the decline of job security in the 1990s: displacement...

Federal Reserve Bank of Chicago 17

Daniel Aaronson and Daniel G. Sullivan

Daniel Aaronson is an economist and Daniel G. Sullivan isa senior economist and vice president at the Federal ReserveBank of Chicago. The authors would like to thank Ann Ferrisfor her very capable assistance.

Introduction and summary

The news media frequently suggest that Americanworkers have suffered a significant decline in jobsecurity during the 1990s. Of course, plannedand actual employment reductions at major cor-porations such as AT&T, IBM, and General Motorshave been important stories. But articles suchas those in the 1996 New York Times book TheDownsizing of America go beyond reporting indi-vidual cases of layoffs to suggest that there hasbeen a fundamental change in the employmentrelationship. According to such articles, workersin general have suffered a loss of job security,and long-term employment relationships are athing of the past. Moreover, such articles claimthat this decreased job security has left workersfeeling more anxious about their futures.

The perception of declining job security isshared by many policymakers and other analysts,who believe worker anxiety to be a major reasonwage inflation in the 1990s has remained modestin the face of historically low levels of unemploy-ment. Perhaps most famously, Federal ReserveBoard Chairman Alan Greenspan testified toCongress in February 1997 that “atypical restrainton compensation increases has been evident fora few years now and appears to be mainly theconsequence of greater worker insecurity.”1

Former U.S. Labor Secretary Robert Reich recentlymade much the same point when he wrote, “Wagesare stuck because people are afraid to ask for araise. They are afraid they may lose their job.”2

Labor economists, however, have often beenskeptical of claims of widespread declines in jobstability and security. They note that media accountsare long on anecdotes and short on evidencebased on nationally representative survey data.Moreover, the most carefully executed studies

using scientifically designed survey data collectedthrough the early 1990s often reached conclusionsquite at odds with media reports. For instance,Diebold, Neumark, and Polsky (1997) concludedthat “aggregate job retention rates have remainedstable.” Similarly, Farber (1998) found that “therehas been no systematic change in the overall distri-bution of job duration over the last two decades.”

More recently, however, researchers havebegun to analyze survey data from the mid-1990s,and conclusions somewhat more in line withmedia reports are emerging. For instance,Neumark, Polsky, and Hansen (1997) reportedthat “there is some evidence that job stabilitydeclined modestly in the first half of the 1990s.Moreover, the relatively small aggregate changesmask rather sharp declines in stability for workerswith more than a few years of tenure.” Similarly,Farber (1997b) concluded that “after controllingfor demographic characteristics, the fraction ofworkers reporting more than ten and more than20 years of tenure fell substantially after 1993 toits lowest level since 1979.” Thus job stability—the tendency of workers and employers to formlong-term bonds—seems to be declining some-what. Moreover, evidence of significant changeis especially apparent in more direct measuresof worker security, such as Farber’s (1997a) tabu-lations of the number of workers reporting invol-untary job loss. The extent of changes in job tenure,turnover, and displacement reported in thesemore recent studies is much too modest to justify

The decline of job security in the 1990s:Displacement, anxiety, and theireffect on wage growth

Economic Perspectives18

the most sensationalistic news reports. Neverthe-less, some decline in job security, especially forworkers who have attained significant seniority,now seems reasonably clear.

In this article, we review some of the findingsof this research on job stability and job security.We then present some new tabulations of ratesof job loss for high seniority workers based onthe Bureau of Labor Statistics’ (BLS) DisplacedWorker Surveys (DWS). Next, we look directly atworkers’ own perceptions of their job securityusing data from the National Opinion ResearchCenter’s General Social Survey (GSS). Finally, weattempt to relate our measures of displacementand worker anxiety to wage growth by examiningtime-series data for the nine U.S. census divisions.

Our tabulations of annual displacement ratesfrom the DWS focus on workers with five or moreyears of tenure. We find that among such workers,job loss due to “shift or position abolished,” whichamong the surveys’ possible reasons for job losscomes closest to capturing the notion of “down-sizing,” increased quite dramatically from annualrates of two tenths or three tenths of a percentthroughout the 1980s to a range of six tenths orseven tenths of a percent in the mid-1990s. Deter-mining the trend in displacement more generallyis complicated by changes in the DWS. However,our preferred estimates suggest that overall dis-placement rates were higher in 1995, the mostrecent year for which we have data, than at anytime since the data began in 1979. We estimate a1995 displacement rate of about 3.4 percent forworkers with five or more years of tenure. By com-parison, the rate for 1982, which was in the middleof a severe recession, was only about 2.5 percent.We consider this a substantial increase in the riskof displacement for high-seniority workers.

We also find that displacement has becomesomewhat more “democratic” in the 1990s. Previ-ously, high-seniority workers who were highlyeducated, were in white-collar jobs, or were em-ployed in the service producing industries wererelatively immune to displacement. More recently,however, displacement rates for these groupshave risen especially fast, while those for somegroups who had high rates of displacement in the1980s, such as those with at most a high schooleducation, those in blue-collar occupations, andthose working in manufacturing, rose less or evenfell relative to their peaks in the early 1980s. Asa result of this increased democratization of dis-placement, many more workers may now con-sider themselves at risk for job loss.

The GSS data suggest that workers’ own per-ceptions of their job security have also declined.The fraction of workers not responding “veryunlikely” to the question, “How likely is it thatyou will lose your job in the next year?,” rosefrom about 31 percent in 1989 to about 40 percentin 1996, the most recent year for which data areavailable. The 1996 figure approximately matchesthe highest reading since this question began tobe asked in 1977. The 1996 reading is especiallyremarkable given that unemployment was gen-erally below 6 percent, while in 1982, when sucha level of anxiety was previously reached, unem-ployment was nearly 10 percent. One should not,however, exaggerate the extent to which workers’anxiety over job loss has increased. The mainchange has been an increase in the number ofworkers responding that it is “not too likely”rather than “very unlikely” that they will losetheir jobs. The percentages of workers respond-ing that it is “fairly likely” or “very likely” haverisen more modestly.

With a few exceptions, the groups of workerswho have experienced the largest increases indisplacement rates have also had the largestincreases in reported probabilities of job loss.For instance, an increase in perceived likelihoodof job loss has been especially great among white-collar workers. Perceived job security has actuallyincreased for blue-collar workers. Another inter-esting finding concerns the relationship betweenworkers’ perceptions of their job security and theuse of computers in their industry. In the early1980s, workers in industries with greater com-puter usage felt more secure on average than otherworkers. By the mid-1990s, however, the relation-ship had reversed, with workers in industrieswith greater computer usage feeling less secure.

Finally, we attempted to judge to what extentour findings of an increase in displacement ratesand workers’ perceptions of their chances ofjob loss are related to changes in aggregatewages. Standard short-run Phillips curve analy-ses such as Gordon (1997) have tended to predicthigher levels of wage inflation than have actuallyoccurred in the last two or three years. Thoughsignificant forecast errors are nothing new forsuch models,3 the importance of the question topolicymakers has led to a great deal of specula-tion as to why wage inflation has remained sub-dued. Our findings and those of other researchers,which suggest that job security has declined inrecent years, add some plausibility to the case


that worker anxiety has played a role in restrain-ing wage inflation.

However, one could easily point to otherrecent changes in labor markets or elsewhere inthe economy that might be affecting wage growth.4

Why should one particular change—that towardsreduced job security—be considered the key fac-tor? To make a more convincing case for the im-portance of worker insecurity, one would wantto observe that in the past, when displacementor anxiety was high relative to unemployment,wage inflation had also been subdued. Moreover,to obtain any sense of the quantitative importanceof worker insecurity in restraining wage inflation,one needs to examine historical evidence. Unfor-tunately, with annual measures of displacementand insecurity that go back only to the late 1970s,there is not much hope of extracting such infor-mation from the U.S. time-series data.

Our strategy is to look cross-sectionally aswell as over time and ask whether census regionsthat have had higher displacement rates or workerperceptions of insecurity have tended to havelower wage growth. Such a strategy parallelsthe “wage curve” analyses of Blanchflower andOswald (1994), although, as in Blanchard andKatz (1997), we employ the traditional Phillipscurve specification in which the change in wages,rather than their level, is related to unemploy-ment and measures of job security.

We pool separate data from the nine censusregions to estimate the effect of displacementrates or perceptions of insecurity on forecasts ofwage inflation. We find that, holding constantthe unemployment rate, higher values of bothdisplacement and worker insecurity are associ-ated with lower wage growth. However, evenwith the additional source of data variation thatcomes from pooling the separate census divisions,our estimates of the magnitude of these effectsare imprecise. Indeed, if we allow for the possi-bility that there may be some permanent, unmea-sured characteristics of regions that are associatedwith different levels of wage growth, we cannotreject the hypothesis that the true effect of jobsecurity on wages is zero and that estimates thesize of those we obtain could have arisen bychance. Nonetheless, our best estimates suggestthat increases in displacement rates and workers’own anxiety about their job security could beresponsible for restraining wage growth byabout three tenths to seven tenths of a percent-age point per year during the mid-1990s. Such

an effect would explain all or most of the puzzleof lower than expected wage inflation.

Previous research on turnoverand displacement

The New York Times describes its book ondownsizing as putting “a human face on a historicpredicament that is as ubiquitous as it is painful.”5

A large body of research demonstrates that jobloss is painful, at least for workers who have at-tained significant tenure.6 For example, Jacob-son, LaLonde, and Sullivan (1993c) found thateven six years after job loss, earnings lossesamong a sample of Pennsylvania workers dis-placed in the early 1980s were still equal to about25 percent of their predisplacement earningslevels. What has been somewhat less clear toresearchers is whether job loss has become anymore ubiquitous in recent years.

It is helpful to divide the relevant researchinto two parts—that on job stability and that onjob security. By stability we mean the tendencyfor workers and firms to develop long-termrelationships. Research on job stability questionsthe many media accounts claiming that such long-term employment relationships have gone theway of buggy whips. By security we mean workers’ability to remain in employment relationshipsas long as their own performance is satisfactory.Research on job security asks whether there hasbeen an increase in involuntary job loss due toreasons beyond workers’ control. Job stabilitydepends on workers’ own choices, in additionto the factors that influence job security. Forinstance, if a group of workers increase theircommitment to the labor force or to their partic-ular employers, then their job stability may riseeven if they are increasingly subject to threatsof displacement. As we shall see, research sug-gests a larger 1990s decline in job security thanin job stability.

In our view, trends in job security are muchmore relevant to the discussion of whether spe-cial factors might be restraining wage inflationthan are trends in job stability. Indeed, if declinesin job stability are less dramatic than declines injob security, it must largely be because workersare less likely to leave jobs voluntarily, and adecreased tendency to quit jobs may itself signalworker insecurity. Nevertheless, we begin witha short account of research on job stability.

The starting point for much of the researchon job stability is the distribution of job tenure.


full years of job tenure

0

5

10

15

0 5 10 15 20 25

percent

Most of what is known about this distributionderives from a series of supplements to the Cur-rent Population Survey (CPS).7 As an illustration,figure 1 displays the distribution of job tenurefor employed men between the ages of 35 and44. These data were collected from the most recentMobility Supplement to the CPS, which was con-ducted in February 1996. The figure shows thatthe most common tenure levels are the shortest—for example, less than one year and between oneand two years—with a roughly monotonic declinein the number of workers with successively longertenure. Nevertheless, there are many workerswith substantial levels of job tenure. For men inthe 35 to 44 age group, the median tenure is about6.1 years. Moreover, about 33 percent of suchworkers have been in their current jobs at leastten years and about 22 percent have been in theircurrent jobs at least 20 years.

