides models as a great advance. These models
deliver a coherent theory of frictional unemployment and its
determinants. They provide an analytical framework for studying cyclical
movements in unemployment, vacancies, job finding rates, and the joint
dynamics of worker flows and job flows. They provide tools for analyzing
search-and-matching behavior by employers and job seekers, and for
studying the implications of search-and-matching frictions for wage
dispersion and individual wage dynamics. These tools are widely used to
study the effects of policies, wage setting arrangements, and other
economic institutions on unemployment and a variety of other labor
market outcomes.
We hope to see these models taken in directions that can explain
large and lasting earnings losses at job displacement. There are
potentially several ways to bring MP-type models closer to the evidence
on the earnings losses associated with job displacement. Models that
incorporate learning about match quality over time (as in Jovanovic
1979), the acquisition of specific skills through learning-by-doing on
the job, and investments in specific training (as in Becker 1962) could
yield substantial earnings losses upon job loss. These three mechanisms
influence match durability and the evolution of surplus values in
ongoing matches. It would be useful to integrate these mechanisms into
MP models of unemployment fluctuations, which have thus far devoted much
greater attention to the forces governing match formation. Robert Topel
(1990) and Derek Neal (1995), among others, argue that specific forms of
human capital play a central role in determining the magnitude of
earnings losses associated with job displacement. Lars Ljungqvist and
Thomas Sargent (1998) build an equilibrium search model that hardwires a
link between job loss and the destruction of human capital, and that
includes further human capital depreciation during unemployment.
Workers may also enjoy rents for reasons apart from
search-and-matching frictions and returns on specific human capital.
Other explanations for worker rents include fairness norms and concerns
about pay equity (Akerlof and Yellen 1990), high pay as a device to
deter shirking (Bulow and Summers 1986), the appropriation of
quasi-rents generated by sunk investments (Grout 1984, Caballero and
Hammour 2005), and worker sharing of product market rents. Paul Beaudry
and John DiNardo (1991) stress the role of long-term contracting and
one-sided commitment as a source of downward wage stickiness. Johannes
Schmieder and von Wachter (2010) consider workers who receive higher
wages as a consequence of tight labor market conditions in the past.
They find evidence that these workers experience higher layoff rates and
lose their wage premiums upon job loss, a pattern of results that
supports the presence of rents. Whether this pattern accounts for larger
earnings losses in recessions, when displacements are more widespread,
is an open question.
Workers who enter the labor market in periods of slack conditions
suffer negative effects on future earnings that persist for 10 years or
more (see, for example, Kahn 2010). Both lasting declines in employer
quality and lasting effects of low starting wages on wage growth within
firms contribute to the persistent negative earnings effects of slack
conditions at entry (see, for example, Oreopoulos, Heisz, and von
Wachter forthcoming). These results are interesting in part because new
entrants have not accumulated job-specific rents and are unlikely to
have accumulated much in the way of specific human capital. Apparently,
weak conditions at the time of labor market entry slow the accumulation
of rents and specific human capital for many years thereafter. Similar
forces could lower the future earnings prospects of workers who are
displaced in recessions and slumps.
ACKNOWLEDGMENTS We thank Bob Hall, Richard Rogerson, the editors,
and conference participants for many helpful comments on an earlier
draft. April Chen, Olga Deriy, and Gregor Jarosch provided outstanding
research assistance. Steven Davis thanks the University of Chicago Booth
School of Business for research support. The authors report no conflicts
of interest.
References
Akerlof, George A., and Janet L. Yellen. 1990. "The Fair
Wage-Effort Hypothesis and Unemployment." Quarterly Journal of
Economics 105, no. 2: 255-83.
Autor, David H., and Lawrence F. Katz. 1999. "Changes in the
Wage Structure and Earnings Inequality." In Handbook of Labor
Economics, vol. 3A, edited by O. Ashenfelter and D. Card. Amsterdam:
North-Holland.
Bartlesman, Eric J., and Mark Doms. 2000. "Understanding
Productivity: Lessons from Longitudinal Microdata." Journal of
Economic Literature 38 (September): 569-94.
Beaudry, Paul, and John DiNardo. 1991. "The Effects of
Implicit Contracts on the Movement of Wages over the Business Cycle:
Evidence from Micro Data." Journal of Political Economy 99, no. 4:
665-88.
Becker, Gary S. 1962. "Investment in Human Capital: A
Theoretical Analysis." Journal of Political Economy 70, no. 5, part
2: 9-49.
Binmore, Ken, Ariel Rubinstein, and Asher Wolinsky. 1986. "The
Nash Bargaining Solution in Economic Modelling." RAND Journal of
Economics 17, no. 2: 176-88.
Blanchard, Olivier J., and Peter Diamond. 1989. "The Beveridge
Curve." BPEA, no. 2: 1-60.
Browning, Martin, and T. F. Crossley. 2001. "Unemployment
Insurance Levels and Consumption Changes." Journal of Public
Economics 80, no. 1: 1-23.
Bulow, Jeremy I., and Lawrence H. Summers. 1986. "A Theory of
Dual Labor Markets with Application to Industrial Policy,
Discrimination, and Keynesian Unemployment." Journal of Labor
Economics 4, no. 3 (part 1): 376-414.
Burdett, Kenneth, and Dale T. Mortensen. 1998. "Wage
Differentials, Employer Size, and Unemployment." International
Economic Review 39, no. 2.
Bureau of Labor Statistics. 2010. "Worker Displacement:
2007-2009." News release. USDL-10-1174.
www.bls.gov/news.release/disp.nr0.htm.
Burgard, Sarah A., Jennie E. Brand, and James S. House. 2007.
"Toward a Better Estimation of the Effect of Job Loss on
Health." Journal of Health and Social Behavior 48, no. 4: 369-84.
Burgess, Simon, and Helene Turon. 2010. "Worker Flows, Job
Flows and Unemployment in a Matching Model." European Economic
Review 54, no. 3 (April): 393-408.
Caballero, Ricardo J., and Mohamad L. Hammour. 2005. "The
Costs of Recessions Revisited: A Reverse-Liquidationist View."
Review of Economic Studies 72, no. 2 (April): 313-41.
Charles, Kerwin Kofi, and Melvin Stephens. 2004. "Disability,
Job Displacement and Divorce." Journal of Labor Economics 22, no.
2: 489-522.
Congressional Budget Office. 2004. "Family Income of
Unemployment Insurance Recipients." Washington (March).
Costain, J., and M. Reiter. 2008. "Business Cycles,
Unemployment Insurance, and the Calibration of Matching Models."
Journal of Economic Dynamics and Control 32, no. 4:1120-55.
Couch, Kenneth A., and Dana W. Placzek. 2010. "Earnings Losses
of Displaced Workers Revisited." American Economic Review 100, no.
1: 572-89.
Davis, Steven J. 2005. Comment on "Job Loss, Job Finding and
Unemployment in the U.S. Economy over the Past Fifty Years" by
Robert Hall. NBER Macroeconomics Annual 20: 139-57.
Davis, Steven J., and John Haltiwanger. 1990. "Gross Job
Creation and Destruction: Microeconomic Evidence and Macroeconomic
Implications." NBER Macroeconomics Annual 5:123-68.
Davis, Steven J., R. Jason Faberman, and John Haltiwanger. 2006.
"The Flow Approach to Labor Markets: New Data Sources and
Micro-Macro Linkages." Journal of Economic Perspectives 20, no. 3:
3-26.
--. 2012. "Labor Market Flows in the Cross Section and over
Time." Journal of Monetary Economics 59, no. 1:1-18.
Davis, S. J., R. J. Faberman, J. C. Haltiwanger, and I. Rucker.
2010. "Adjusted Estimates of Worker Flows and Job Openings in
JOLTS." In Labor in the New Economy, edited by K. Abraham, M.
Harper, and J. R. Spletzer. University of Chicago Press.
Den Haan, Wouter, Garey Ramey, and Joel Watson. 2000. "Job
Destruction and the Experiences of Displaced Workers."
Carnegie-Rochester Conference Series on Public Policy 52: 87-128.
Eliason, M., and D. Storrie. 2009. "Does Job Loss Shorten
Life?" Journal of Human Resources 44, no. 2: 277-302.
Elsby, Michael, Ryan Michaels, and Gary Solon. 2009. "The Ins
and Out of Cyclical Unemployment." American Economic Journal:
Macroeconomics 1, no. 1: 84-110.
Eyigungor, Burcu. 2010. "Specific Capital and Vintage Effects
on the Dynamics of Unemployment and Vacancies." American Economic
Review 100, no. 3: 1214-37.
Farber, Henry. 1999. "Alternative and Part-Time Employment
Arrangements as a Response to Job Loss." Journal of Labor Economics
17, no. 4, part 2 (October): S142-S169.
Foster, Lucia, John Haltiwanger, and C. J. Krizan. 2001.
"Aggregate Productivity Growth: Lessons from Microeconomic
Evidence." In New Directions in Productivity Analysis, edited by
Edward Dean, Michael Harper, and Charles Hulten. University of Chicago
Press.
Frey, Bruno S., and Alois Stutzer. 2002. "What Can Economists
Learn from Happiness Research?" Journal of Economic Literature 40,
no. 2 (June): 402-35.
Gertler, Mark, and Antonella Trigari. 2009. "Unemployment
Fluctuations with Staggered Nash Wage Bargaining." Journal of
Political Economy 117, no. 1: 38-86.