Figure 2 shows how median job tenure haschanged over time for men and women in threeage groups, 25 to 34, 35 to 44, and 45 to 54. Thesedata are derived from CPS Mobility Supplementsconducted in January of 1963, 1966, 1968, 1973,1978, 1983, 1987, and 1991 and February of 1986.8

Not surprisingly, older workers typically havelonger job tenures than younger workers. Also,men typically have longer tenures than women,who are more likely to have interrupted theircareers for family reasons. However, our primaryinterest is in the aggregate trends in these data.

For men, especially those in the two highest agegroups, median job tenures declined from 1991to 1996, which is consistent with claims of decreasedjob stability. However, women’s job tenure rosefor all age groups. So, overall, there has been rel-atively little change in median job tenure duringthe 1990s. Moreover, the drop in male tenure forthe two oldest groups seems to be mainly a con-tinuation of a trend that was evident throughoutthe 1980s. Thus, it is difficult to conclude that jobstability has suffered more than a modest declinein the 1990s.9

Farber (1997b) shows that similar, thoughsomewhat more dramatic, changes took placeduring the 1990s at the high end of the tenuredistribution. In particular, he shows that thepercentages of workers reporting more thanten and 20 years of tenure declined significantlybetween 1991 and 1996. The proportion of workersaged 35 to 64 with ten or more years of tenuredeclined from 38.3 percent to 35.4 percent. Formen the decline was more dramatic, from 44.3percent to 40.0 percent, while among women,the decline was from 31.4 percent to 30.3 per-cent. Similar drops were reported for workerswith different educational levels. However, theoccupational groups that have historically hadthe highest long-term employment levels, suchas managerial, professional and technical, andblue-collar workers, had the largest declines inthe 1990s. Similarly, the declines were greatest

in industries, such as transportation,communications, and public utilities,in which long-term employment hadbeen most common. As a result, thefrequency of long-term employmentis now more similar across occupa-tions and industries.

Farber’s (1997b) results, as wellas the trends in median job tenureshown in figure 2 suggest that jobstability among men has declinedmodestly during the 1990s. For women,job stability has either declined verymodestly or continued to rise, depend-ing on whether it is measured bymedian tenure or the proportion ofworkers with high tenure levels. Forboth sexes, the changes appear toomodest and gradual to support thesensationalistic media reports pro-claiming the end of long-term em-ployment relationships.

FIGURE 1

Distribution of job tenure for men aged 35–44,February 1996

Source: Authors� calculations based on data from U.S. Department of Commerce,Bureau of the Census, Current Population Survey, Mobility Supplement, February 1996.


To reach a tenure of ten years, a worker mustsurvive from the first year of a job into the second,from the second to the third, and so on for tenyears. Thus, the distribution of job tenures and,in particular, the fraction of workers with ten ormore years of tenure can be thought of as depend-ing on a sequence of survival probabilities goingback many years. This means that if the proba-bility of remaining in a job for another year wereto have suddenly dropped sometime in the ear-ly 1990s, it would take a number of years for thischange to show up fully in the tenure distribution.In this case, results such as Farber’s (1997a) andthose shown in figure 2 might not reveal the full

extent of change. For this reason, it is of interest toexamine job survival or retention probabilities.10

Diebold, Neumark, and Polsky (1997) andNeumark, Polsky, and Hansen (1997) have care-fully analyzed the trend in retention rates fromthe mid-1980s to the mid-1990s, using data ontenure distributions from CPS supplements forevery four years from 1983 to 1995. The authorsestimate four-year retention rates—the probabil-ity that a worker will remain in a job an additionalfour years—by dividing the number of workerswith a given set of characteristics and a certainnumber of years of tenure in one survey by thenumber of workers with those characteristicsbut four fewer years of tenure in the survey fouryears earlier. Adjustments are made for a num-ber of potential problems, including nonresponseto the survey, the tendency of workers to roundtheir tenure to a multiple of five years, the differ-ing levels of unemployment at the time of thesurveys, and the special nature of the tenure dataderived from the February 1995 CPS Supplementon Contingent Work.

Table 1 contains some representative resultsfrom Neumark, Polsky, and Hansen (1997).11

Evidently, the trend over time in retention ratesdepends to a great extent on workers’ initial levelof tenure. For workers with less than two yearsof initial tenure, four-year retention probabilitiesare estimated to have increased from 32.9 percentfor the 1983–87 period to 34.6 percent for the1987–91 period to 39.1 percent for the 1991–95period. However, for workers with two to lessthan nine years of tenure, rates first declinedthen rose slightly. The strongest evidence of adecline in job stability comes from the group ofworkers who initially had between nine and 15years of tenure. The retention rate for theseworkers declined from 81.6 percent for 1987–91to 74.8 percent for 1991–95. Retention rates alsodeclined sharply between 1987–91 and 1991–95

1

2

3

4

1960 ’70 ’80 ’90 2000

years

Total

Men

Women

Ages 25-34

Ages 35-44

Ages 45-54

2

4

6

8

1960 ’70 ’80 ’90 2000

years

Total

Men

Women

3

7

11

15

1960 ’70 ’80 ’90 2000

years

Total

Men

Women

Median tenure, ages 25–54

FIGURE 2

Note: Shaded areas indicate recessions.

Source: Authors� calculations based on data from U.S.Department of Labor, �Employee tenure in the mid-1990s,�January 30, 1997, and Farber (1998).

Four-year job retention rate estimates

Initial tenure 1983�87 1987�91 1991�95

0 to < 2 32.9% 34.6% 39.1%

2 to < 9 58.6 54.8 56.4

9 to < 15 82.7 81.6 74.8

15 plus 63.0 70.2 63.3

Weighted average 53.9 53.6 54.4

Source: Derived from Neumark, Polsky, and Hansen (1997).

TABLE 1


for workers with 15 or more years of tenure, butreturned to rates observed in the 1983–87 period.The weighted average rate was quite stable, fallingjust 0.3 percentage points from 1983–87 to 1987–91and then increasing 0.8 percentage points from1987–91 to 1991–95.

The results on retention probabilities areconsistent with those on tenure levels in suggestingsome modest declines in job stability for workerswith several years of tenure. Several researchershave reported more dramatic declines in jobstability during the 1980s and/or 1990s. Forexample, Boisjoly et al. (1994), Rose (1995), andMarcotte (1996) report evidence of declining jobstability from the Panel Study of Income Dynamics(PSID) data. Similarly, Swinnerton and Wial (1995)reported significant declines in job retention ratesin the 1991–95 period. However, in our view thecombined results of Diebold et al. (1996), Swin-nerton and Wial (1996), and Jaeger and Stevens(1997) show that the more dramatic declinesreported in the literature were largely the resultof researchers failing to take account of occasionalchanges in survey question wording. The mostcareful analyses of job stability trends imply thatthere have been at most modest declines in stabilityin the late 1980s and 1990s.12

Research suggests, we believe, larger declinesin measures of job security. Farber (1997a) ana-lyzes data from the seven Displaced Worker Surveys(DWS), CPS supplements that are described inthe next section. He finds that “rates of job lossare up substantially relative to the standard ofthe last decade, particularly when some consid-eration is given to the state of the labor market.”He also finds that displacement rates increasedmost for several groups, such as the more edu-cated and those in white-collar occupations, thathave traditionally had relatively low levels ofdisplacement, which implies that displacementhas become somewhat more democratic. Changesin the reasons workers give for their job lossalso point to especially large increases in whatthe media might mean by “downsizing.” Finally,Farber finds that the consequences of displace-ment in terms of time spent unemployed andreduced wage rates upon reemployment appearto be mainly a function of the business cycle. Theconsequences of displacement were worse duringthe recessions of the early 1980s and 1990s, butthere is little evidence of a secular increase inthe seriousness of displacement.

Valetta (1997) also finds an increasing numberof dismissals in data from the PSID, which is

consistent with Farber’s results (1997a). More-over, Valetta’s finding that the increase is con-centrated among workers with higher levels oftenure is consistent with our finding below thatthe increase in displacement rates for workerswith five or more years of tenure has been espe-cially dramatic.

Displacement trends forhigh-seniority workers

Below, we present new measures of the rateof job displacement for workers with five ormore years of tenure. We had two main goals.First and most important, we wanted the measuresto be comparable over time so that we couldaccurately judge whether displacement was in-creasing. Given changes in the underlying surveymethodology, this is not completely straightfor-ward and, despite our best efforts, it is possiblethat certain of our measures change over timefor reasons that have nothing to do with actualchanges in the rate of job displacement. Our sec-ond goal was to create annual time series, thehighest frequency possible, so as to be better ableto examine the relationship between displacementand wage inflation.

Our measures of displacement are basedprimarily on the Bureau of Labor Statistics’DWS. These surveys were conducted as supple-ments to the CPS in January of even years from1984 to 1992 and in February 1994 and 1996.13

For the purposes of the survey, displacement isdefined as involuntary job loss not related to aworker’s performance. Thus, displacement ex-cludes quits and cases in which workers are dis-charged for poor performance.14 The surveys areretrospective, asking individuals whether theyhave experienced job loss any time over the lastfive years in the case of the 1984 to 1992 surveysand over the last three years in the case of the1994 and 1996 surveys. Thus, our earliest infor-mation on displacement is for 1979 and our lat-est is for 1995.

For workers who report that they were dis-placed in the relevant time period, the DWS asksfor the specific reason for their displacement.The possible responses are:

■ Plant or company closed down or moved,■ Insufficient work,■ Position or shift abolished,■ Seasonal job completed,■ Self-operated business failed, and■ Some other reason.


This list of reasons is less than ideal. For ex-ample, insufficient work might be the reason whyone of the other events occurred. A plant mayhave closed because there was insufficient workto do. Position or shift abolishment is probablysupposed to cover instances of “corporate down-sizing,” but it is possible that those in nine to fivework environments will be confused by the ref-erence to shifts. In any case, it lumps together in-stances of complex “re-engineering” exercises,which presumably reflect long-run organiza-tional changes, with closings of shifts in factories,which are more likely to be associated withshort-run declines in demand. The seasonal joband self-employment categories don’t corre-spond to many people’s conception of job dis-placement and, in fact, make up only a trivialfraction of the job loss that we consider. Finally,perhaps because of some of the ambiguities ofthe preceding categories, “other” is a commonresponse. In fact, growth in the “other” categoryis responsible for a large percentage of the totalgrowth of displacement of high-seniority workers.

The first difficulty we face in constructing aconsistent measure of worker displacement isthat the DWS only collects information, such asthe year of displacement, the worker’s tenure,and other characteristics of the lost job, for atmost one incident of displacement over the rele-vant period. If workers were displaced twice ormore in the same period, they are instructed toanswer the additional questions for the lost jobon which they had the highest tenure. This inev-itably leads to some undercounting of incidentsof displacement. Moreover, as Farber (1997a) notes,the change in the length of the period over whichthe DWS asks workers to report on displacementcreates a problem of comparability over time,since the undercounting problem is more severewhen the interval covered is five years.

Farber’s approach to this problem is to exam-ine only displacement that occurred in the lastthree years of the five-year periods covered bythe 1984 to 1992 surveys. As he notes, these ratesare still not comparable to rates computed fromthe three-year intervals of the 1994 and 1996 sur-veys, because some workers may lose a job inyear one or two of the five-year period beforethe survey and then lose another job in year three,four, or five. If the workers had accumulatedless tenure on the second lost job than they hadon the first, they would be recorded as losing ajob in the 1994 and 1996 surveys, but not in the

last three years before the 1984 to 1992 surveys.Farber’s solution is to use PSID data to quantifythe frequency of job loss patterns and adjust ratesin the DWS to offset them.15

Our approach is to restrict our analysis toincidents of job displacement in which the affect-ed workers had five or more years of tenure.Obviously, it is not possible to lose two such jobsin one three- or five-year interval, so the numberof such job loss incidents should be correctly tal-lied no matter whether the year is part of a three-or five-year interval in the DWS. Of course, wewill miss all displacement incidents in whichworkers had less than five years of tenure. How-ever, the consequences of job loss are not likelyto be particularly great for workers with littletenure and, thus, our measure may capture themost important forms of job displacement.

The DWS gives us estimates of the numberof workers with five or more years of tenure whoare displaced in a particular year. To calculate adisplacement rate, we need to divide this estimateof the number of high-tenure displaced workersby the number of high-tenure workers who wereat risk in that year. We derive the latter figure asthe product of the level of total employment andthe fraction of total employment accounted forby workers with five or more years of tenure.

Our estimated displacement rate is rd

n ftt

t t

55

5= ,

where dt5 is the number of workers with five or

more years of tenure displaced in year t, nt is to-tal employment, and ft

5 is the fraction of employ-ment accounted for by workers with five ormore years of tenure.

As noted above, we derive estimates of dt5

from the DWS. To estimate nt , we use the CPSoutgoing rotation files. The outgoing rotationsare those CPS members who are in the fourthand eight months of their eight-month participa-tion, about 25 percent in a given month. Poolingthe outgoing rotations for all 12 months of theyear yields a large data set that can be used toestimate employment levels quite precisely. Toestimate ft

5, we use the CPS tenure supplementsdescribed earlier. As noted, these were conduct-ed in 1981, 1983, 1987, 1991, and 1996. To com-pute displacement rates for 1979 and 1980, weuse the value of ft

5 from 1981. For other years inwhich there was no supplement, we interpolatelinearly from the preceding and succeeding tenuresupplements. Because the fraction of workers


with five years of tenure changes very slowlyrelative to the number of displaced workers, thisinterpolation causes no problems.

There is another problem with the 1994 and1996 DWS. The follow-up questions on the detailsof the displacement episode are not asked ifworkers do not give one of the first three stan-dard reasons for displacement. This is unfortu-nate because, as we have already noted, a nontrivialand growing number of workers report “other”as their reason for displacement. To ignore workersnot responding with one of the three standardreasons would, we feel, significantly skew ourresults.16 However, for workers giving a non-standard reason, we do not know whether theyhad five years of tenure or in what year they losttheir job.

To deal with this problem, we estimatedstatistical models to gauge the percentage of dis-placed workers giving nonstandard reasonswho had five years of tenure and, of those, thepercentage who were displaced in each of thethree years covered by the surveys. The detailsof our procedure are contained in box 1. The ideais to use the displaced workers giving nonstand-ard reasons in 1992 to determine which workercharacteristics reported in the basic CPS wereassociated with having five years of tenure andthen to use those characteristics to estimate the

percentage of displaced workers reporting non-standard reasons in the 1994 and 1996 surveyswho had five years of tenure. Similarly, we usedthe workers reporting standard reasons and fiveyears of tenure in the 1994 and 1996 surveys todetermine the characteristics associated withbeing displaced in each of the three years coveredby the surveys and then used those characteris-tics to predict which year workers reportingnonstandard reasons were displaced.

Table 2 shows our basic results for overalldisplacement rates. The rows of the table corre-spond to the years in which displacement occurred.The first five columns of the table correspond tothe number of years after the displacement yearthat the displacement rate was measured. Forexample, the only information on the 1979 dis-placement rate comes from the 1984 DWS, whichwas conducted with a lag of five years. The esti-mated rate, 0.96 percent, is thus shown in thecolumn headed five-year lag. For the majorityof years, we have multiple measures of the dis-placement rate. For example, for 1985 we haveestimates from the 1986, 1988, and 1990 DWS.These rates, shown in the columns for one-,three-, and five-year lags, are estimated to be2.29 percent, 1.79 percent, and 1.43 percent,respectively.

Percent displaced among workers with five or more years tenure:Alternative measurement lags

TABLE 2

OverallYear One-year lag Two-year lag Three-year lag Four-year lag Five-year lag estimate

1979 0.96 1.46

1980 1.23 1.69

1981 1.59 1.23 1.92

1982 2.27 1.84 2.52

1983 2.41 1.66 1.52 2.25

1984 1.75 1.22 1.80

1985 2.29 1.79 1.43 2.22

1986 1.99 1.56 2.18

1987 1.83 1.34 1.34 1.84

1988 1.44 1.18 1.61

1989 1.71 1.62 1.85

1990 1.90 2.11

1991 2.76 2.35 2.83

1992 2.40 2.67

1993 2.92 2.31 2.89

1994 2.21 2.46

1995 3.44 3.44

Source: Authors� calculations from data of the U.S. Department of Labor, Bureau of Labor Statistics.


The results for 1985 illustrate the final diffi-culty we face in constructing an annual measureof job displacement for workers with five ormore years of tenure. That is, displacement rateestimates tend to drop as the time since the sur-vey increases. Workers seem to forget incidentsof displacement as time passes, a phenomenonnoted previously by Topel (1990) and others. Asa result, it is inappropriate to simply average thevarious measures to arrive at an overall dis-placement rate for that year. For instance, if wewere to directly compare the single estimate for1979 with the single estimate for 1995, we wouldbe comparing a rate measured with a five-yearlag with a rate measured with a one-year lag.Thus, the comparison would reflect not onlydifferences in actual displacement rates betweenthe years, but also the tendency of rates measuredwith a greater lag to be lower.

Table 2 reveals that estimated displacementrates tend to drop on average by about 11 per-cent for each additional year that the survey lagsthe year of displacement. Our solution to this

problem, which is described in detail in box 1, isessentially to adjust rates based on lags greaterthan one year upward by about 11 percent foreach additional year that the survey lags theyear of displacement. Our final estimates of theannual displacement rates, which are shown inthe last column of table 2, are averages of all theadjusted rates for the year in question. For in-stance, the estimated 11 percent annual declinesuggests that if the rate for 1979 had been mea-sured in 1980, it would have been 1.46 percent,rather than 0.96 percent. Thus, our final estimatefor 1979 is 1.46 percent. In a year with multiplemeasurements, the measures are adjusted bydifferent amounts, depending on how long afterthe year of displacement the survey was taken.For instance, the estimate for 1985 obtained witha one-year lag is left at 2.29 percent, but the rateobtained with a three-year lag is adjusted up from1.79 percent to 2.18 percent to reflect the addi-tional two years since the survey; the rate with afive-year lag is adjusted from 1.43 percent to 2.17percent to reflect the additional four years since

the survey. The adjusted rates of 2.29percent, 2.18 percent, and 2.17 percentare then combined to obtain the finalestimate of 2.22 percent.17

The final results are plotted overtime as the black line in figure 3, panelA. The overall displacement rate forworkers with five years of tenure roseduring the recessions of the early 1980sfrom 1.5 percent in 1979 to a peak of2.5 percent in 1982. It then declinedduring the economic expansion thatfollowed to a low of about 1.6 percentin 1988. Then, in the 1990s, it roserather dramatically. It is not surpris-ing that the rate should have risenduring the recession of 1990–91, butthe 1991 rate, at over 2.8 percent, was0.3 percentage points higher than in1982, even though by most measuresthe 1982 recession was much moresevere. More noteworthy is the fail-ure of the displacement rate to declineduring the expansion of the mid-1990s.Indeed, in 1995 the rate shot up to3.4 percent, its highest ever reading.The high overall displacement ratesthat we estimate for the mid-1990s areconsistent with the view that job secu-rity declined significantly for workerswith five or more years of tenure.

0.0

0.5

1.0

1.5

1980 ’83 ’86 ’89 ’92 ’95

percent

Plant or companyclosed down

or moved

Position orshift abolished

Slack work

B. Standard reason

0

1

2

3

4

1980 ’83 ’86 ’89 ’92 ’95

percent

Overall rate(all reasons)

Standard rate(standard reasons)

A. Overall and standard reasons

FIGURE 3

Displacement rates, workers with 5 years tenure

Note: Shaded areas indicate recessions.Source: Authors� calculations based on data from U.S. Department ofLabor, Bureau of Labor Statistics, Displaced Worker Survey, 1984�96.


Constructing an annual index of displacement

BOX 1

from 1992 to 1994 and 1996, and 2) the distri-bution of year of displacement conditionalon the independent variables is the same forworkers displaced due to standard and non-standard reasons.

The final task in computing annual dis-placement rates is to combine rates measuredfor the same year by different surveys into asingle overall rate. We did this by estimatingthe following simple statistical model:

log ( ) ,r s tst t st= + − − +α γ ε1

where rst is the displacement rate for year tmeasured by the survey in year s, and εst isan error term assumed to have constantvariance and to be uncorrelated across ob-servations. The parameter γ measures therate at which estimates of displacement ratesdecline as time between displacement andthe survey increases. Its estimate correspondsto an approximately 11 percent rate of decline.The overall rate is captured by the year ofdisplacement effects, α t. Specifically, the esti-mate of the rate corresponding to a one-yearlag between displacement and the survey isexp(αt). These are the estimates shown in thefinal column of table 2 and plotted in figure 3.

In order to compute estimates for sepa-rate demographic groups, we expanded theabove model to

log ( ) ,r s tdst dt dst= + − − +α γ ε1

where rdst is the rate for demographic groupd in year t as measured by the DWS of years. The demographic specific rates are thenexp(αdt). We also computed estimates of dis-placement rates adjusted for changes in theage and sex distribution. These were basedon models of the form

log ( ) ,r s tdstk dt k dstk= + + − − +α β γ ε1

where rdstk is the rate for the age and sexgroup k. The presence of the βk controls forchanges in the age and sex distribution thatmight affect estimates of overall rates. How-ever, the adjusted rates were similar enoughto the unadjusted rates that we only reportthe latter.

1See, for example, Maddala (1983) for an explanation of thelogistic regression model discussed below.

To estimate the fraction of workers report-ing nonstandard reasons for displacementin the 1994 and 1996 DWS who had five ormore years of tenure, we estimated a logisticregression model using the sample of suchworkers in the 1992 DWS.1 The dependentvariable in this model was an indicator forhaving five years of tenure and the indepen-dent variable consisted of dummy variablesfor the nine census regions, sex, ten-year agecategories, race, marital status, educationless than high school, high school graduate,some college, and college degree, as well aspart-time status, one-digit occupation, andone-digit industry of the person’s job as re-ported in the main CPS. We then used theestimates of the parameters of this model,along with the equivalent characteristics forworkers reporting nonstandard reasons fordisplacement in 1994 and 1996 to form anestimate of the probability that such workershad five or more years of tenure at the timeof their job loss.

We estimated the fraction of such workersthat were displaced in each of the three pos-sible years covered by the 1994 and 1996 sur-veys by estimating a multinomial logisticregression model on the sample of 1994 or1996 displaced workers reporting standardreasons for displacement. In this model, thedependent variables were indicators for theyear of displacement and the independentvariables were the same as in the modelabove. The parameter estimates were thenused to estimate the probability that work-ers were displaced in each of the three yearscovered by the 1994 and 1996 surveys.

In computing displacement rates basedon the 1994 and 1996 surveys, we thencounted all workers reporting nonstandardreasons for displacement as displaced in allthree possible years. However, we multipliedthe weights for such individuals by the esti-mated probabilities of having five or moreyears of tenure and of being displaced in theyear in question. This procedure should pro-vide estimates of displacement rates amongthose with five years of tenure that are con-sistent over time if 1) the relationship betweenthe probability of five-year tenure and theindependent variables remains constant


The colored line in figure 3, panel A showsthe rate of displacement due to the first threestandard reasons in the survey. Comparing thetwo lines, it is clear that a large part of the signif-icant mid-1990s increase is due to an increase inthe number of displaced workers giving “other”as their reason for displacement. However, eventhe colored line, which is not dependent on theimputations of tenure and year of displacementdescribed in box 1, suggests that there has beensome decline in security, especially given the levelof unemployment. The rate of displacement forstandard reasons is estimated to be higher in 1995than it was in 1982, even though the unemploy-ment rate was below 6 percent during most of1995, while it was nearly 10 percent in 1982. Thus,even when limited to displacement for standardreasons, our results suggest a noticeable declinein job security.