Grout, Paul A. 1984. "Investment and Wages in the Absence of
Binding Contracts." Econometrica 52 (March): 449-60.
Gruber, Jonathan. 1997. "The Consumption Smoothing Benefits of
Unemployment Insurance." American Economic Review 87, no. 1:
192-205.
Hagedorn, Marcus, and Iourii Manovskii. 2008. "The Cyclical
Behavior of Equilibrium Unemployment and Vacancies Revisited."
American Economic Review 98, no. 4: 1692-1706.
Hall, Robert E. 1995. "Lost Jobs." BPEA, no. 1: 221-73.
--. 2005. "Employment Fluctuations with Equilibrium Wage
Stickiness." American Economic Review 95, no. 1: 50-65.
Hall, Robert E., and Paul R. Milgrom. 2008. "The Limited
Influence of Unemployment on the Wage Bargain." American Economic
Review 98, no. 4: 1653-74.
Hertweck, Matthias S. 2010. "Endogenous On-the-Job Search and
Frictional Wage Dispersion." WWZ Discussion Paper no. 2010/02.
University of Basel.
Hornstein, Andreas, Per Krusell, and Giovanni L. Violante. 2010.
"Frictional Wage Dispersion in Search Models: A Quantitative
Assessment." Federal Reserve Bank of Richmond, Princeton
University, and New York University (August 10).
Jacobson, Louis, Robert LaLonde, and Daniel Sullivan. 1993.
"Earnings Losses of Displaced Workers." American Economic
Review 83, no. 4: 685-709.
Jovanovic, Boyan. 1979. "Job Matching and the Theory of
Turnover." Journal of Political Economy 87, no. 5: 972-90.
Kahn, Lisa. 2010. "The Long-Term Labor Market Consequence of
Graduating College in a Bad Economy." Labor Economics 17, no. 2:
303-16.
Kennan, John. 2009. "Private Information, Wage Bargaining and
Employment Fluctuations." Review of Economic Studies 77, no. 2:
633-64.
Kilponen, Juha, and Juuso Vanhala. 2011. "The Sensitivity of
Job Destruction to Vintage and Tenure Effects." Bank of Finland
Discussion Paper no. 1080, revised. Helsinki.
Kletzer, Lori. 1989. "Returns to Seniority after Permanent Job
Loss." American Economic Review 79, no. 3: 536-43.
Kodrzycki, Yolanda K. 2007. "Using Unexpected Recalls to
Examine the Long-Term Earnings Effects of Job Displacement."
Federal Reserve Bank of Boston Working Paper no. W07-2. Boston, Mass.
Ljungqvist, Lars, and Thomas J. Sargent. 1998. "The European
Unemployment Dilemma." Journal of Political Economy 106, no. 3
(June): 514-50.
Meyer, Bruce D., Wallace K. C. Mok, and James X. Sullivan. 2010.
"The Under-Reporting of Transfers in Household Surveys: Its Nature
and Consequences." Working Paper no. 15181. Cambridge, Mass.:
National Bureau of Economic Research.
Mortensen, Dale, and Eva Nagypal. 2007. "More on Unemployment
and Vacancy Fluctuations." Review of Economic Dynamics 10, no. 3:
327-47.
Mortensen, Dale, and Christopher Pissarides. 1994. "Job
Creation and Job Destruction in the Theory of Unemployment." Review
of Economic Studies 61 (July): 397-415.
Neal, Derek. 1995. "Industry-Specific Human Capital: Evidence
from Displaced Workers." Journal of Labor Economics 13, no. 4:
653-77.
Oreopoulos, Philip, Andrew Heisz, and Till von Wachter.
Forthcoming. "Short- and Long-Term Career Effects of Graduating in
a Recession." American Economic Journal: Applied Economics.
Oreopoulos, Philip, Marianne Page, and Ann Huff Stevens. 2008.
"The Intergenerational Effects of Worker Displacement."
Journal of Labor Economics 26, no. 3: 455-83.
Pissarides, Christopher. 2009. "The Unemployment Volatility
Puzzle: Is Wage Stickiness the Answer?" Econometrica 77, no.
5:1339-69.
Poletaev, Maxim, and Chris Robinson. 2008. "Human Capital
Specificity: Evidence from the Dictionary of Occupational Titles and
Displaced Worker Surveys, 1984-2000." Journal of Labor Economics
26, no. 3: 387-420.
Polivka, Anne E., and Stephen M. Miller. 1998. "The CPS after
the Redesign: Refocusing the Economic Lens." In Labor Statistics
Measurement Issues, edited by John Haltiwanger, Marilyn E. Manser, and
Robert Topel. University of Chicago Press for the National Bureau of
Economic Research.
Postel-Vinay, Fabien, and Jean-Marc Robin. 2002. "Equilibrium
Wage Dispersion with Worker and Employer Heterogeneity."
Econometrica 70, no. 6 (November): 2295-2350.
Ramey, Garey. 2008. "Exogenous vs. Endogenous
Separations." Working paper. University of California, San Diego
(October).
Rege, Mari, Kjetil Telle, and Mark Votruba. 2009. "The Effect
of Plant Downsizing on Disability Pension Utilization." Journal of
the European Economic Association 7, no. 4: 754-85.
Robin, Jean-Marc. 2011. "On the Dynamics of Unemployment and
Wage Distributions." Econometrica 79, no. 5 (September): 1327-55.
Ruhm, Christopher. 1991. "Are Workers Permanently Scarred by
Job Displacements?" American Economic Review 81: 319-23.
Rupp, Kalman, and David Stapleton. 1995. "Determinants of the
Growth in the Social Security Administration's Disability Programs:
An Overview." Social Security Bulletin 58, no. 4: 43-70.
Schmieder, Johannes, and Till von Wachter. 2010. "Does Wage
Persistence Matter for Employment Fluctuations? Evidence from Displaced
Workers." American Economic Journal: Applied Economics 2, no. 3:
1-21.
Schmieder, Johannes, Till von Wachter, and Stefan Bender. 2009.
"The Effects of Unemployment Insurance on Labor Supply and Search
Outcomes: Regression Discontinuity Estimates from Germany."
Department of Economics Discussion Paper Series no. DP0910-08. Columbia
University.
Schoeni, Robert, and Michael Dardia. 2003. "Estimates of
Earnings Losses of Displaced Workers Using California Administrative
Data." PSC Research Report no. 03-543. Population Studies Center,
Institute for Social Research, University of Michigan.
Shimer, Robert. 2004. "The Consequences of Rigid Wages in
Search Models." Journal of the European Economic Association 2, no.
2-3: 469-79.
--. 2005. "The Cyclical Behavior of Equilibrium Unemployment
and Vacancies." American Economic Review 95, no. 1: 25-49.
--. 2007. "Reassessing the Ins and Outs of Unemployment."
Working Paper no. 13421. Cambridge, Mass.: National Bureau of Economic
Research.
--. 2010. Labor Markets and Business Cycles. Princeton University
Press.
Silva, Jose Ignacio, and Manuel Toledo. 2009. "Labor Turnover
Costs and the Behavior of Vacancies and Unemployment."
Macroeconomic Dynamics 13, S1: 76-96.
Stephens, Melvin, Jr. 2004. "Job Loss Expectations,
Realizations, and Household Consumption Behavior." Review of
Economics and Statistics 86, no. 1 (February): 253-69.
Stevens, Ann Huff. 1997. "Persistent Effects of Job
Displacement: The Importance of Multiple Job Losses." Journal of
Labor Economics, 15, no. 1, part 1: 165-88.
Stevens, Ann, and Jesamyn Schaller. 2011. "Short-Run Effects
of Parental Job Loss on Children's Academic Achievement."
Economics of Education Review 30, no. 2: 289-99.
Sullivan, Daniel, and Till von Wachter. 2009. "Job
Displacement and Mortality: An Analysis Using Administrative Data."
Quarterly Journal of Economics 124, no. 3: 1265-1306.
Topel, Robert. 1990. "Specific Capital and Unemployment:
Measuring the Costs of Worker Displacement." Carnegie-Rochester
Series on Public Policy 33 (Autumn): 181-214.
Von Wachter, Till. 2010. "Long-Term Unemployment: Causes,
Consequences and Solutions." Testimony before the Joint Economic
Committee of the U.S. Congress, April 29. Columbia University.
Von Wachter, Till, and Elizabeth Weber Handwerker. 2009.
"Variation in the Cost of Job Loss by Worker Skill: Evidence Using
Matched Data from California, 1991-2000." Columbia University.
Von Wachter, Till, Elizabeth Weber Handwerker, and Andrew Hildreth.
2008. "Estimating the 'True' Cost of Job Loss: Evidence
Using Matched Data from California 1991-2000." Center for Economic
Studies Working Paper no. 09-14. Washington: U.S. Bureau of the Census.
Von Wachter, Till, Jae Song, and Joyce Manchester. 2011.
"Long-Term Earnings Losses Due to Mass-Layoffs during the 1982
Recession: An Analysis Using Longitudinal Administrative Data from 1974
to 2008." Columbia University.
Wightman, Patrick. 2009. The Effect of Parental Job Loss on
Children. Ph.D. dissertation, Irving B. Harris Graduate School of Public
Policy Studies, University of Chicago.
STEVEN J. DAVIS
University of Chicago
TILL VON WACHTER
Columbia University
(1.) See, for example, Jacobson, Lalonde, and Sullivan (1993),
Couch and Placzek (2010), and von Wachter, Song, and Manchester (2011).