Figure 3, panel B displays separate displace-ment rates for the three standard reasons. Evi-dently, the rate due to firms or plants closing ormoving has declined somewhat in the 1990s,while the rate due to slack work has remainedrelatively high, given the state of the businesscycle. However, the most notable feature of fig-ure 3, panel B is the sharp increase beginningin 1990 of the displacement rate due to shifts orpositions being abolished. This rate, which prob-ably comes the closest to capturing corporatedownsizing, was between 0.2 percent and 0.3percent from 1979 to 1989, but rose to more than0.8 percent in 1995. This two hundred or threehundred percent increase seems to represent arather significant break from history.

Figure 4, panel A shows the overall dis-placement rate for men and women. For most ofthe period covered by our data, women were lesssubject to displacement than men, with the typi-cal gap in rates being five tenths or six tenths ofa percentage point. In the last three years, how-ever, the gap has been much smaller, about onetenth of a percentage point. Thus, by our mea-sure, women have suffered a larger decline injob security than men. This finding highlightsthe difference between the displacement ratesestimated here and the trends in median tenurediscussed earlier. Median tenure has generallybeen increasing for women relative to men.However, tenure levels are measures of stability,reflecting workers’ own commitment to the laborforce and individual employers in addition to forc-es beyond workers’ control, such as displacement.

In the comparison of male and female tenurelevels, workers’ own choices are likely the moreimportant factor. For this reason, we would arguethat displacement rates are the better measure ofworker insecurity.

Figure 4, panel B displays displacement ratesfor white and black workers. Once we restrictthe sample to workers with five or more yearsof tenure, the difference between the races is rel-atively minor. Still, there have been some changesover time. Early in the period covered, especiallyduring the recession of the early 1980s, blackshad noticeably higher displacement rates. How-ever, by the end of the period, whites had higherrates of displacement.

Figure 4, panel C shows the breakdownbetween those with a college degree and thosewithout a college degree. Although displace-ment rates for college graduates remain muchlower than those for workers without collegedegrees, the gap has narrowed considerably inthe 1990s. Until 1990, displacement rates for col-lege graduates never exceeded 1.3 percent andthe gap between them and non-graduates wasoften a percentage point or more. In the 1990s,displacement rates for college graduates roseespecially sharply, to levels of more than twicetheir previous peak. Thus, the gap in displace-ment rates between those with a college degreeand those without has narrowed considerably,though rates for college graduates remain signif-icantly lower.

Figure 4, panel D shows displacement ratesfor blue-collar and white-collar workers. Thoughdisplacement rates for high-tenure blue-collarworkers remain about a percentage point higherthan those for white-collar workers, the gap hasclearly shrunk during the 1990s. For instance,even the recessions of the early 1980s had littleeffect on displacement rates for high-tenurewhite-collar workers, but since 1988, their ratesof job loss have approximately doubled. By con-trast, the recessions of the early 1980s caused amajor increase in blue-collar displacement tolevels only slightly lower than in recent years.Even more dramatic differences in displace-ment trends are observed between more nar-rowly defined occupations. For example, 1995displacement rates for laborers are significantlylower than in 1982, while those for professionaland technical workers are approximately threetimes higher.


Figure 5 shows estimated displace-ment rates for workers in goods pro-ducing and service producing industries.Again, a large gap in rates in the 1980shas narrowed appreciably. Displace-ment rates for those in goods produc-ing industries are still significantlylower than in 1982, but rates for thosein the service producing industries areabout two and half times greater. Evenso, workers in goods producing indus-tries remain significantly more at riskfor displacement than those in serviceproducing industries. More dramaticchanges can be identified for certainindustries. For instance, displacementrates for workers in the finance indus-tries rose from about 0.5 percent to 1.0percent in the 1980s to 2.8 percent in 1995.

The results in figures 4 and 5 allpoint to the general increase in high-tenure displacement rates having beenaccompanied by a kind of democrati-zation, in which those who had beenrelatively immune to job displacementhave seen the fastest increase in displace-ment. Previously, those with a collegeeducation, in white-collar jobs, or inservice producing industries mighthave considered themselves immuneto job loss. Given the increase in dis-placement rates that we have estimat-ed for these groups, this is probablyno longer the case for many suchworkers. Thus, the number of workerswho feel at risk may have increasedeven more than the increase in thedisplacement rate would suggest.

Workers� perceptions of jobsecurity: The NORC-GSS

In a series of recent papers, Manski(1990, 1993) has observed that researchersknow a great deal about the outcomesthat individuals or groups experiencebut much less about the outcomes thatthey expect. This assertion is particu-larly relevant for job security research,which to date has focused on the mea-surement of displacement rates, ten-ure distributions, and other measuresof actual employment outcomes. How-ever, a primary issue in this literatureconcerns measuring perceptions of

0

1

2

3

4

1980 ’83 ’86 ’89 ’92 ’95

percent

Men

Women

A. Men and women

1

2

3

4

1980 ’83 ’86 ’89 ’92 ’95

percent

White

Black

B. White and black workers

0

1

2

3

4

1980 ’83 ’86 ’89 ’92 ’95

percent

Non-graduates

Collegegraduates

C. College graduates and non-graduates

0

1

2

3

4

5

1980 ’83 ’86 ’89 ’92 ’95

percent

Blue-collar

White-collar

D. Blue-collar and white-collar workers

FIGURE 4

Displacement rates among workers with five or moreyears tenure, by demographic category

Note: Shaded areas indicate recessions.Source: Authors� calculations based on data from U.S. Department of Labor,Bureau of Labor Statistics, Displaced Worker Survey, 1984�96.


the risk of future economic harm. Therefore,these measures are indirect, in the sense thatexpectations about risk, which are subjective innature, must be inferred from individual orgroup realizations.18

The General Social Survey (GSS) data setallows us to address the perceptions questiondirectly. Up to now, this data set has receivedsome attention in the popular press but littleamong researchers studying job security.19 TheGSS is a nationally representative annual sur-vey conducted by the National Opinion ResearchCenter (NORC). The survey asks a series ofdemographic and employment questions, in-cluding, in most years since 1977, two questionsabout job security. Respondents are asked1) “Thinking about the next 12 months, howlikely is it that you will lose your job or be laidoff—very likely, fairly likely, not too likely, ornot at all likely?” and 2) “About how easy wouldit be for you to find a job with another employ-er with approximately the same income andfringe benefits that you now have? Would yousay very easy, somewhat easy, or not easy at all?”

Data are available for 13 years between 1977and 1996 covering roughly 10,000 individuals.Several years (1980, 1984, 1987) are missingbecause NORC did not ask the job securityquestions, and other years (1979, 1981, 1992)are missing because the GSS was not conducted.The sample includes all respondents who arecurrently employed, English-speaking, andaged 18 to 64. It is important to note that the

sample makes no restriction on ten-ure because such information is notgiven in the GSS. Therefore, the job se-curity perceptions sample is not strictlycomparable to the displacement ratesample discussed earlier. This probablyaccounts for some different trendsamong subsamples of the population.

The GSS does have some impor-tant limitations.20 Most noteworthy isthat each GSS survey year consists ofan independently drawn nationallyrepresentative sample of the popula-tion. Thus, unlike other national sur-veys such as the PSID, the GSS doesnot allow us to observe the sameindividuals across time. Surveys thatfollow individuals allow the use ofpanel data techniques to control forunmeasured individual-specific char-acteristics, such as ability or ambition,

that change across the business cycle and are cor-related with the other variables in the model.Such a survey format would allow us to investi-gate the future employment dynamics of workersand examine whether job anxiety predicts futurejob displacement or wage loss.

The easiest way to see how perceptions of jobsecurity have changed over time is to graphicallyexamine the responses to the GSS questions. Fig-ure 6, panels A and B show the distribution ofresponses to the two questions from 1977 to 1996.Each line represents a separate response exceptthe highest line in panel A, which is the sum ofthe very, fairly, and not too likely responses.Between 30 percent and 40 percent of workers feelsome degree of insecurity about losing their jobin the next year, although only 10 percent of respon-dents feel very or fairly sure that job loss will occur.Between 35 percent and 50 percent of workersrespond that it would not be easy to find a compa-rable job.

As with the displacement rates, the responsesare fairly cyclical through the early 1990s. Usingthe job loss likelihood question, job security de-clined during recessions in the early 1980s and1990s and increased during the expansion of the1980s. But since 1991, the percentage of workerswho answer that they are not at all likely to losetheir job has fallen, despite the strong and widelyfelt expansion of the economy. Amazingly, in1996, the fraction of workers who answered thatthey had some concern about their job’s future

0

1

2

3

4

5

1980 ’83 ’86 ’89 ’92 ’95

percent

Goodsproducing

Serviceproducing

Goods producing and service producing

FIGURE 5

Displacement rate, by industry

Note: Shaded areas indicate recessions.Source: Authors� calculations based on data from U.S. Department of Labor,Bureau of Labor Statistics, Displaced Worker Survey, 1984�96.


0

20

40

60

1977 ’82 ’85 ’88 ’90 ’93 ’96

B. Comparable job at same pay and benefits?percent able to find comparable job

Not easy

Somewhateasy

Easy

1977 ’82 ’85 ’88 ’90 ’93 ’960

20

40

60

A. Lose your job in the next year?percent

Verylikely

Fairlylikely

Not too likely

Very, fairly, ornot too likely

FIGURE 6

Potential for job loss and comparable reemployment

Source: Authors� calculations based on data from the NationalOpinion Research Center, General Social Survey, 1997.

was equal to the percentage that answered thisway during the severe 1982–83 recession. How-ever, most of this is due to an increase in thepercentage of workers who answer they are nottoo likely to lose their job. Therefore, while therehas been a noticeable shift in worker anxietyduring this expansion, most of the change is dueto workers acknowledging some, albeit a slight,likelihood of losing their job over the next year.

The job comparability question also tends tobe cyclical, but showed signs of breaking thistrend during the initial phase of the 1990s expan-sion. Beginning in 1988, the percentage of workerswho answered that it would not be easy to finda comparable job at the same pay and benefitsmonotonically increased, peaking at almost 46percent in the 1994 survey. However, in the 1996survey, the “not easy to find a comparable job”response declined and the percentage answeringit was easy to find a comparable job increased.

Therefore, through 1996, workersseemed somewhat less concernedabout their chances of finding a com-parable job, but somewhat more con-cerned about the likelihood of losingtheir current job.21

Figures 7 and 8 show the trends inthese two series by gender, race, edu-cation, industry, and occupation. Thetwo primary differences between theGSS results and the displacement rateresults are exhibited in figure 7, panelsA and B. Panel A shows that there isno male–female gap in perceptions ofjob security throughout the sample pe-riod, while the displacement ratesshowed a large male–female gap thatnarrowed over time. Figure 7, panel Bdisplays a large black–white gap inworker anxiety that has narrowedsomewhat over time, while figure 4showed no significant difference indisplacement rates by race, exceptduring the 1982 recession. We believethat a substantial portion of these dif-ferences may be due to the differenttenure restrictions in the samples.That is, while the results on displacedworkers come from a sample of work-ers with five years of tenure, we cannotmake comparable restrictions on theGSS sample. The importance of thisrestriction is evident in other research.For example, Fairlie and Kletzer (1997)

use the DWS to estimate displacement rate gapsbetween black and white workers but make nosample restrictions based on tenure. They find a30 percent gap in displacement rates between theraces from 1982 and 1991. Likewise, between1982 and 1991, the black–white gap in the GSSjob loss data is 29 percent.