(2.) We review the evidence and provide citations to the relevant
literature in section III. See also von Wachter (2010).
(3.) The BED contains longitudinally linked records for all
businesses covered by state unemployment insurance agencies, making it
virtually a census of nonfarm private business establishments.
(4.) To deal with weaknesses in the JOLTS sample design, Davis and
others (2012) rely on BED data to track the cross-sectional distribution
of establishment-level growth rates over time. They combine micro data
from the BED and the JOLTS to obtain the layoff series in figure 1. To
extend the layoff series back in time before the advent of the JOLTS,
they use the BED to construct synthetic, JOLTS-like layoff rates. Davis
and others (2010) discuss sample design issues in the JOLTS and develop
the adjustment methodology implemented by Davis and others (2012).
(5.) See Bartlesman and Doms (2000) and Foster, Haltiwanger, and
Krizan (2001) for reviews of the evidence on reallocation and
productivity growth.
(6.) This pattern holds in earlier postwar U.S. recessions as well.
See, for example, Blanchard and Diamond (1989), Davis and Haltiwanger
(1990), Davis, Faberman, and Haltiwanger (2006), and Elsby, Michaels,
and Solon (2009).
(7.) As an example, the Conference Board uses new claims for UI
benefits in constructing its Leading Economic Index. See Conference
Board, "Global Business Cycle Indicators,"
www.conference-board.org/data/bcicountry.cfm?cid=1.
(8.) Figure 2 cumulates weekly UI claims over 12 months, but the
calculations otherwise follow the same approach as in figure 1. The BDS
job destruction series are available at an annual frequency and extend
further back in time than the BED-based job destruction series in figure
1, but they are not as timely. Because the BDS series reflect 12-month
changes in establishment-level employment, they are not directly
comparable to the BED-based job destruction series based on 3-month
changes.
(9.) Tabulations in Davis and others (2006) based on BED and JOLTS
data indicate that most employment reductions are achieved through
layoffs when firms contract by 30 percent or more.
(10.) The very high rates of initial UI claims in the early 1980s
should be interpreted with caution. Temporary layoffs were a major
phenomenon in the early 1980s, unlike in later recessions, and many
temporarily laid-off workers qualified for UI benefits. Since few
temporary layoff spells last more than a full year, and given that our
definition of a mass layoff excludes temporary firm-level fluctuations,
temporary layoffs play little role in our job displacement measure. For
similar reasons, temporary layoffs have little impact on the annual job
destruction measures.
(11.) In calculating the data for this figure, we allow the at-risk
population to change from year to year. For some purposes it is more
appropriate to consider the cumulative displacement rate for a fixed
at-risk population. Consider, for example, the population of male
workers younger than 50 with 3 or more years of job tenure at firms with
at least 50 employees as of 1979, working in industries with continuous
SSA coverage. By our criteria 16 percent of this fixed population
experienced a job displacement event during 1980-85.
(12.) We include displacements that occur in y + 1 and y + 2 in the
sample for displacement year y to raise the number of observations of
displaced workers, and to align the inclusion windows for displaced and
control group workers. Note that this approach smooths the estimated
earnings effects of job displacement from one displacement year to the
next, which works against finding differences between recessions and
expansions.
(13.) Since our sample window stops in 2008, for displacement years
after 1988 we do not observe 20 years of earnings data after a
displacement. For these years, the postdisplacement dummies are included
up to the maximum possible number of years.
(14.) For 1980 the baseline is years 5 and 6 before displacement,
and for 1981 it is years 6 and 7 before displacement. We also drop the
dummy variable for the first calendar year in each regression. These
zero restrictions, two for the baseline and one for the first calendar
year, resolve the potential collinearity among the dummy variables in
equation 1.
(15.) If the out-year earnings recovery is faster for displacements
that occur in booms, this choice understates the cyclical differences in
the cost of job loss.
(16.) Similarly, we calculate the corresponding mean of actual
annual earnings before and after displacement by first obtaining the
average for each displacement year, [[bar.e].sup.act., y.sub.t] and then
averaging over the years belonging to expansions and recessions.
(17.) The online appendix, accessible on the Brookings Papers web
site, www.brookings. edu/economics/bpea.aspx, under "Past
Editions," contains additional results by age group.
(18.) Stephens (2004) provides survey-based evidence that
subjective assessments of job loss probabilities have considerable
predictive power for future layoffs at the individual level, even when
conditioning on standard demographic variables that are correlated with
layoff risks. Nevertheless, his main empirical specification yields no
evidence of a relationship between job loss expectations and household
consumption conditional upon losing a job.
(19.) There appear to be few previous efforts to evaluate whether
equilibrium search-and-matching models can account for the earnings
losses associated with job displacement. An exception is Den Haan,
Ramey, and Watson (2000). Davis (2005) provides some
back-of-the-envelope calculations. The loss of earnings potential upon
job loss is an important element in the theoretical model of high
European unemployment rates developed by Ljungqvist and Sargent (1998).
(20.) See, for example, Shimer (2004, 2010), Hall (2005), Gertler
and Trigari (2009), and Kennan (2009). Mortensen and Nagypal (2007),
Ramey (2008), Pissarides (2009), Burgess and Turon (2010), and Eyigungor
(2010), among others, propose alternative resolutions to the
unemployment volatility puzzle.
(21.) We could refine the treatment-control comparisons in table 4
by replicating the employment stability criterion used for controls in
section II. This type of refinement may make sense in future research.
Given the uniformity of wages and the small consequences of job loss in
table 4, however, we do not think the basic MP model can explain the
evidence on earnings losses or rationalize strong concerns about job
loss and unemployment.
(22.) See the online appendix for a version of table 5 that adopts
their calibration, which is meant to match features of the British
economy from 1964 to 1999.
(23.) In practice, empirical treatment-control comparisons do not
perfectly condition on the idiosyncratic component of jobs and match
values. However, as long as the empirical specification at least partly
captures a disproportionate loss of marginal jobs in the wake of a
negative aggregate shock, the composition effect we highlight here will
also be present in the empirical estimates of earnings losses associated
with job loss in a recession.
(24.) Postel-Vinay and Robin (2002) consider a different model with
search on the job and heterogeneity in productivity on both sides of the
labor market. Employers have all the bargaining power, and newly
reemployed workers start at the bottom of the wage distribution after an
unemployment spell. When an employed worker finds an attractive outside
opportunity, the incumbent employer may respond with a successful
counteroffer (a wage increase). Thus, the model of Postel-Vinay and
Robin also yields a prolonged earnings recovery path after job loss that
is tied to search on the job, but wage gains may or may not coincide
with job changes.
Comments and Discussion
COMMENT BY
ROBERT E. HALL The crisis of 2008 and its aftermath caused large
increases in the incidence of involuntary job loss and large increases
in subsequent earnings losses, as replacement jobs have become much
harder to find. In this paper Steven Davis and Till von Wachter
contribute to research on this key topic in two ways: by providing a
detailed analysis, based on von Wachter's earlier work, of data on
the earnings of individual workers following mass layoffs; and by
examining the leading class of models of unemployment and labor
turnover, that developed by Peter Diamond, Dale Mortensen, and
Christopher Pissarides, to compare the consequences of job loss in those
models with the findings of the new empirical work.
What do the authors mean by the "cost of job loss"? They
measure it as the difference in subsequent earnings between workers who
retained their jobs during a mass layoff and those who suffered layoffs.
The entire focus is on personal rather than social loss: if workers who
are highly paid relative to their productivity suffer layoffs and are
immediately hired elsewhere at normal wages and the same productivity,
it is a private loss--a transfer of rents--but not a social loss.
Measuring the social loss would involve a host of issues for which
appropriate data are lacking, and even some, such as the right choice of
social welfare function, that bring in deep conceptual disagreements. By
defining the "cost of job loss" as they do, the authors pose a
question that is potentially answerable based on the excellent data they
use.
A related point is that the paper focuses on measuring only the
losses caused by mass layoffs rather than those caused by the
fundamental underlying forces that result in, among other things, mass
layoffs. Again, the reader has to decide how to relate the information
the authors extract from the data to the deeper issues of the harm to
society from, for example, policy failures that have dramatically
increased labor market volatility. The authors effectively exploit their
comparative advantage in providing good summaries of the sample evidence
that economists interested in improving policy need to know about.
As the authors discuss, but seemingly only as an afterthought,
individual workers do not suffer layoffs as a result of random
selection; the data are drawn from normal experience and not from a
controlled experiment. Thus, the issue of potential biases from
nonrandom selection arises. Employers have an incentive to discharge
workers whose pay is high relative to their productivity. Consider the
results of a study comparing the earnings of victims of mass layoffs
with the earnings of workers at firms with no mass layoffs. The victims
are differentially workers who have high wages relative to productivity.
In subsequent employment, these workers are likely to receive pay closer
to the norm for their productivity. And the same thing would have
happened to them, to some extent, without the layoffs. Thus, not all the
decline in earnings observed among layoff victims compared with workers
at firms without layoffs is the result of the layoffs--part would have
occurred anyway. What the statistical procedure measures is the sum of
the causal effect and the selection effect. Therefore, comparing
laid-off workers with those at firms without layoffs exaggerates the
consequences of layoff, because the subsequent wage growth of those not
laid off will be faster, on average, than the growth that the layoff
victims would have experienced but for the layoff. Productivity is
largely unobserved, so the control variables in the regressions do not
fully adjust for the problem.
This proposition has a flip side that is helpful in measuring the
selection effect. The workers at firms with mass layoffs who are not
laid off are also subject to a selection effect. They tend to be the
workers with low wages relative to productivity. This condition, too,
would tend to disappear over time, so the same statistical procedure
applied to compare nonvictims at firms with layoffs with workers at
firms without layoffs would show a positive effect after the layoff. It
would also tend to show wage declines, relative to workers at firms
without layoffs, prior to the layoff.
My figure 1 shows what the results would look like if selection
were part of the story. The earnings of victims at layoff firms would
tend to rise prior to the layoff, fall dramatically after the layoff,
and recover subsequently. The earnings of nonvictims at firms with
layoffs would fall prior to the layoffs and rise later, again relative
to earnings at nonlayoff firms.
The paper covers this point very briefly, by citing evidence from
von Wachter's earlier paper (von Wachter, Song, and Manchester
2011). Rather than compare the victims with the nonvictims, however,
that paper compares the results for victims alone with the results for
all employees of firms with mass layoffs. After adjustment for the
fraction of the workers who were victims, the results of the second
approach were similar to those for victims only. This is a quite
roundabout way of making the comparison I have suggested--their finding
implies that no favorable selection effect operated among the
nonvictims. As it happens, however, von Wachter did undertake the
comparison I had in mind and was kind enough to send me the results,
which are presented in figure 2. That figure shows essentially no effect
for nonvictims. Rather than tracking the paths in figure 1, the paths
for nonvictims are flat.
[FIGURE 1 OMITTED]
Although the evidence in figure 2 is impressive--and surely
deserves to be in this paper in place of the brief and opaque summary of
the evidence presented instead--it is not completely dispositive,
because it rests on the identifying hypothesis that the forces that
caused the layoffs had no effects on the firm. Subsequent earnings
differences among those not laid off combine the favorable effect of
selection with the unfavorable effects of continuing employment at a
firm that has suffered a large reversal resulting in mass layoffs. The
finding that the nonvictims had roughly zero earnings effects means that
the two effects offset each other, not that they are both zero.
Another possibility is that mass layoffs occur in firms that have
permanent unobserved characteristics that make them more susceptible to
mass layoffs. Von Wachter and others (2011) test this hypothesis by
including firm fixed effects in their regressions. The resulting
estimates capture only the earnings losses within the workers of each
firm, because the fixed effects pick up differences across firms.
Because the within-firm earnings losses remain substantial, although
smaller, the authors reach the reasonable conclusion that at least that
amount of losses is actually attributable to the mass layoffs.
[FIGURE 2 OMITTED]
One way to think about the selection issues in general is to
consider the following hypothetical. A survey asks, "At any time in
your career, were you laid off from a job that you had held at least 3
years?" An econometrician includes a dummy for a yes answer in a
Mincer log wage regression for a sample of 55-year-olds and gets a
coefficient of -0.06 with a standard error of 0.01. Most of us would
interpret this finding in terms of a
selection-and-unobserved-characteristics story as well as a
cost-of-layoff story.
Nonetheless, there is no serious doubt in my mind that a mass
layoff inflicts substantial personal earnings losses on its victims for
at least a few years and probably more. I think there is more doubt
about the permanent loss, which could arise from selection.
The second major contribution of the paper is to confront the
leading model of unemployment and labor turnover--the
Diamond-Mortensen-Pissarides (DMP) model--with the findings of
substantial earnings losses among victims of mass layoffs. It is easy to
explain the basic issue here, although the authors defer this
explanation until well into their exposition of the plumbing of the
model. Evidence on the cost of recruiting suggests that, right after a
new hire, an employer has about $1,000 invested in the worker. (1) The
bargaining structure of the DMP model interprets this amount as the
employer's capitalized share of the surplus the job generates. If
the bargain splits the surplus roughly equally, the worker has a similar
stake in the job. The worker's loss from a layoff that occurs
immediately after the hire would thus be about $1,000, which is far
below the figure that the paper calculates for the typical layoff
occurring 3 or more years after the hire.
The DMP model as normally developed is focused on unemployment and
is exceedingly stripped down with respect to how the typical employment
relationship evolves after the hire. Ali that matters for the analysis
of unemployment is the present value of the expected margin the employer
will earn from the relationship from the difference between the
worker's productivity and the worker's wage. Given the
objective of the model, it is no shortcoming that the model cannot
generate realistically big figures for the consequences of job loss.
In their conclusion, the authors lay out some of the ideas from
labor economics that would belong in a master model of the employment
relationship that deals both with the issues that gave rise to the DMP
model and with many issues of governance of the ongoing relationship. I
think the paper performs an important service in making it clear that
the master model faces an important challenge in explaining how workers
move from having, on average, only a roughly $1,000 stake in a brand-new
job to having around $100,000 at stake after more than 3 years of
tenure. The implied gradient of accumulation of the worker's share
of job-specific capital is remarkably steep, and thus a real challenge
to empirical model builders.
REFERENCES FOR THE HALL COMMENT
Hall, Robert E., and Paul R. Milgrom. 2008. "The Limited
Influence of Unemployment on the Wage Bargain." American Economic
Review 98, no. 4: 1653-74.
Von Wachter, Till, Jae Song, and Joyce Manchester. 2011.
"Long-Term Earnings Losses Due to Mass-Layoffs during the 1982
Recession: An Analysis Using Longitudinal Administrative Data from 1974
to 2008." Columbia University.
(1.) Hall and Milgrom (2008) report that recruiting cost is 0.43
day of pay per day a vacancy is held open. Daily earnings of the average
American worker are $153, so the daily vacancy cost is $66. According to
JOLTS data, it takes 16 days for a vacancy to be filled, so the cost of
recruiting one new worker is 16 x $66 = $1,066. Under the zero-profit
condition of the DMP model, the value of the employer's share of
the surplus is the cost of recruiting the worker. With a symmetric Nash
bargain, the worker's share has the same value.
COMMENT BY
RICHARD ROGERSON Steven Davis and Till von Wachter have written a
very nice paper that one hopes will motivate much follow-up research. I
would summarize the broad theme of the paper as follows. The data reveal
a lot of heterogeneity in the nature of unemployment experiences.
Although many, perhaps most, unemployment spells are relatively short
and seem not to be associated with any persistent negative outcomes
beyond the short-term loss in income, the displacement events that are
the authors' focus are associated with substantial long-term
losses. (1) Much recent work on unemployment dynamics has focused on
accounting for movements in the level of unemployment and its breakdown
into inflows and outflows. Davis and von Wachter argue that a
"good" theory of unemployment should account not only for
changes in unemployment inflows and outflows but also for the nature of
individual unemployment spells, in particular the experiences of the
group they define as "displaced workers." They go on to show
that a large set of commonly used models of unemployment dynamics fail
in this regard. This paper can then be interpreted as issuing a
challenge to researchers to develop richer models of wage and
unemployment dynamics.
I think this general message is an important one. My comments will
focus on two broad points. First, in the context of documenting the
facts regarding the earnings losses of displaced workers, I will note
some additional information that would be useful to have. Second,
regarding the need for models of unemployment dynamics that reflect the
experiences of displaced workers, I will argue that the research agenda
should be broadened somewhat, relative to what the authors call for.
Beyond simply asking the models to also account for the experiences of
displaced workers, I think the larger issue is to develop a unified
theory of worker flows and wage dynamics more generally.
EARNINGS LOSSES OF DISPLACED WORKERS: EMPIRICAL FINDINGS The
previous literature on the earnings losses of displaced workers has
documented that workers with at least moderate job tenure who are
displaced from medium-size and large firms in mass-layoff events suffer
long-lasting decreases in earnings. The authors contribute to this
literature by documenting how these losses vary with the state of the
economy at the time of displacement. The main finding is that the
present value of losses for a worker displaced during a recession is
roughly twice that for a worker displaced during an expansion. I think
this is an interesting finding, but that some additional information
would be valuable, some of which also applies to the earlier literature.
I note in advance that, for the most part, I am abstracting from
constraints imposed by data availability.
The data used by the authors allow them to measure earnings losses
after displacement but do not allow them to decompose these losses into
the separate components due to unemployment, reductions in hours in
subsequent employment, and reductions in subsequent wages. There is some
suggestion that persistent earnings losses are not dominated by the
first component, but it is surely relevant for short-term earnings
losses, and hence for total present-value losses, and its importance may
well vary with the business cycle. In view of the models that the
authors consider in section IV of the paper, it is necessarily of
interest to know more about the exact role of unemployment in accounting
for these losses. Beyond that, there is still a lot of scope for changes
in working hours to play a significant role. If they do, it might also
be of interest to explore how the change in hours is correlated with
other variables. For example, given that many households have two
earners, displacement of one member could lead to a reallocation of
market work across members. The interesting issue here is the
possibility that individuals might choose to work fewer hours after
experiencing a displacement, which implies that there is an endogenous
component to the earnings losses.
A second issue is that, like most of the related literature, the
authors' analysis focuses entirely on mean earnings losses relative
to a control group. It would be worth knowing more about the
distribution of earnings losses and how they correlate with other
factors. If there are compositional differences between displaced
workers in expansions and recessions, one would like to know whether
these differences can account for the cyclical variation in earnings
losses measured by the authors. For example, the authors' data work
reveals that recessions are times when relatively more long-tenured
workers are displaced, and if losses are increasing in tenure, this
could explain part of the gap. Similarly, if the ultimate losses depend
upon how quickly an individual is able to secure employment following
displacement, it is of interest to assess the extent to which cyclical
changes in unemployment duration can account for the higher earnings
losses during recessions. More generally, how does the distribution of
losses differ between recessions and expansions? Is there simply a shift
in the distribution as one moves through the cycle, or are there notable
changes in the shape of the distribution?