On the other hand, figure 7, panels C and Dlook quite similar to the displacement rate re-sults, pointing again to a democratization of jobinsecurity. White-collar and college-educatedworkers were relatively immune to job anxietyduring the 1970s and 1980s, but have experiencedsubstantial increases in job insecurity during the1990s. The change has been large enough to basical-ly eliminate the gap in job insecurity betweencollege graduates and non-graduates. Blue-collarworkers still feel less secure than white-collarworkers, but the gap is less than half what it wasin the 1970s and early 1980s.


As shown in figure 7, panel E, jobsecurity has declined during the 1990sin the service sector but has remainedrelatively flat (other than a temporarydrop in 1993) in the goods sector. Mostof the decline in the service industryarises from the services, finance, in-surance, and real estate (FIRE), andgovernment sectors. Analogous to thedisplacement rate findings, perceptionsof job security have dropped substan-tially in FIRE, with roughly 50 percentfewer workers saying that they are notat all likely to lose their job in the nextyear. The lack of movement amonggoods producing sectors hides somevariance between specific industries.In particular, job insecurity (measuredby the probability of losing your job)in manufacturing has doubled since1989, surpassing the level of anxietywitnessed in 1982. In 1996, job insecu-rity in the manufacturing sector wassubstantially higher than in all othermajor industries. The goods sector hasnot increased because agriculture andconstruction workers have experiencedcorresponding declines in job insecu-rity over the past few years.

Figure 8, panel A shows the per-centage of male and female workerswho believe it is not easy to find a com-parable job with the same pay and ben-efits. This graph shows a small butpersistent male–female gap that is elimi-nated in 1996. Using this job securitymeasure, most of the 1990s increase inanxiety appears to be due to femaleworkers. Figure 8, panel B shows thata large black–white gap during the 1970sand early 1980s had disappeared bythe end of the 1980s.

Figure 8, panels C and D againdisplay the diminishing gap in workeranxiety between college-educated andnon-college-educated workers andwhite- and blue-collar workers, respec-tively. Panel D shows that the histori-cally large difference between white-and blue-collar employees all but van-ished in 1996. White-collar workersare among the few groups that did notexperience a sharp drop in anxiety

20

30

40

50

60

1977 ’82 ’85 ’88 ’90 ’93 ’96

A. By genderpercent very, fairly, or not too likely

Male

Female

20

30

40

50

60

1977 ’82 ’85 ’88 ’90 ’93 ’96

B. By racepercent very, fairly, or not too likely

Black

White

20

30

40

50

60

1977 ’82 ’85 ’88 ’90 ’93 ’96

C. By educationpercent very, fairly, or not too likely

Non-graduates

College graduates

20

30

40

50

60

1977 ’82 ’85 ’88 ’90 ’93 ’96

D. By occupationpercent very, fairly, or not too likely

White-collar

Blue-collar

20

30

40

50

60

1977 ’82 ’85 ’88 ’90 ’93 ’96

E. By industrypercent very, fairly, or not too likely

Goods

Services

FIGURE 7

Likelihood of job loss, by demographic category



20

30

40

50

60

70

1977 ’82 ’85 ’88 ’90 ’93 ’96

A. By genderpercent not easy

Male

Female

20

30

40

50

60

70

1977 ’82 ’85 ’88 ’90 ’93 ’96

B. By racepercent not easy

White

Black

1977 ’82 ’85 ’88 ’90 ’93 ’96

C. By educationpercent not easy

Non-graduates

Collegegraduates

20

30

40

50

60

70

1977 ’82 ’85 ’88 ’90 ’93 ’96

D. By occupationpercent not easy

White-collar

Blue-collar

20

30

40

50

60

70

1977 ’82 ’85 ’88 ’90 ’93 ’96

E. By industrypercent not easy

Goods

Services

20

30

40

50

60

70

FIGURE 8

Likelihood of finding comparable employment,by demographic category


about finding a comparable job in1996, reflecting increased anxietyamong professional workers. In 1996,42 percent of professional workers re-sponded that it was not easy to find acomparable job, up from 30 percentin 1989, matching the percentage thatanswered that way during the 1982recession.

Finally, the 1996 drop in anxietyabout finding a comparable job ismainly from the goods sector (figure8, panel E). Anxiety about finding acomparable job for service sector em-ployees peaked in 1994, but remainsslightly above the levels seen duringthe last expansion. Nearly every groupbelieved it would be easier to find acomparable job in 1996 than in 1994,exceptions being government employ-ees and professional and sales workers.

Controlling for populationcharacteristics

The results presented thus far arebased on raw data. However, otherchanges in the work force during thelast 20 years, including shifts in theage and educational distribution ofU.S. workers, may be confounding thetime trends in worker anxiety. Is thetrend in worker anxiety by industry,occupation, or education the sameeven after simultaneously controllingfor multiple characteristics of the pop-ulation? To find out, we estimate“ordered probit” regressions, an ap-propriate statistical technique for thisproblem because it accounts for thediscrete and ordered nature of the jobsecurity questions. The details of theestimation procedure are describedin box 2.

Table 3 reports the coefficients,standard errors, and marginal effectsfrom a specification that uses the like-lihood of losing your job as the depen-dent variable and industry, occupation,year, gender, race, age, marital status,education, and region dummies ascontrols.22 The marginal effects mea-sure the impact of a change in somevariable, say whether the individualis a sales worker, on the probability


Ordered probit regressions

BOX 2

The ordered probit model is based on a latent regression such as

where yi* is the unobserved job insecurity of person i, xi are demographic and other individu-

al characteristics of person i and εi is a person-specific error term. The parameter β is a vectorof coefficients that measure the average impact of the demographic variables on the level ofjob security. While we do not observe yi

*, we do observe the k possible answers allowed by thesurvey, as represented by yi :

y y

y y

y y

y k y

i i

i i

i i

i k i

= ≤

= ≤ <= ≤ <

= ≤−

0 0

1 0

21

1 2

1

if

if

if

if

*

*

*

* .

µµ µ

µM

For example, in the GSS likelihood of losing your job question, yi = 0 corresponds to an-

swering “not at all likely to lose my job,” while yi = 3 corresponds to the “very likely to lose my

job” answer. The µi’s are unknown intercept parameters to be estimated in the model.Assuming a normal distribution in the error term, we can calculate the probability of each

of the k answers as

2

0

0 1

1

0

1

1

) ( )

( )

( ) ( ) ,

( )

Prob

if

if

if

y j x

x j

x x j k

x j kj j

k

= =+ =+ − + < ≤ −

− + =

RS|

T|−

−

ΦΦ Φ

Φ

µ βµ β µ β

µ β

where Φ is the standard normal cumulative distribution function. From equation 2, we cancalculate the marginal effect (the impact of a change in the x variable on the probability ofevent y occurring) by

∂ =∂

= +

∂ =∂

= + − +

∂ =∂

= − +

−

−

prob

prob

prob

( )( )

)( )

( ( ) ( ))

( )( ( )) ,

yx

x

y jx

x x

y kx

x

j j

k

0

3

1

0

1

1

φ µ β β

φ µ β φ µ β β

φ µ β β

where φ is the standard normal density function and the x variables are measured at theirmean value. Equation 3 essentially calculates the effects of changes in the covariates on the cellprobabilities.

It should be noted that many of the independent variables in our models are 0–1 indica-tors, such as whether the individual is a college graduate. In this case, the marginal effect iscalculated as the difference between the cell probabilities when the event occurs (a collegegraduate) and when the event does not occur (not a college graduate):

Prob Prob( , ) ( , ),y j x y j x= ′ − = ′1 0

where x′,1 is the vector of covariates where the college graduate variable is set to 1 and x′,0 isthe vector of covariates where the college graduate variable is set to 0.

1) y xi i i* ,= +β ε


Marginal effect on base case probability

Standard Not at all Not too Fairly VeryCoefficient error likely likely likely likely

Base case probabilitya 0.697 0.225 0.046 0.032

Agriculture 0.069 0.097 0.024 �0.014 �0.005 �0.005Construction �0.322 0.070* �0.120 0.062 0.027 0.031Manufacturing �0.234 0.054* �0.086 0.046 0.019 0.021Transportation, �0.058 0.066 �0.021 0.012 0.004 0.004 communications, and utilitiesWholesale trade �0.021 0.084 �0.007 0.004 0.002 0.002Retail trade �0.128 0.057* �0.046 0.026 0.010 0.010Finance, insurance, and real estate 0.060 0.069 0.021 �0.012 �0.004 �0.004Services �0.079 0.050 �0.028 0.016 0.006 0.006

Professional, technical 0.148 0.047* 0.049 �0.030 �0.010 �0.009Managerial 0.264 0.049* 0.085 �0.053 �0.017 �0.015Sales 0.119 0.056* 0.040 �0.024 �0.008 �0.008Craftsman �0.014 0.052 �0.005 0.003 0.000 0.000Operative or laborer �0.165 0.047* �0.060 0.033 0.013 0.014Service worker 0.126 0.048* 0.042 �0.026 �0.009 �0.008

1978 0.125 0.062* 0.042 �0.026 �0.009 �0.0081982 �0.194 0.061* �0.071 0.039 0.016 0.0171983 �0.236 0.060* �0.087 0.046 0.019 0.0211985 �0.089 0.060 �0.032 0.018 0.007 0.0071986 �0.018 0.062 �0.006 0.004 0.001 0.0011988 0.034 0.069 0.012 �0.007 �0.002 �0.0021989 0.062 0.070 0.021 �0.013 �0.004 �0.0041990 0.022 0.070 0.008 �0.004 �0.002 �0.0021991 �0.147 0.067* �0.053 0.029 0.012 0.0121993 �0.184 0.066* �0.067 0.037 0.015 0.0161994 �0.121 0.057* �0.044 0.024 0.009 0.0101996 �0.151 0.056* �0.055 0.030 0.012 0.013

Female �0.035 0.029 �0.012 0.007 0.003 0.003Black �0.265 0.039* �0.098 0.052 0.022 0.024Other race �0.095 0.070 �0.034 0.019 0.007 0.007Age 18�24 0.031 0.043 0.011 �0.006 �0.002 �0.002Age 44�65 0.130 0.029* 0.044 �0.027 �0.009 �0.008Never married �0.111 0.034* �0.040 0.022 0.009 0.009Divorced �0.024 0.037 �0.009 0.005 0.002 0.002Separated �0.198 0.064* �0.072 0.039 0.016 0.017Widowed �0.181 0.076* �0.066 0.036 0.014 0.015High school dropout �0.124 0.038* �0.044 0.025 0.010 0.010College graduate 0.024 0.038 0.008 �0.005 �0.002 �0.002Graduate school graduate 0.055 0.056 0.019 �0.011 �0.004 �0.004

New England 0.120 0.064 0.040 �0.024 �0.008 �0.008Mid-Atlantic 0.008 0.046 0.003 �0.002 �0.000 �0.000East North Central 0.093 0.044* 0.032 �0.019 �0.006 �0.006East South Central 0.040 0.059 0.014 �0.008 �0.003 �0.003South Atlantic 0.045 0.044 0.016 �0.009 �0.003 �0.003West North Central 0.077 0.055 0.026 �0.016 �0.005 �0.005West South Central �0.032 0.052 �0.011 0.006 0.002 0.002Mountain �0.116 0.058* �0.042 0.023 0.009 0.009

Intercept 1b 0.517 0.077*Intercept 2b 1.420 0.078Intercept 3b 1.849 0.079*

Log likelihood �18,316Sample size 9,935

* = significant at the 5% level.aBase case is a white, married male, aged 25 to 44, high school graduate, who worked a clerical government job in the Pacificregion in 1977. Industry, occupation, region, and year dummies are relative to government (industry), clerical (occupation),Pacific (region), and 1977 (years).bEach response has its own intercept. See box 2 for details. The three intercept terms are used to compute the marginal effects for the four categories of responses (final four columns).