Since the measured gap in earnings losses between recessions and
expansions reflects changes relative to a control group, it is also
relevant to ask what fraction of the gap is accounted for by changes in
the control group's earnings. It is certainly possible that the
wage gains of the control group are quite different following a
recession than following an expansion.
Finally, it is of interest to compare the cyclical gap in earnings
losses for displaced workers with other, related measures of how labor
market outcomes differ with the state of the business cycle. In
particular, a related literature has found that college students who
graduate during a recession face persistent earnings losses relative to
those who graduate during expansions. Although displaced workers and
college graduates are very different populations, they face the common
problem of needing to find employment. It would be worth knowing how the
magnitudes of these effects compare, and more generally the extent to
which these two empirical findings reflect the same underlying economic
forces. Put somewhat differently, the cyclical variation in earnings
losses for displaced workers may not reflect anything special about
displaced workers. Rather, it may simply be that individuals who find
themselves in need of a job at a time when aggregate conditions are bad
experience substantial long-term earnings losses relative to what the
same individuals would experience under better conditions.
EARNINGS LOSSES FOR DISPLACED WORKERS: MODELS Having documented a
new fact, the authors next assess the extent to which existing models of
labor market dynamics can account for it. This seems a reasonable way to
proceed, yet there is a sense in which the authors are getting a little
ahead of where the current literature is. Although their main new
empirical finding is about the cyclical variation in earnings losses for
displaced workers, the fundamental phenomenon of interest is not
cyclical in nature. That is, even during periods in which economic
aggregates are stable, some long-tenured workers are displaced and
suffer large and persistent losses in earnings. Even if cyclical
variation in earnings losses were the ultimate issue of interest, a
natural first test would be to assess whether existing models are able
to empirically account for the key features of displacement during
stable periods. If they are, one would then proceed to ask whether they
can also account for the cyclical variation in earnings losses. However,
the main conclusion from this part of the paper is that existing models
fail the first test. In other words, some benchmark models of
unemployment flows are unable to account not only for the authors'
new finding but also for key findings of the preexisting literature.
Before commenting on the exact exercises that the authors go
through in this section, I want to take a step back and offer a somewhat
broader view. In thinking about what types of models offer promise for a
better understanding of earnings losses for displaced workers, I think
it is important to view the empirical papers on this topic as a
subliterature within the broader literature on wage dynamics. One
prominent strand of this literature, including, for example, papers by
David Card (1994), Martin Floden and Jesper Linde (2001), and Eric
French (2005), uses panel data to estimate statistical models of wage
dynamics of the following general form:
log [w.sub.it+1] = [bar.[w.sub.i]], + [beta][X.sub.it+1] +
log[z.sub.it+1] + [[epsilon].sub.it+1],
where [[bar.w].sub.i] is an individual fixed effect, [X.sub.it+1]
is a vector of observable individual characteristics,
[[epsilon].sub.it+1] is a random disturbance (possibly measurement
error), and [Z.sub.it+1] is a persistent idiosyncratic shock that
evolves according to log [Z.sub.it+1] = [rho] log [Z.sub.it] +
[[eta].sub.it+1], where [[eta].sub.it+1] is another random disturbance,
distributed normally with mean zero and standard deviation
[[sigma].sub.[eta]]. (2)
Although estimates vary among papers in this literature, there is a
clear consensus that p is close to 1 and that the variance of [eta] is
substantial. (3) For example, Floden and Linde (2001) estimate that
[rho] = 0.914 and [[sigma].sub.[eta]] = 0.206. (4) To fix ideas, I adopt
these estimates for the [z.sub.it], process and consider a population of
individuals in which everyone is otherwise identical, all of the
coefficients in [beta] are zero, and the variance of [[epsilon].sub.it],
is also zero; that is, I assume that [z.sub.it] is the sole source of
earnings dynamics.
To analyze how the literature on earnings losses associated with
displacement fits within this framework, I simulate outcomes for a
sample of 10,000 workers for 40 years and ask the following question:
during the initial 25 years, what fraction of individuals experience a
wage decrease of 25 log points that persists for at least 15 years? The
answer is 36 percent, or a bit over 1 percent per year on average. If
instead one looks for changes that persist for at least 20 years
(implying that the focus is now on the initial 20 years), the answer is
22 percent, smaller than the previous number but still a bit larger than
1 percent per year.
These outcomes mimic the kinds of earnings losses that the authors
document. But what is noteworthy about this statistical model of wages
is that, by symmetry, one will also find that a similar fraction of
workers experience a wage increase of 25 log points that persists for at
least 15 (or 20) years. A simple but critical message from this exercise
is that large and persistent shocks to wages, both positive and
negative, are common. Displacement is just one instance of sudden,
large, and persistent negative changes, albeit an important one. Put
somewhat differently, I think the key to understanding earnings losses
for displaced workers in particular is to understand idiosyncratic wage
dynamics more generally.
Before proceeding further, it will be useful to think about what
the wage shocks in these statistical earnings models represent. There
are two key issues. One concerns whether these idiosyncratic shocks are
really just proxies for unmeasured heterogeneity. Abstracting from this
possibility, the second issue concerns the extent to which these shocks
reflect changes in the marginal value product of individuals as opposed
to changes in wages holding marginal value product fixed. Sorting these
issues out is beyond the scope of this comment, and I will proceed under
the assumption that at least a substantial part of these wage shocks
reflects changes in wages, holding marginal value products fixed. In
what follows I will refer to these shocks as "luck shocks." In
the specific case of the earnings losses that the authors measure, I
think this interpretation seems reasonable--it is hard to tell a story
in which a large group of workers displaced from a given firm experience
a negative shock to the true value of their productivity relative to the
workers who were not displaced and remained at the firm.
The key point is that the stand that one takes on how to interpret
the shocks has implications for what types of models one pursues and the
corresponding issues involved. If one interprets them as shocks to the
marginal value product of workers, then there is not much of a challenge
theoretically. Instead, the key challenge is to document that workers
are truly hit with large, persistent shocks to their productivity. Lars
Ljungqvist and Thomas Sargent (1998) provide one example of a model that
stresses idiosyncratic shocks to productivity and can presumably produce
outcomes that qualitatively resemble displacement. However, they do not
provide any direct evidence on the shocks.
Alternatively, if one interprets the shocks in wage equations as
primarily reflecting "luck shocks," the key challenge for
modeling is to understand why luck plays such a large and persistent
role in wage determination at the individual level. Obviously, a model
in which workers are always paid the value of their marginal product
will not suffice. A key recent paper in this regard is that by Andreas
Hornstein, Per Krusell, and Giovanni Violante (2011), who study the
ability of a wide variety of search models to generate substantial
variation in wages for identical workers in steady state. Although
search is not a necessary ingredient for a theory of wage dispersion, it
is a natural candidate to consider, and there is a long tradition of
viewing search and wage dispersion as being intimately connected.
With this information as background, let me now comment on the
specific exercises that the authors undertake. Because simple search
models in the tradition of Dale Mortensen and Christopher Pissarides
have become the leading framework for modeling the flows of workers
between employment and unemployment, and displacement is a separation
between a worker and a firm, it is tempting to think that these simple
models are a good starting point for thinking about the earnings losses
associated with displacement. However tempting this may be, it turns out
to be a poor choice of starting point. I argued previously that the key
to understanding the earnings losses of displaced workers is a theory of
wage dispersion. But in the steady state of the simplest version of
Mortensen-Pissarides models, such as the specification used by Robert
Shimer (2005), all workers earn the same wage, and the only source of
earnings dynamics is the flow of workers into and out of unemployment.
That is, there are no individual wage dynamics in steady state. As such,
this type of model is a clear nonstarter for thinking about why
identical workers can be paid so differently. Put somewhat differently,
although the simplest version of the Mortensen-Pissarides model is
useful for thinking about the forces that shape the flows into and out
of unemployment, it completely abstracts from the issue of wage
dispersion. The richer specification of Mortensen and Pissarides (1994)
does include idiosyncratic shocks to match productivity and so does
generate some wage dispersion for identical workers in steady state. But
as Hornstein and others (2011) show, the extent of dispersion is
minimal.
Hornstein and others (2011) argue that from the perspective of
generating wage dispersion for identical workers, the most promising
search models are those that feature on-the-job search. The model by
Simon Burgess and Helene Turon (2010) that Davis and von Wachter study
does include on-the-job search and thus is a reasonable starting point.
In fact, they show that this model can generate more substantial
earnings losses for displaced workers than the earlier models, although
far less than what is found in the data. In terms of illustrating the
underlying economics, I think a preferable benchmark would be the
somewhat simpler and more transparent job ladder model of Kenneth
Burdett and Mortensen (1998). (5)
The Burdett-Mortensen search model features identical workers,
identical firms, exogenous layoffs, and on-the-job search. In the
steady-state equilibrium, firms pay different wages and have different
employment levels, with high-wage firms being larger. The model
generates simple yet interesting wage dynamics: an unemployed worker
accepts the first job offered, and after that accepts any job offer that
pays a higher wage. In this sense a worker moves up the job ladder over
time. However, because of layoff shocks, workers face a risk of moving
back into the unemployment pool and needing to start the ladder over
again. Qualitatively, this model seems promising. Getting an offer from
a high-wage firm is a persistent positive "luck shock,"
whereas being laid off is a persistent negative "luck shock,"
the size of which depends on the worker's wage at the time of
layoff. There is a strong connection between the incidence of positive
and negative "luck shocks"--the workers who experience the
largest negative shocks are exactly those who have previously
experienced the largest positive shocks.
Although I think the Burdett-Mortensen model is useful for
illustrating some key ideas, it will not be able to generate the
persistence of the losses that Davis and von Wachter find in the data,
since laid-off workers in this model will move up the earnings
distribution just as do workers who enter the labor force for the first
time. Put somewhat differently, although I think wage dispersion is
intimately related to the earnings losses of displaced workers, a model
with sufficient wage dispersion alone is not enough to generate the
kinds of persistent losses found by Davis and von Wachter.
Hornstein and others (2011) argue that empirically reasonable
versions of this job ladder model can generate roughly an order of
magnitude more wage dispersion than the standard
Mortensen-Pissarides-style models. Although this is still significantly
less dispersion than is found in the data, the calculations in Hornstein
and others (2011) suggest substantially more wage dispersion than the
Davis-von Wachter calibration of the Burgess-Turon model. In particular,
Davis and von Wachter report that the ratio of the maximum to the
minimum wage in their calibrated model is about 1.5, whereas Hornstein
and others report that one could justify a calibration of a job ladder
model in the spirit of Burdett and Mortensen that can generate a ratio
of more than 1.5 for the mean wage relative to the minimum wage.
Understanding the economics behind these differences is potentially
important. Although greater wage dispersion in this model will
presumably lead to greater earnings losses for displaced workers, it is
important to note that the calculations carried out by Hornstein and
others (2011) do not explicitly relate to the estimated earnings losses
of displaced workers.
SUMMARY Davis and von Wachter argue that a good theory of
unemployment should necessarily be consistent with the evidence on
earnings losses for displaced workers. I am sympathetic to this argument
and hope that this paper serves to motivate additional work on
developing richer models of labor market dynamics. I would stress two
points. First, in my view the most promising direction for building
models that can generate substantial earnings losses for displaced
workers is to build them in ways that generate substantial wage
differences for identical workers. Second, a substantial amount of work
remains to be done to build useful and coherent models that can account
for the joint behavior of worker flows and earnings dynamics. Although
some existing search models are promising in terms of generating wage
dispersion for identical workers, factors other than search may also
play a role. The importance of unions in the United States is dwindling,
but the loss of union jobs may well account for part of the earnings
loss for displaced workers. Rigidities in organizational pay structures
may also help explain why large gaps can emerge between individual
productivity and individual wages. Finally, the models studied in this
paper assume that workers are risk neutral. Analysis of the welfare
consequences of displacement will surely require a framework that allows
for risk-averse workers and asset accumulation.
REFERENCES FOR THE ROGERSON COMMENT
Burdett, Kenneth, and Dale Mortensen. 1998. "Wage
Differentials, Employer Size and Unemployment." International
Economic Review 39: 257-73.
Burgess, Simon, and Helene Turon. 2010. "Worker Flows, Job
Flows and Unemployment in a Matching Model." European Economic
Review 54: 393-408.
Card, David. 1994. "Intertemporal Labor Supply: An
Assessment." In Advances in Econometrics, edited by Christopher
Sims. Cambridge University Press.
Floden, Martin, and Jesper Linde. 2001. "Idiosyncratic Risk in
the United States and Sweden: Is There a Role for Government
Insurance?" Review of Economic Dynamics 4: 406-37.
French, Eric 2005. "The Effect of Health, Wealth and Wages on
Labour Supply and Retirement Behavior." Review of Economic Studies
72: 395-427.
Hansen, Gary, and Ayde Imrohoroglu. 1992. "The Role of
Unemployment Insurance in an Economy with Liquidity Constraints and
Moral Hazard." Journal of Political Economy 100:118-42.
Hornstein, Andreas, Per Krusell, and Giovanni Violante. 2011.
"Frictional Wage Dispersion in Search Models: A Quantitative
Assessment." American Economic Review 101: 2873-98.
Ljungqvist, Lars, and Thomas Sargent. 1998. "The European
Unemployment Dilemma." Journal of Political Economy 106:514-50.
Mortensen, Dale, and Christopher Pissarides. 1994. "Job
Creation and Job Destruction in the Theory of Unemployment." Review
of Economic Studies 61: 397-415.
Rogerson, Richard, and Martin Schindler. 2002. "The Welfare
Cost of Worker Displacement." Journal of Monetary Economics 49:
1213-34.
Shimer, Robert. 2005. "The Cyclical Behavior of Equilibrium
Unemployment and Vacancies." American Economic Review 95: 25-49.
(1.) Using quantitative models, Hansen and Imrohoroglu (1992) show
that "typical" unemployment spells have relatively small
welfare consequences at the individual level, whereas Rogerson and
Schindler (2002) show that displacement of older workers is much more
costly at the individual level.
(2.) One limitation of this simple statistical model of earnings
dynamics is that it abstracts from the role that worker mobility plays
in the process. The phenomenon of displacement that Davis and von
Wachter stress is obviously about a strong connection between certain
types of worker turnover and earnings shocks. Other papers in the
literature have expanded these models to incorporate mobility, but more
work is clearly needed.
(3.) Related to an issue that was raised in the previous section,
this literature has studied both earnings and wage dynamics. The main
message is that they display similar properties, in that they also
exhibit large and persistent idiosyncratic fluctuations. Put somewhat
differently, unemployment dynamics account for a very small portion of
idiosyncratic changes in earnings.
(4.) Although Floden and Linde include individual fixed effects in
their specification, in estimation they assume that individual fixed
effects are captured by observables and hence are subsumed into X.
(5.) See Hornstein and others (2011) for a broader discussion and
citations of many other related papers.
GENERAL DISCUSSION Edward Lazear expressed surprise at the result
that individuals experience larger losses following job loss during a
recession than during an expansion. In theory, he argued, an
idiosyncratic job loss during an expansion should send a more negative
signal to employers about that worker than the same job loss during a
recession, when layoffs are prompted by worsening macroeconomic
conditions. That idea led Lazear to wonder whether individuals who lose
their job during a recession spend more time out of work than those laid
off during an expansion. Such a difference might indicate that greater
skill depreciation in the former group is what causes their greater
earnings losses.
Lazear then returned to a puzzle brought up by the discussants: if
the potential lifetime earnings losses from a recession layoff are so
large, why don't these individuals try to make up for some of these
losses by investing more in their human capital? He suggested that the
lack of evidence of such investment indicated some type of selection
effect at work.
Till von Wachter built on Lazear's last comment, arguing that
if selection effects were an important source of earnings losses, one
would expect that workers laid off idiosyncratically during expansions
should experience larger losses than those laid off during recessions.
He then reported results of another test of selection effects he had
conducted, comparing the earnings losses of workers laid off from firms
that eliminated a large fraction of their payrolls with those of other
laid-off workers. If selection effects were important, the first group
should have experienced smaller earnings losses than the second, more of
whom were likely to have been laid off idiosyncratically. In the data,
however, this turned out not to be the case.
Von Wachter saw this lack of evidence for selection effects as
pointing toward a different theory in which cyclical pressures cause an
economywide drop in starting wages; when laid-off workers return to
work, their wages are recalibrated to this lower level, and these lower
wages "stick" to the worker for a long time. In this theory,
all workers hired during a recession, whether they were laid off earlier
in the recession or not, would experience the persistent negative
impacts of starting work during a period with low average wages.
Justin Wolfers was struck by the correlation between the present
value of job loss and the unemployment rate. He pointed out that
unemployment duration is also correlated with the unemployment rate, and
he wondered whether the unemployment rate or the unemployment duration
structure was the more important determinant of the present value of job
loss. He suggested that an interesting experiment would be to compare
the present value of job loss during the recent recession with that in
the 1980s recession, since the latter period exhibited a larger
unemployment rate shock but a smaller shock to the duration structure.
Christopher Carroll viewed the result that the size and persistence
of earnings losses varied across the business cycle as an important
contribution to the literature, in part because it could help explain
the dynamics of the business cycle. The risk of large, persistent
earnings losses amounts to greater uncertainty about future earnings,
which might lead individuals to increasing their saving rates, which, in
turn, could lead to a shortfall in aggregate demand.
Betsey Stevenson noted that the human capital loss resulting from a
layoff could fall into either of two categories: workers might have
built up firm-specific capital that they cannot find another employer to
make use of, or their general skills might deteriorate quickly during
periods of unemployment. The distinction, she argued, was important for
policy. If firm-specific skills are the major issue, an appropriate
policy response might be to subsidize firms to keep workers employed
through downturns. But if atrophy of general skills is the greater
concern, a better response might be job training programs aimed at
mitigating those skill losses. Responding to Wolfers, Stevenson
cautioned against simple comparisons of workers' earnings losses
across recessions, since the age structure of the workforce, and thus
the average tenure of workers, have changed over time.