Note: Dependent variable is the likelihood of losing your job in the next year. The possible answers are not at all likely,not too likely, fairly likely, and very likely.

Source: Authors� calculations based on data from the National Opinion Research Center, General Social Survey, various years.

Likelihood of losing your current job in the next year: Ordered probit analysis

TABLE 3


of the individual responding to the job securityquestions in a particular way. The results are re-ported relative to a base case white, married,male, high school graduate, aged 25 to 44, whoworked in a clerical job in the government in1977. The first row shows that the probability of thebase case person responding that he is not at alllikely to lose his job is 69.7 percent. The thirdrow, Construction, reveals that the probability ofa not at all likely response from a clerical workerin the construction industry is 12.0 percentagepoints lower (or 57.7 percent) than for a clericalworker in the government in 1977.

Overall, the table shows that many of thecharacteristics that look significant in the univari-ate graphs, such as occupation and race, remainsignificant indicators, even after controlling forthe demographic and employment variables.This is also true of ordered probit regressionswhere the comparable job question is the depen-dent variable. Table 3 gives further detail on thespecific industry and occupational groups thattraditionally experience higher levels of job anx-iety, including workers in the construction andmanufacturing sectors and operatives and laborersin all industries. While the probability of mana-gerial workers responding that they are not atall likely to lose their job is 78.2 percent, the sameprobability for an operative or laborer is 63.7percent. These industry and occupational differ-ences are statistically significant.

We can test whether job security has changedover this expansion by calculating time trendeffects within the ordered probit framework.The rows labeled 1978 to 1996 in table 3 showthe results of such an exercise. As with the sim-ple univariate graphs, perceptions of job securityhave been quite low since 1991 when measuredby the likelihood of job loss. Controlling for de-mographic, industry, and occupation shifts cannotexplain the recent high insecurity felt by workers.Job anxiety remains on the order of that seenduring the last two recessions.

Also, we stratified the sample by gender, race,education, occupation, and industry and ran sep-arate ordered probit regressions for differentcategories of workers. The purpose of this exer-cise is to see whether the time trends reportedin the graphs still exist after controlling for otherdemographic, industrial, and occupational struc-tural shifts. By running separate regressions foreach demographic group, we allow the parame-ters on other covariates to change across groups.

This flexible specification allows, say, the effectof being married to exert a different influence onperceptions of job security for high school drop-outs and college graduates. However, the maininferences from these results (not shown) do notchange much. The recent trend in increased jobanxiety arises primarily from better educatedand white-collar workers. On the other hand,workers who are high school dropouts are moresecure about their job in 1996 than at any othertime since 1977, with the exception of 1989, theend of the 1980s expansion. Managerial and pro-fessional workers have witnessed increases injob insecurity, while there is no statistical trendapparent in other detailed occupations. Increasedanxiety appears in manufacturing, services, andgovernment, while construction workers, whohave traditionally had a high probability of job lossbecause of the seasonal nature of the work, haveseen an increase in job security during the 1990s.

Lastly, we looked at the perceptions of jobloss among a few nonstandard groups of workers.Table 4 reports the coefficients, standard errors,and marginal effects from additional variablesthat are asked (sometimes periodically) in theGSS or are computed from other data sources.The first group of variables is other work charac-teristics, including union membership, the sizeof the employee’s work site, whether the firmpays fringe benefits, whether the organizationhas gone through a merger or reorganizationduring the last five years, computer usage in theindustry, and employment conditions in the in-dustry and region. The second group of variableslists several hardships that individuals haverecently experienced, including poverty, unem-ployment, work problems, financial problems,other hardships, and health problems. The letterat the beginning of each row indicates a separateregression that includes all of the demographicand employment variables reported in table 3.The sample sizes from each of these regressionsare reported in column 1. Since some questionsbegin to be asked after 1977, all marginal effectsare calculated relative to the same base case per-son in the first year that the question is includedin the GSS.

The first row shows that union members arelikely to be more insecure about their future jobprospects than nonunion members even aftercontrolling for compositional differences in oc-cupation and industry between the groups. Thisfinding may be confounded by the choice to join


Effect of other variables on likelihood of losing your current job in the next year:Ordered probit analysis

TABLE 4

Marginal effect on base case probability

Sample Standard Not at all Not too Fairly VeryOther variables size Coefficient error likely likely likely likely

Work characteristicsa. Union member 4,761 �0.124 0.049* �0.040 0.025 0.008 0.007b. Current organization merged 550 �0.209 0.132 �0.080 0.039 0.019 0.022c. Current organization reorganized 552 �0.123 0.120 �0.046 0.024 0.011 0.012d. No fringe benefits 463 �0.491 0.196* �0.193 0.077 0.042 0.074

Size of work sitee. 1�9 employees 3,079 0.158 0.079* 0.060 �0.032 �0.013 �0.015e. 10�49 employees 3,079 0.107 0.078 0.041 0.021 �0.009 �0.011e. 100�499 employees 3,079 0.076 0.080 0.029 �0.015 �0.006 �0.008e. 500+ employees 3,079 0.031 0.081 0.012 �0.006 �0.003 �0.003f. Region unemployment ratea 9,935 �0.086 0.019* �0.026 0.016 0.005 0.005g. Industry unemployment ratea 9,935 �0.029 0.020 �0.010 0.006 0.002 0.002h. Industry computer useb 9,529 �0.193 0.078* �0.070 0.039 0.015 0.015

Work and other problemsi. Below the poverty line 7,620 �0.352 0.051* �0.132 0.066 0.031 0.035j. Unemployment spell in last 5 yrs. 3,946 �0.448 0.045* �0.135 0.090 0.026 0.020k. Unemployment spell in last 10 yrs. 5,752 �0.403 0.035* �0.135 0.083 0.029 0.023l. Problems at work 281 0.004 0.184 0.001 �0.000 �0.000 �0.000l. Financial problems 281 �0.124 0.216 �0.043 0.028 0.010 0.006l. Other hardships 281 �0.535 0.219* �0.201 0.110 0.052 0.039m. Not healthy 5,292 �0.229 0.047* �0.087 0.043 0.020 0.025

* = significant at the 5% level.aIndustry and regional unemployment rates are calculated from the March 1977�96 Current Population Survey.bIndustry computer use is from Autor, Katz, and Krueger (1997). They calculate the share of computer users by three-digit Standard IndustrialClassification codes from the October 1984, 1989, and 1993 Current Population Survey. The computer use data are linearly interpolatedbetween 1984 and 1993, set at 1984 levels in year prior to 1984, and at 1993 levels in years post 1983.

Notes: The letter at the beginning of each row indicates a separate regression that includes all of the control variables listed in table 3. The basecase is the same as table 3. If the variables listed in this table were not reported in 1977, the base case is the first year the question was asked.Except for the unemployment rates and the computer usage variable, the marginal effects are reported as the difference between the base casewhere the characteristic is not present (say, the respondent is not a union member) and where the characteristic is present (is a union member).The size of work site coefficients are relative to a company with 50�99 employees. See box 2. The base case probabilities, which are not reported,are available upon request.

Source: Authors' calculations based on data from the National Opinion Research Center, General Social Survey, 1997.

a union (workers who are more insecure abouttheir future employment are more likely to joinunions) and therefore suffers from what econo-metricians call endogeneity bias. It is a bit sur-prising, because much of the research on unionssuggests that union workers are less sensitiveto business cycles because wages and employ-ment are set in multiyear contracts. However,union wages have been growing slower thannonunion wages recently, suggesting that workersare concerned enough about job security thatthey are willing to trade off wage growth formore security. Furthermore, the decline inunion membership over the last few decadescould signal reduced bargaining power of unionemployees.

In a way, the union results are similar to somelimited evidence on workers in organizations

undergoing change. In 1991, the GSS includedinformation on whether a respondent’s firmhas gone through a merger or a reorganization.Because we only have one year of data, the pre-cision of the point estimates is low. Neverthe-less, the magnitude of the marginal effects isconsistent with stories about restructuring anddownsizing leading to more insecurity duringthe 1990s. Unfortunately, because the sample isredrawn each year, we cannot test whether theseworkers have faced greater job loss frequenciesin subsequent years.

In regressions that add the size of the em-ployee’s work site, the results suggest that thosewho work at smaller sites are less likely to beconcerned about their job. However, the questionasks the size of the work site not the size of theorganization.


compensation during this tight labor marketexpansion is one of the driving forces behindpublic policy concerns about job security. Manyanalysts argue that workers have sacrificedwage growth for a more secure relationshipwith their current employers.

To investigate this question, we follow anestimation strategy pursued in Blanchflowerand Oswald (1994), among others. In particular,we look cross-sectionally as well as over timeand ask whether census regions that have hadhigher displacement rates or worker perceptionsof insecurity have tended to have lower wagegrowth. The regressions that we use are similarto the original Phillips curve, which posits anegative relationship between the rate of wagechange and the contemporaneous unemploy-ment rate.24

We use three wage measures, annual, weekly,and hourly earnings, that are computed fromthe 1977 to 1996 March CPS. However, our pre-ferred wage measure is hourly earnings, becauseit does not confound changes in wages withchanges in hours worked. This is an importantdistinction because annual hours are highly cor-related with the job security measures (as wellas unemployment rates). Therefore, we have tobe careful to distinguish a wage effect from a laborsupply (hours) effect. This is probably less of aconcern with the weekly measure.

We include controls for one of two job secu-rity measures (a security index calculated fromthe GSS and a displacement rate calculated fromthe DWS), the contemporaneous unemploymentrate (calculated from the March CPS), and time-and location-specific indicator variables. The time-and location-specific variables account for unex-plained characteristics of wages that are commonacross time and regions (essentially, they allowthe intercept term to vary over time and region).For example, the time variables will account forchanges in productivity growth and expectationsof inflation that are common across regionswithin the U.S. Box 3 gives the technical detailsof our estimation procedure.

Table 5 highlights some of our findings. PanelsA and B are from separate regressions. The secu-rity measure is the insecurity index in panel A,and the log displacement rate in panel B.

The wage effect is reasonably consistent acrossthe two job security measures. The coefficients(since the variables are measured in logs, thecoefficients are elasticities) suggest that, using theannual earnings measure, a 10 percent increase

When we add industry and regional unem-ployment rates to the statistical model, we find,not surprisingly, that workers in regions and, toa much smaller extent, industries that are expe-riencing higher unemployment are less secureabout their own prospects. (The unemploymentrates are calculated from the March CPS.)

Finally, we add a variable that measures theshare of computer users in each individual’sthree-digit SIC industry. The data are compiledby Autor, Katz, and Krueger (1997) using theOctober 1984, 1989, and 1993 CPS. As part of theEducation Supplements, the three CPS surveys askedworkers whether they used a computer at work,where a computer is defined as a desktop terminalor PC and not a hand-held data device or electron-ic cash register. We interpolate computer sharesby industry between the 1984 and 1993 end datesand hold years before 1984 and after 1993 constantat the 1984 and 1993 levels. Surprisingly, the re-sults suggest that workers in industries that aremore computer intensive are less secure abouttheir jobs, after controlling for demographics,time, industry, and occupation.23 When the com-puter usage variable is interacted with the timedummies, it becomes apparent that this computerindustry–job insecurity correlation is driven bythe 1993 to 1996 period. Prior to the 1990s, thereis a positive relationship between working in acomputer-intensive industry and job security.Unfortunately, we have no data on whether theindividual respondents are computer users.