John Haltiwanger said that in data he had examined, the duration of
joblessness was a strong determinant of wages upon reemployment:
separated workers who remained jobless for more than a short while
experienced much larger wage losses than those who were reemployed
quickly. He suggested that the large wage cuts that the long-term
unemployed often accept upon reemployment could help explain the higher
average persistent earnings losses observed among workers displaced
during recessions. Von Wachter countered that he had conducted a similar
analysis with German data and did not find large earnings losses to be
associated with time spent out of work. Results from separate work with
Jae Song and Joyce Manchester using U.S. data suggested that an
important predictor of earnings losses following joblessness was
switching industries.
Robert Gordon wished the paper had distinguished more carefully
between firm-specific human capital and pure economic rent. A piece of
evidence in support of the idea that displaced workers suffered
significant rent losses, as opposed to loss of firm-specific skills, was
the fact that many workers losing mid-level jobs experience earnings
losses that far exceed the cost of retraining for a similar job. This
suggests that, before being laid off, these workers were being paid much
more than their marginal product. A vivid example of this phenomenon is
the two-tier wage system at General Motors and Chrysler, in which
high-tenured employees are paid double the amount that more recent hires
are paid for the same position. Davis responded that the empirical part
of the paper was agnostic on the issue of whether rent or firm-specific
human capital was a greater source of earnings losses for displaced
workers. This agnosticism was deliberate, since he and von Wachter
thought the data they examined did not allow them to distinguish between
different sources.
Jesse Rothstein expressed skepticism that firm-specific human
capital losses could account for anywhere near the magnitude and
persistence of earnings losses among those laid off with less than five
years of tenure. Why wouldn't these workers rebuild this human
capital over the course of several years upon being reemployed?
Von Wachter responded to a query about the average tenure of
laid-off workers by explaining that missing data in the early years of
the sample made it impossible to compute an overall average. He and
Davis had, however, compared workers with different predisplacement job
tenure who were displaced during different recessions and found that
these workers experienced similar earnings losses. In other words,
displaced workers' earnings losses appear to be larger in
recessions even for shorter-tenured workers.
Stephanie Aaronson remarked that much of the discussion had focused
on the possibility of within-firm selection of individuals for layoff
according to varying levels of employee rents. She suggested that an
alternative type of selection, at the firm level, could be driving the
results. Perhaps the division of rents between workers and capital
varies at the firm level, for structural reasons such as varying
contracting schemes. Then an economy-wide shift of rents from workers to
capital during a recession might explain some of the persistent earnings
losses that workers experience.
David Romer observed that the labor literature had moved away from
the idea that certain aspects of hiring practices, not involving
search-and-matching considerations, caused some jobs to pay large,
persistent rents, which a worker would lose when laid off. He suggested
that one could view this paper as a test of the idea that such rents are
significant. He also thought it might be time for researchers to revisit
the classic literature on interindustry wage differentials and
efficiency wages, in which large, persistent rents are possible. Gregory
Mankiw agreed with Romer, adding that it might be worthwhile to add a
section to the paper looking at the results from an efficiency wages
perspective, to see if it provided a more interesting way of
understanding the data than the currently popular Mortensen-Pissarides
model.
Steven Davis pointed to an aspect of the data presented by Robert
Hall that confounded simple theories of rents or selection to explain
persistent earnings losses: for seven consecutive years before the
mass-layoff event, the wages of those who ended up being laid off during
the recession were virtually the same as those who were not. Von Wachter
added that in separate work with Song and Manchester he had tried to
predict workers' wage losses based on their moves from high-wage
industries to low-wage industries and found that the fraction of total
wage losses that they explained was surprisingly low. Losing a job in
any industry during a major recession led to large earnings losses.
Robert Shiller suggested that the results could be explained by
workers' choices, following a layoff, to move from unpleasant jobs,
such as those requiring long hours, to more pleasant ones. Von Wachter
agreed that it was reasonable to question the extent to which pure
earnings losses translated into utility losses for individuals. To
investigate this question, he, Song, and Manchester had examined a range
of outcomes, including job stability, health, and children's
schooling outcomes, and found that all of these outcomes worsened
following job loss.
Table 1. Present-Value Earnings Losses after Mass-Layoff Events,
Men 50 or Younger with at Least 3 Years Prior Job Tenure,
1980-2005 (a)
PDV of average loss at
displacement
As a mul-
% of all tiple of As % of
years predis- PDV of
from placement counter-
1980 annual factual
Subgroup (b) to 2005 Dollars earnings earnings (c)
All 100 77,557 1.71 11.9
Displaced in 88 72,487 1.59 11.0
expansion year
Displaced in 12 109,567 2.50 18.6
recession year
Displaced in year
with unemployment
rate:
<5.0% 23 50,953 1.06 9.9
5.0-5.9% 35 71,460 1.56 10.9
6.0-6.9% 13 71,006 1.58 10.7
7.0-7.9% 21 89,792 2.07 14.4
[greater than or 8 121,982 2.82 19.8
equal to] 8.0%
Source: Authors' calculations using equation 2 and estimates from
equation 1.
(a.) PDVs are calculated over 20 years of job displacement at an
annual discount rate of 5 percent. Mass-layoff events are defined
as in section I. See text for further description. Dollar figures
are in dollars of 2000.
(b.) When a year contains both expansion and recession months or
monthly unemployment rates that fall in different ranges, that
year's values are allocated proportionally to the number of months
in each cyclical state or range.
(c.) Counterfactual earnings are what the displaced worker would
have earned over the same 20 years had he not been displaced.
Table 2. Present-Value Earnings Losses after Mass-Layoff Events,
Various Groups, 1980-2005 (a)
PDV of average loss at displacement
As a mul-
tiple of As % of
predisplace- PDV of
ment annual counfactual
Group (b) Dollars earnings earnings (c)
Women 21-50, 3 or more years
tenure
All years 38,033 1.5 10.9
Expansion years only 33,164 1.3 9.5
Recession years only 68,782 3.3 20.6
Men 21-50, 6 or more years
tenure
All years 106,900 2.0 12.9
Expansion years only 100,543 1.8 11.9
Recession years only 148,400 3.0 20.0
Men 21-30, 3 or more years
tenure
All years 50,240 2.1 9.8
Expansion years only 39,639 1.7 7.8
Recession years only 117,322 4.0 22.0
Men 31-40, 3 or more years
tenure
All years 49,599 1.2 7.7
Expansion years only 42,555 1.0 6.5
Recession years only 93,833 2.2 16.0
Men 41-50, 3 or more years
tenure (d)
All years 98,519 1.8 15.9
Expansion years only 95,716 1.7 15.1
Recession years only 116,515 2.2 21.9
Men 51-60, 3 or more years
tenure (e)
All years 99,288 1.8 24.0
Expansion years only 97,934 1.7 23.1
Recession years only 108,248 2.1 31.1
Source: Authors' calculations using equation 2 and estimates from
equation 1.
(a.) PDVs are calculated over the 20 years following displacement
as described in table I, except as noted below. Dollar figures are
in dollars of 2000.
(b.) Ages and years of tenure are as of time of displacement.
Values for years containing both expansion and recession months or
monthly unemployment rates that fall in different ranges are
calculated as described in table 1.
(c.) Counterfactual earnings are what the displaced worker would
have earned over the same 20 years had he or she not been
displaced.
(d.) PDVs are calculated over 15 years.
(e.) PDVs are calculated over 10 years.
Table 3. Percent of Employed Adults Who Worry They Will Experience an
Adverse Job-Related Event in the Near Future
Adverse event (a)
Cut in Cut in Cut in
Survey month hours it ages benefits Layoff
August 1997 15 17 34 20
August 2003 15 17 31 19
August 2004 14 17 28 20
August 2005 13 14 28 15
August 2006 16 19 30 17
August 2007 12 14 29 14
August 2008 14 16 27 15
August 2009 27 32 46 31
August 2010 25 26 39 26
August 2011 30 33 44 30
Source: Reproduced from Gallup Polling data at
www.gallup.com/poll/1720/ work-work-place.aspx and
www.gallup.com/poll/149261/
Worries-Job-Cutbacks-Return-Record-Highs.aspx.
(a.) Based on polling of workers employed full or part time. The
survey question is "Next, please indicate whether you are worried
or not worried about each of the following happening to you,
personally, in the near future. How about Jthe following are
rotated) that your hours at work will be cut back'? that your wages
will be reduced'? that your benefits will be reduced'? that you
will be laid off?"