The bottom of table 4 reports the parametersfrom the “problem” variables. Many of thesecoefficients are significant and negative, suggest-ing that job insecurity goes hand-in-hand withother work- and non-work-related problems.Furthermore, the large effects from previousunemployment suggest that these workers aremore prone to insecurity than those who havenot experienced a spell of unemployment in thelast ten years. This result suggests that past jobloss may be a reasonable indicator of future anx-iety and is consistent with studies that use dis-placement rates as an indicator of job security.However, the results on past unemploymentmay be driven by unobserved characteristics,such as ability.

Job security and wage growth

While a number of papers have measuredrecent trends in job security or stability, nonethat we are aware of attempt to link these trendsto wage growth. Yet the allegedly slow rise in


in the job security measure results in a 0.2 per-cent decline in wage growth. This is statisticallysignificant at conventional levels. However, thehourly wage coefficient implies that about half ofthis decrease is a wage effect and the other half isan hours-worked effect. Furthermore, the wage

effect is imprecisely enough estimated that wecannot reject the hypothesis that the true effect iszero. However, if this effect is real, the impact onnominal wage growth during the 1990s has beenfairly large. Referring to figure 3, displacementrates rose from around 2.0 percent in the 1980s

The impact of job security on wage growth

BOX 3

The wage-job security relationship is esti-mated from a regression of the form:

4) w y U wrt rt rt rt r t rt= + + + + +−α λ ϕ 1 v v v ,

where wrt is the log wage, yrt is a measure ofthe level of job security, and Urt is the log un-employment rate. The variables are aggregat-ed into a market r at time t. We aggregateindividuals into the nine census regions sincegeographical labor markets smaller than re-gions are not available in the GSS. The wageand unemployment measures are computedfrom the 1977 to 1996 March Current Popula-tion Surveys (CPS). The annual earnings mea-sure is a sum of all income earned in theprevious year. The weekly earnings measureis calculated as annual earnings divided bythe number of weeks worked in the previousyear. The hourly earnings measure is calcu-lated as annual earnings divided by the num-ber of weeks worked in the previous yeartimes the number of hours worked per weekin the previous year.

There are several ways to estimate equa-tion 4. Perhaps the simplest way is to aver-age all individuals in market r at time t anduse the cell means as the observation unit.This is essentially what was done for thedisplacement rates (see box 1). However, thestandard errors will be biased downwardbecause common unmeasured factors of in-dividuals may be attributed to local employ-ment conditions (Moulton, 1990). Instead,for the job loss likelihood index, we use a“two-step” procedure. In the first step, weestimate ordered probit regressions likeequation 1 in box 2 but augment them to al-low a calculation of a region-specific securityindex for each year. In particular, we regressthe loss likelihood responses, yi

*, on demo-graphics (Xit), and year interacted with re-gion dummies (Ritvt):

y X vit it it t itR* .= + +β δ ε

The vector of region–year dummy coeffi-cients (δ) is equivalent to the mean residualsby year and region and can be interpreted asindexes of job insecurity, after controlling fordifferences in education, gender, age, in-come, and marital status of workers in partic-ular areas. We also run the final wageequations with job security indexes that arenot demographically adjusted and find thatthis adjustment does not make a significantdifference to the inferences.

The wage variables are estimated from alog wage equation of the form:

ln w Xirt = +β µirt irt .

Again, the X matrix controls for educa-tion, marital status, and other standard humancapital controls. We use the mean residual byregion and year (µrt ) as a measure of the wageadjusted for these demographics.

In the second step, we estimate ordinaryleast squares regressions of the mean residu-al from the first stage wage equation (µrt ) onthe contemporaneous unemployment rate(Urt), the lagged dependent variable, the secu-rity index (δrt), and region and year indicatorvariables (or fixed effects):

µ αδ λ ϕµrt rt rt rt r t rtU v v v= + + + + +−1 .

The sample size varies depending on thejob security index that is used. The regres-sions that include the displacement rate arerun on 17 years (17 years times nine regionsequals 153 observations) and the regressionsthat include the security index are run on 13years (117 observations).


Relationship between wages and job security

TABLE 5

Change in Change in Change inlog annual earnings log weekly earnings log hourly earnings

No region Region No region Region No region Regioncontrols controls controls controls controls controls

A. (sample size = 117)

Insecurity index �0.018 �0.010 �0.013 �0.007 �0.013 �0.008(0.012) (0.013) (0.012) (0.013) (0.010) (0.011)

Log unemployment rate �0.029 �0.048 �0.029 �0.045 �0.024 �0.040(0.012) (0.015) (0.012) (0.015) (0.010) (0.013)

B. (sample size = 153)

Log displacement rate �0.021 �0.018 �0.020 �0.017 �0.010 �0.008(0.008) (0.008) (0.008) (0.008) (0.007) (0.007)

Log unemployment rate �0.028 �0.052 �0.028 �0.051 �0.026 �0.044(0.009) (0.011) (0.009) (0.012) (0.008) (0.011)

Notes: All regressions include year controls. See text for explanation of variables and sample. The unit of observationis the nine census regions.

Sources: Authors� calculations based on data from the Bureau of Labor Statistics, Current Population Survey, 1996;the Bureau of Labor Statistics, Displaced Worker Survey, 1984�96; and the National Opinion Research Center,General Social Survey, 1997.

to 2.75 percent in the early 1990s to almost 3.5percent in 1995. Using a job insecurity wageelasticity estimate of –0.01, this suggests that jobinsecurity lowered wage growth by 0.3 percent-age points a year in the early 1990s and roughly 0.7percentage points in 1995, relative to what wouldhave happened if displacement rates had stayedat the 1980s level. The job anxiety index grew ap-proximately 25 percent during the 1990s, suggest-ing a 0.3 percentage point decline in wages per yearfrom the results in table 5, panel A. However, ourestimates for hourly wages cannot reject the pos-sibility that these effects arose purely by chance.

Our analysis is just a first step in estimatingthe impact of job security on wage inflation. Thereis much more work to be done on this question.First, we plan to explore micro-data-based tech-niques to solve technical problems associatedwith having two measures, such as wage growthand job security, that are jointly determined.Second, as it is currently measured, the securityindex encompasses a fair amount of noise ormeasurement error. This measurement errorleads to a downward bias in the wage–securityrelationship. Finally, a key question is causation.Does high job security cause high wages or viceversa? This question could be examined by esti-mating vector autoregressive models, which allowa flexible relationship between wages, unemploy-ment, and job security.

Conclusion

Our review of the literature and our newresults on displacement for high-tenure workersreveal a modest decline in job stability and a largerdecline in job security, especially for workerswith higher levels of job tenure. Apparently,some of the increases in displacement that havebeen observed in the 1990s have been offset bydeclines in quit rates. The higher displacementrates suggest that workers have more reason tobe worried about their job security in the 1990s,and the lower quit rates suggest they may beless confident about their job prospects. Consis-tent with these findings, our tabulations ofworkers’ evaluations of their chances of job lossreveal a noticeable increase in the proportion ofworkers who feel that they are at least at somerisk of job loss.

When we relate variations in displacementrates and anxiety levels over time and acrosscensus divisions to the corresponding variationsin wage growth, we find estimates of the effectof insecurity on wages that would be largeenough to explain all or most of the puzzle ofslow wage growth in the 1990s. Of course, theseestimates are rather imprecise and may evenhave arisen by chance. Still, we believe that theseresults add to the case for worker insecurity hav-ing restrained wage growth and justify furtherresearch on the topic.


1Greenspan (1997).

2Reich (1997).

3See, for example, Staiger, Stock, and Watson (1997).

4The manner in which workers and employers are matched toeach other has changed quite noticeably in the 1990s. The pro-cess may have been made more efficient by the rapid expansionof the temporary services industry. (See Segal and Sullivan,1995, 1997). Also, Internet job postings may make interregionaljob search more efficient. Such developments may reduce thelikelihood of bottlenecks and spot labor shortages that contrib-ute to inflationary pressures. Or, in the language of the short-run Phillips curve, we can argue that they have reduced thenatural rate of unemployment independently of any increase inworker anxiety.

5New York Times (1996).

6Notable contributions to this literature include Podursky andSwaim (1987), Kletzer (1989), Topel (1990), Ruhm (1991), andJacobson et al. (1993a, 1993b, 1993c). Fallick (1996) and Kletzer(1997) provide recent surveys of this literature.

7The CPS is a monthly mini-census of about 45,000 householdsthat is the source for such familiar statistics as the unemploy-ment rate. When appropriately weighted to account for thescientifically designed sampling procedures employed by theBLS, the CPS yields nationally representative estimates.

8The figures are taken from U.S. Department of Labor (1997)and Farber (1998).

9Tenure data were also collected as part of the CPS Pension andBenefits Supplements of May 1979 and April 1993. The latter,which found higher median tenure for most age and sexgroups than the Mobility Supplement of January 1991, support-ed the conclusion of Farber (1998) that job durations were rela-tively stable in the 1980s and 1990s. However, the tenure datafrom the Pension and Benefits Supplements are based on a slight-ly different question than those from the Mobility Supplements.The latter asks how long workers have been continuously em-ployed by their current employers, while the former simplyasks how long workers have been employed by their currentemployers. Omitting the condition that the employment becontinuous could raise the tenure estimates. Suppose a workerwas employed by a firm for five years, left for two years, andthen returned for another five. In the Mobility Supplements, inwhich the question refers to continuous employment, theworker is likely to report a tenure of five years. However, in thePension and Benefits Supplements, in which the question simplyasks workers how long they have worked for their employers,the worker is likely to report a tenure of ten years. Thus, it ispossible that some or all of the higher tenure reading in the1993 survey was due to the omission of the word continuous inthe key question rather than an actual increase in job stability.

Changes in question wording also complicate the interpreta-tion of the trends in figure 2. Before 1983, the Mobility Supple-ments asked workers when they started working for theircurrent employers. Tenure was then calculated based on work-ers’ responses. Since 1983, workers have been asked how longthey have continuously worked for their current employers,which yields the tenure information directly. Of course, ifworkers correctly answered all questions, it would make nodifference whether tenure was solicited directly or calculatedfrom the start date of their jobs. But workers do not always re-port accurately; figure 1 shows that workers have a tendency toreport tenures that are multiples of five years. In the earlierMobility Supplements, there was a tendency to report start yearsthat were multiples of five, such as 1960, 1965, and so on. Thischange compromises the comparability of the data over time.

10Hall (1982) studied retention probabilities using data froma single cohort which, as shown by Ureta (1992), requires astable rate of job beginnings, as well as a stable set of retentionprobabilities.

11Neumark, Polsky, and Hansen (1997) report several alterna-tive estimates of retention probabilities. Those shown in table 1are, we believe, their preferred estimates.

12The lack of a major decline in job stability is also consistentwith the work of Stewart (1997), who analyzed the March CPSannual demographic files and found no increase in the rate ofjob change from the previous calendar year.

13See Hipple (1997).

14Of course, since this is a survey of individuals, there may beinstances in which respondents misreport by saying, for exam-ple, that they were displaced when they were actually fired.Such mismeasurement is possible with any household survey.

15The PSID has a number of advantages for studying turnoverand displacement. Unfortunately, the sample size is too smallto estimate disaggregated rates. Thus, Farber computes a singleset of adjustment factors that he applies to all workers.

16Tabulations, such as those in Hipple (1997), that do not countworkers displaced for nonstandard reasons do not show an in-crease in the current period comparable to what we find below.

17Our procedure yields what we believe to be consistent com-parisons across years. It is possible, however, that the overalllevel of the estimates could be off by some constant percentage.Suppose, for example, that the reason the displacement ratesdecline by 11 percent for each additional year that the surveylags the year of displacement is not that the rates measuredwith a one-year lag are correct and the other years reflectforgetting, but that the rates measured five years later arecorrect and earlier surveys reflect “spurious remembering.”(In our opinion, this is a much less likely scenario but one thatwe obviously can’t rule out.) Then our estimates will all betoo high by about 52 percent (11 percent compounded for fouryears), but the pattern across years will be unaffected.

NOTES


18Some observers (for example, Neumark and Polsky, 1997) ar-gue that attitudinal questions about job security may not pro-vide convincing evidence of actual job loss if perceptions areformed from misinformation. They point out that much of thereporting on job security relies on anecdotal evidence and,therefore, is not based on random sampling. It is always possi-ble to find someone who is struggling, even in a booming econ-omy. During the early 1990s, the recession hit journalists andeditors, as well as other white-collar workers especially hard,perhaps resulting in more stories about displacement thanwere warranted. Since press reports may help form percep-tions of the chance of job loss among readers, there is the dan-ger that we might observe an increase in perceptions of jobinsecurity that has little to do with actual job loss.

19An exception is Schmidt and Thompson (1997). Several poll-ing agencies, such as Gallup and Yankelovich, and surveyorganizations, such as the University of Michigan SurveyResearch Center (SRC), have also been soliciting perceptionsof worker security over the last two decades. See Otoo (1997)for an analysis of the SRC data. Dominitz and Manski (1997)describe the new Survey of Economic Expectations, which asksrespondents their level of concern about losing their job, losingpart of their income, losing their health insurance, and beingvictimized by a burglary. However, this survey began in 1994and therefore provides no information on longer-term trends injob security.

20See Dominitz and Manski (1997) for a criticism of the wordingof the GSS questions.

REFERENCES

Autor, David, Lawrence Katz, and Alan Krue-ger, 1997, “Computing inequality: Have comput-ers changed the labor market?,” National Bureauof Economic Research, working paper.

Blanchard, Olivier, and Lawrence Katz, 1997,“What we know and do not know about the nat-ural rate of unemployment,” Journal of EconomicPerspectives, Winter, pp. 51–72.

Blanchflower, David, and Andrew Oswald, 1994,The Wage Curve, Cambridge, MA: The MIT Press.

Boisjoly, Johanne, Greg J. Duncan, and TimothySmeeding, 1994, “Have highly-skilled workersfallen from grace? The shifting burdens of invol-untary job losses from 1968 to 1992,” Northwest-ern University, working paper.

Card, David, 1995, “The wage curve: A review,”Journal of Economic Literature, June, pp. 785–799.

Diebold, Francis, David Neumark, and DanielPolsky, 1997, “Job stability in the United States,”Journal of Labor Economics, Vol. 15, No. 2, pp.206–233.

, 1996, “Comment on Kenneth A.Swinnerton and Howard Wial, ‘Is job stabilitydeclining in the U.S. economy?’,” Industrial andLabor Relations Review, Vol. 49, No. 2, pp. 348–352.

Dominitz, Jeffrey, and Charles Manski, 1997,“Perceptions of economic insecurity: Evidencefrom the Survey of Economic Expectations,” PublicOpinion Quarterly, Vol. 61, No. 2, Summer, pp.261–287.

Fairlie, Robert, and Lori Kletzer, 1997, “Jobs lost,jobs regained: An analysis of black/white differ-ences in job displacement in the 1980s,” Univer-sity of California, Santa Cruz, working paper.

21The Gallup poll also found that the number of workers an-swering “not at all likely” to lose their job decreased during the1990s. On the other hand, Yankelovich, which asks whether los-ing your job worries you, found little change in response be-tween 1992 and 1995. But both polls found some job securitydifferences across the 1990s by education or income. Likewise,Otoo (1997) found significant increases in job anxiety between1988 and 1995 using the SRC micro data.

22Additional results, including an analogous table for the com-parable job question, are available from the authors upon re-quest.

23With no such controls, the correlation between industry com-puter usage and job security is positive and highly significant.

24Blanchflower and Oswald estimate models that relate thelevel of wages to unemployment. But using wage levels impliesthat the coefficient on a lagged wage variable is less than one,whereas most of the literature (see Card, 1995, and Blanchardand Katz, 1997) has found that it is close to one. When we ranlevel wage equations, we also found the coefficient on thelagged dependent variable to be one, suggesting that the rela-tionship is really between unemployment and security andchanges in wages. The change specification also avoids techni-cal problems associated with having a lagged dependent vari-able on the right-hand side with a regional fixed effect andserial correlation in the error term.


Jaeger, David, and Ann Huff Stevens, 1997,“Is job stability in the United States falling?Trends in the Current Population Survey and PanelStudy of Income Dynamics,” Yale University,working paper.

Kletzer, Lori, 1998, “Job displacement: What dowe know, what should we know?,” Journal ofEconomic Perspectives, forthcoming.

, 1989, “Returns to seniority after per-manent job loss,” American Economic Review, Vol.79, June, pp. 536–543.

Maddala, G. S., 1983, Limited-Dependent and Qual-itative Variables in Econometrics, Cambridge:Cambridge University Press.

Manski, C., 1993, “Adolescent econometricians:How do youth infer the returns to schooling?,”in Studies of Supply and Demand in Higher Education,C. Clotfelder and M. Rothschild (eds.), Chicago:University of Chicago Press.

, 1990, “The use of intentions data topredict behavior: A best case analysis,” Journalof the American Statistical Association, Vol. 85, pp.934–940

Marcotte, David, 1996, “Has job stability declined?Evidence from the Panel Study of Income Dynamics,”Northern Illinois University, working paper.

Moulton, Brent, 1990, “An illustration of a pitfallin estimating the effects of aggregate variableson micro units,” Review of Economics and Statistics,Vol. 72, No. 2, May, pp. 334–338.

National Opinion Research Center, 1997, GeneralSocial Survey, retrieved from the Internet atwww.icpsr.umich.edu/gss, September.

Neumark, David, and Daniel Polsky, 1997,“Changes in job stability and job security: Anec-dotes and evidence,” Michigan State University,working paper.

Neumark, David, Daniel Polsky, and DanielHansen, 1997, “Has job stability declined yet?New evidence for the 1990s,” paper prepared forFebruary 1998 Russell Sage Foundation Confer-ence on Changes in Job Stability and Job Security.

Fallick, Bruce, 1996, “A review of the recent em-pirical literature on displaced workers,” Industrialand Labor Relations Review, Vol. 50, No. 1, pp. 5–16.

Farber, Henry, 1998, “Are lifetime jobs disap-pearing: Job duration in the United States 1973–93,” Labor Statistics Measurement Issues, JohnHaltiwanger, Marilyn Manser, and Robert Topel(eds.), Chicago: University of Chicago Press,forthcoming.

, 1997a, “The changing face of job lossin the United States, 1981–95,” Brookings Papers onEconomic Activity: Microeconomics.

, 1997b, “Trends in long term em-ployment in the United States, 1979–96,” paperprepared for the third public German-AmericanAcademic Council Federation Symposium,“Labor Markets in the USA and Germany,”held at Bonn, Germany, June 10–11.

Gordon, Robert, 1997, “The time-varying NAIRUand its implications for economic policy,”Journal of Economic Perspectives, Vol. 11, Winter,pp. 11–32.

Greenspan, Alan, 1997, “Monetary policy,” testi-mony and report before the U.S. House Com-mittee on Banking, Housing, and Urban Affairs,104th Congress, 1st session, February 26.

Hall, Robert, 1982, “The importance of lifetimejobs in the U.S. economy,” American EconomicReview, Vol. 72, pp. 716–724.

Hipple, Steven, 1997, “Worker displacement inan expanding economy,” Monthly Labor Review,Vol. 120, No. 12, December.

Jacobson, Louis, Robert LaLonde, and DanielSullivan, 1993a, “Earnings losses of displacedworkers,” American Economic Review, Vol. 72,September, pp. 685–709.

, 1993b, The Costs of Worker Dislocation,Kalamazoo, MI: W.E. Upjohn Institute for Employ-ment Research.

, 1993c, “Earnings losses of high se-niority displaced workers,” Economic Perspec-tives, Federal Reserve Bank of Chicago, Vol. 17,No. 6, November/December, pp. 2–20.


New York Times, 1996, The Downsizing of America,New York: Random House.

Otoo, Maria Ward, 1997, “The sources of workeranxiety: Evidence from the Michigan survey,”Federal Reserve Board of Governors, workingpaper.

Podursky, Michael, and Paul Swaim, 1987, “Jobdisplacement earnings loss: Evidence from theDisplaced Worker Survey,” Industrial and Labor Rela-tions Review, Vol. 41, October, pp. 17–29.

Reich, Robert, 1997, Locked in the Cabinet, NewYork: Alfred Knopf.

Rose, Stephen, 1995, “Declining job security andthe professionalization of opportunity,” NationalCommission for Employment Policy, researchreport, No. 95-04.

Ruhm, Christopher, 1991, “Are workers perma-nently scarred by job displacements?,” AmericanEconomic Review, Vol. 81, No. 1, pp. 319–324.

Schmidt, Stefanie, and Christopher Thompson,1997, “Have workers’ beliefs about job securitybeen keeping wage inflation low? Evidencefrom public opinion data,” Milken Institute,working paper.

Segal, Lewis, and Daniel Sullivan, 1997, “Thegrowth of temporary services work,” Journal ofEconomic Perspectives, Vol. 11, Spring, pp. 117–136.

, 1995, “The temporary work force,”Economic Perspectives, Federal Reserve Bank ofChicago, Vol. 19, No. 2, March/April, pp. 2–19.

Staiger, Douglas, James Stock, and Mark Watson,1997, “The NAIRU, unemployment, and mone-tary policy,” Journal of Economic Perspectives, Vol.11, Winter, pp. 33–49.

Stewart, Jay, 1997, “Has job stability increased?Evidence from the Current Population Survey:1975–1995,” U.S. Department of Labor, Bureauof Labor Statistics, working paper.

Swinnerton, Kenneth, and Howard Wial, 1996,“Is job stability declining in the U.S. economy?Reply to Diebold, Neumark, and Polsky,” Indus-trial and Labor Relations Review, Vol. 49, No. 2,January, pp. 352–355.

, 1995, “Is job stability declining inthe U.S. economy?,” Industrial and Labor RelationsReview, Vol. 48, No. 2, January, pp. 293–304.

Topel, Robert, 1990, “Specific capital and unem-ployment: Measuring the costs and consequenc-es of job loss,” in Carnegie Rochester ConferenceSeries on Public Policy, Vol. 33, Allan H. Meltzerand Charles I. Plosser (eds.), Amsterdam:North-Holland, pp. 181–214.

U.S. Department of Commerce, Bureau of theCensus, 1996, Current Population Survey, Wash-ington, DC, February.

U.S. Department of Labor, 1997, “Employeetenure in the mid-1990s,” Washington, DC,press release, January 30.

U.S. Department of Labor, Bureau of LaborStatistics, 1984–96, Displaced Worker Survey,Washington, DC.

Ureta, Manuelita, 1992, “The importance of life-time jobs in the U.S. economy, revisited,” Ameri-can Economic Review, Vol. 82, March, pp. 322–335.

Valletta, Robert, 1997, “Declining job security,”Federal Reserve Bank of San Francisco, work-ing paper.

the decline of job security in the 1990s: displacement...

Documents