Table 4. Present-Value Income and Earnings Losses Associated with
Job Loss in the Basic Mortensen-Pissarides Models (a)
Percent
Basic MP model version
Nash version,
Nash version, Hagedorn and
standard Manovskii (2008)
calibration (b) calibration
Range of mean PDV 0.20 to 0.22 0.044 to 0.047
income losses over five
aggregate states (d)
Simulation outcomes (e)
All aggregate paths
Mean unemployment rate 6.6 6.7
Monthly job finding rate (f) 43 43
Mean PDV income loss (g) 0.23 0.05
10th-90th percentile -0.55 to 1.07 -0.29 to 0.40
range, income losses
Mean PDV earnings loss (h)
10th-90th percentile
range, earnings losses
Aggregate boom paths (i)
Mean unemployment rate 6.5 6.4
Monthly job finding rate (f) 43 44
Mean income loss (g) -0.19 -0.26
10th-90th percentile -0.84 to 0.56 -0.39 to -0.11
range, income losses
Mean PDV earnings loss (h)
10th-90th percentile
range, earnings losses
Aggregate bust paths (j)
Mean unemployment rate 6.7 7.0
Monthly job finding rate (f) 43 41
Mean income loss (g) 0.66 0.37
10th-90th percentile 0.02 to 1.38 0.26 to 0.51
range, income losses
99th-percentile 2.18 0.66
income loss
Mean PDV earnings loss (h)
10th-90th percentile
range, earnings losses
99th-percentile
earnings loss
Basic MP model version
Credible bargaining
version, Hall and
Milgrom (2008)
calibration (c)
Range of mean PDV 0.20 to 0.23
income losses over five
aggregate states (d)
Simulation outcomes (e)
All aggregate paths
Mean unemployment rate 6.7
Monthly job finding rate (f) 43
Mean PDV income loss (g) 0.23
10th-90th percentile -0.51 to 1.04
range, income losses
Mean PDV earnings loss (h) 1.28
10th-90th percentile -2.62 to 5.72
range, earnings losses
Aggregate boom paths (i)
Mean unemployment rate 6.4
Monthly job finding rate (f) 44
Mean income loss (g) -0.12
10th-90th percentile -0.75 to 0.60
range, income losses
Mean PDV earnings loss (h) 1.14
10th-90th percentile -2.73 to 5.53
range, earnings losses
Aggregate bust paths (j)
Mean unemployment rate 7.0
Monthly job finding rate (f) 42
Mean income loss (g) 0.59
10th-90th percentile -0.08 to 1.35
range, income losses
99th-percentile 2.20
income loss
Mean PDV earnings loss (h) 1.42
10th-90th percentile -2.49 to 5.87
range, earnings losses
99th-percentile 10.81
earnings loss
Notes to table 4:
Source: Authors" calculations.
(a.) All present-value calculations use a 5 percent annual discount
rate. All model calibrations follow Hall and Milgrom (2008) in
their choice of parameter values and the transition matrix of a
five-state Markov process for aggregate shocks. See the online
appendix for a more detailed description of the simulations and
calculations.
(b.) Calibration is similar to that in Shimer (2005) and Hall
(2005).
(c.) Model entails sequential bargaining with disagreement costs a
la Binmore, Rubinstein, and Wolinsky (1986). Calibration is that of
Hall and Milgrom (2008).
(d.) We compute the present-value income losses in the top row
directly from value functions. For each aggregate state, we
calculate the difference between the asset value of employment and
the asset value of unemployment and express the difference relative
to the asset value of employment. Performing this calculation for
the five aggregate states yields the reported ranges.
(e.) Each indicated model is simulated for 1,000 draws of the
aggregate path, with each draw starting from the middle aggregate
state and evolving according to the aggregate transition matrix.
Each draw is simulated for 5,000 working days (20 years at 250
working days per year). The realized paths are tracked for 5,000
day--1 job losers and 1,000 day--1 employed persons on each of the
1,000 aggregate paths.
(f.) Calculated as [theta][[SIGMA].sup.25.sub.i=1]
[(1 - [theta]).sup.i-1], where [theta] is the daily job finding rate,
assuming 25 job seeking days per month.
(g.) For the income calculations, an individual receives the
imputed income value of leisure if unemployed on a given day, and
the annuity value of the wage bargain if employed. At the end of
the simulation horizon, each individual is assigned the asset value
associated with that individual's state on day 5,000. This results
in a realized income path plus terminal value for each individual,
which is then used to compute the realized PDV of income for an
unemployed worker as of day 1. This quantity is then compared with
that of the mean realized present-value income of the day--1
employed persons on the same aggregate path.
(h.) For the earnings calculations, each individual is assigned
zero earnings if unemployed and the annuity value of the wage
bargain if employed. To focus on PDV earnings over a 20-year
horizon comparable to the empirical estimates in section Il, the
terminal value is set to zero at the end of the 5,000-day
simulation horizon. The PDVs of the realized earnings paths for
individuals who become unemployed on day 1 are then compared with
the mean realized present-value earnings for 1,000 individuals who
remain employed on day 1 on the same aggregate path. Because
earnings loss statistics are very similar across all three variants
of the MP model, results are reported only for the credible
bargaining version of the basic MP model.
(i.) The 1,000 aggregate paths are ranked by realized mean PDV
income or earnings loss. This panel reports statistics for the
paths ranked from 90 to 110 (the 20 paths nearest the 10th
percentile) by this metric.
(j.) Statistics are reported for the paths ranked from 890 to 910
(the 20 paths nearest the 90th percentile) by mean PDV income or
earnings loss.
Table 5. Present-Value Income and Earnings Losses Due to Job Loss in
the Burgess-Turon Model (a)
Percent
Aggregate state (b)
Good Middle
Mean PDV of loss due to
idiosyncratic shocks resulting
in job loss
Income (c) (percent of 0.39 0.35
employment asset value)
Earnings (d) (percent of PDV of 2.44 2.54
counterfactual earnings over
20 years)
Job finding rate
Quarterly 82.5 73.7
Monthly 44.1 35.9
Aggregate state transition
Good [right Middle [right
arrow] middle arrow] bad
PDV income losses (c) (percent of
employment asset value)
Mean loss due to idiosyncratic 0.63 0.57
shocks that result in job loss,
comparison with own past (c)
Mean loss due to aggregate 0.25 0.22
shock that results in job
loss, comparison with own past
Inflow-weighted average (f) 0.61 0.55
Mean loss due to idiosyncratic 0.35 0.32
shocks that result in job
loss, comparison with
control group (g)
Mean loss due to aggregate 0 0
shock that results in job
loss, comparison with
control group (h)
Inflow-weighted average 0.33 0.30
PDV earnings losses (d) (percent of
PDV of countrerfactual earnings
over 20 years)
Mean loss due to idiosyncratic 2.85 3.08
shocks that result in job loss,
comparison with own past
Mean loss due to aggregate 2.15 2.57
shock that results in job loss,
comparison with own past
Inflow-weighted average 2.81 3.05
Mean loss due to idiosyncratic 2.54 2.71
shocks that result in job loss,
comparison with control group
Mean loss due to aggregate shock 0 0
that results in job loss,
comparison with control group
Inflow-weighted average 2.39 2.55
Aggregate state (b)
Bad
Mean PDV of loss due to
idiosyncratic shocks resulting
in job loss
Income (c) (percent of 0.32
employment asset value)
Earnings (d) (percent of PDV of 2.71
counterfactual earnings over
20 years)
Job finding rate
Quarterly 64.9
Monthly 29.5
Aggregate state
transition
Good [right
arrow] bad
PDV income losses (c) (percent of
employment asset value)
Mean loss due to idiosyncratic 0.84
shocks that result in job loss,
comparison with own past (c)
Mean loss due to aggregate 0.47
shock that results in job
loss, comparison with own past
Inflow-weighted average (f) 0.80
Mean loss due to idiosyncratic 0.32
shocks that result in job
loss, comparison with
control group (g)
Mean loss due to aggregate 0
shock that results in job
loss, comparison with
control group (h)
Inflow-weighted average 0.29
PDV earnings losses (d) (percent of
PDV of countrerfactual earnings
over 20 years)
Mean loss due to idiosyncratic 3.26
shocks that result in job loss,
comparison with own past
Mean loss due to aggregate 2.57
shock that results in job loss,
comparison with own past
Inflow-weighted average 3.19
Mean loss due to idiosyncratic 2.71
shocks that result in job loss,
comparison with control group
Mean loss due to aggregate shock 0
that results in job loss,
comparison with control group
Inflow-weighted average 2.42
Source: Authors' calculations.
(a.) Burgess and Turon's (2010) search-and-matching model differs
from the basic MP model in capturing search on the job, a
distinction between job flows and worker flows, heterogeneity in
wages and match surplus values, and spikes in aggregate job
destruction. It also adopts a different vacancy creation process
that gives content to the concept of a job apart from the
employer-worker match. Job destruction and job loss arise from
negative aggregate shocks and sufficiently bad idiosyncratic
shocks. We depart from Burgess and Turon's calibration (which was
designed to match features of the U.K. economy) by increasing the
arrival rate of idiosyncratic shocks (from 0.15 to 0.25) and the
efficiency of the matching function (from 0.6 to 1.1). These
changes yield more rapid flows through the unemployment pool and
higher monthly job finding rates, roughly in line with U.S.
outcomes. The unemployment rate is 5.2 percent in the middle state
in our calibration. See the text for further description of the
model and the online appendix for a detailed explanation of the
loss calculations and the underlying simulations.
(b.) Results are for a period of time corresponding to 3 months
with no change in the aggregate state.
c. Calculated from value function comparisons.
(d.) Calculations rely on simulations of aggregate and individual
paths over 20-year horizons (80 quarters), where earnings are set
to the wage if the individual is employed and to zero if not. The
wage when employed depends on the aggregate state and the
idiosyncratic productivity level of the job.
(e.) "Own past" comparisons calculate losses relative to the job
loser's predisplacement employment value evaluated at the old
aggregate state and expressed relative to the same employment
value. The value of unemployment is calculated at the new aggregate
state.
(f.) Inflow-weighted averages of PDV losses associated with
idiosyncratic and aggregate shocks. The weights are given by the
share of job loss due to idiosyncratic shocks during the quarter
and the share triggered by a negative aggregate shock.
(g.) "Control group" comparisons calculate losses relative to the
job value evaluated at the new aggregate state. The value of
unemployment is also calculated at the new aggregate state.
(h.) Calculations that result in zero loss do so because all
workers in the lower tail of the productivity distribution lose
their jobs when hit by a negative aggregate productivity shock, and
all get the value of unemployment in the new state.