Abstract
Distance and direction measures are constructed and used to contrast occupational mobility following involuntary job displacement and total occupational mobility. Displacement involves specific capital loss. Some voluntary occupational mobility, for example, promotions, reflects augmented skills rather than specific human capital loss. Wage losses following displacement are strongly related to distance and direction. This is reflected in a downward shift in the skill portfolio. By contrast, the skill portfolio change in total occupational mobility shows a neutral or modest upward pattern, suggesting limited or no specific human capital loss from voluntary occupational mobility. The mean distance in occupational mobility following displacement declined significantly in the 1980s and 1990s suggesting the labor market was more efficiently reemploying workers following displacement, lowering displacement costs in that period.
I. Introduction
The main goal of this paper is to gain a better understanding of the relationship between occupational mobility and specific human capital. This paper builds on recent literature employing data on skills used or tasks performed in different occupations. This literature has already provided evidence that some occupations are quite similar in the sense of skills used or tasks performed, suggesting that specific human capital losses arising from mobility across these “close” occupations may be negligible.1 However, in assessing potential loss of specific human capital from occupational mobility, it is also important to be able to distinguish the direction of a move. Without this distinction, moving to a different occupation in which, for example, all the skills are used at a higher level (positive direction), would entail a loss of specific capital no different from a change to different skills or a move to an occupation at lower levels of all the same skills (negative direction).2 Some forms of occupational mobility, such as the voluntary movement up a career ladder, typically result in wage gains. Other forms of mobility, such as occupational mobility following an involuntary job displacement due to plant closure, are often associated with wage losses. Even though a distance measure might show a similarity for both types of mobility, the implications for specific human capital are likely to be different due to the difference in direction.
The paper develops and uses measures of occupation distance and direction, based on underlying skill portfolios, and distinguishes between different types of occupational mobility in a way that is relevant for understanding the implications for specific human capital. There are two main contributions. First, the paper shows that combining a direction measure with distance measures is particularly useful for characterizing which kind of occupational mobility results in specific human capital losses and associated wage losses. The analysis suggests that a large amount of occupational mobility involves little or no specific human capital loss, and may in fact be associated with human capital accumulation. Second, the paper presents evidence suggesting that the magnitude of specific human capital losses from occupational mobility following involuntary job loss (displacement) may have been in a secular decline throughout the 1980s and 1990s.
The structure of the paper is as follows. Section II provides a framework relating occupation, skills, tasks, and specific human capital. Occupation distance and direction are defined within this framework, and the implications for various forms of occupational mobility are discussed. A key element of the framework is a relatively low dimension skill vector that is used to characterize a large number occupations. Section III details construction of this vector for three-digit occupations and distance and direction measures between them based on analysts'ratings in the Dictionary of Occupational Titles (DOT). The datasets used are the Displaced Worker Surveys (DWS), the March Current Population Surveys (MCPS), and the basic monthly Current Population Surveys (CPS). Section IV describes these datasets and uses them to contrast the basic distance and direction patterns for occupational mobility following involuntary job loss (DWS) compared to all mobility (MCPS, CPS). The empirical distributions of occupation distance from the DWS and MCPS are compared to simulated distance distributions assuming that workers are randomly displaced (or randomly move) from a job and are then randomly assigned to a new job. The direction of the mobility is very different in the DWS compared to the MCPS or CPS. While the direction is modestly positive for total mobility (MCPS, CPS), it is strongly negative for mobility following involuntary job loss (DWS).
Section V examines the relationship between distance and direction measures and wage losses from occupational mobility following involuntary job loss. It shows the importance of direction in explaining wage losses. The main result is that increased distance only has a negative effect on wages if the direction is negative. Specific human capital loss from mobility between occupations that are further apart in terms of their skill mix is important when the direction is negative, but not otherwise. Section VI shows that the (negative direction) distance between occupations following involuntary job loss declined substantially up to the onset of the recession in the early 2000s. In contrast, overall occupational mobility was little changed over the same period. The reason for this decline in distance following displacement is unclear, but does not appear to be due to changes in the type of workers being displaced or the industry from which they are displaced, suggesting that the labor market was more efficiently reemploying workers following displacement throughout the 1980s and 1990s. Additional support for this interpretation is that, coincident with the decline in distance, there was also a decline in weeks without work following displacement over the same period. The estimates suggest that average wages losses were 7% lower for full-time workers and 8% lower for all workers as a result of the decline in (negative direction) distance.
Both distance and average wage losses show a similar cyclical pattern over the whole 1984–2010 period. The onset of the recession in the early 2000s and the Great Recession starting in 2008 reverses the pattern of decline of the earlier period in both distance and weeks without work following displacement. It is too early to tell if the decline seen in the 1980s and1990s will continue after a full recovery from the Great Recession. Some conclusions are presented in Section VII.
II. Occupations, Skills, and Specific Human Capital
The measures of occupation distance and direction developed and analyzed in this paper are based on characterizations of occupations in terms of their associated skills or tasks. This follows a recent literature that focuses directly on skills and tasks in a job, based on data that go beyond the standard industry and occupation coding.3 There is no single interpretation of the relation between skills and tasks in the literature or the relation to occupations. This section presents a simple theoretical framework, based on the previous literature, that helps to clarify the concepts of occupation distance and direction and their relation to skills, tasks, and specific human capitalas usedinthis paper. The basic concept is that workers, at any point in time, have a vector of skills that are transferable across occupations. The occupations produce output through occupation-specific bundling of tasks associated with these skills.4
A. Skills, Tasks, and Occupations
Workers start their career with some endowed, low dimension skill vector, s = (s1, s2, . sn), which may evolve over a career. The skills, s, give the worker the ability to supply amounts of associated tasks. Output is produced by tasks through J occupations. Workers generate (log) output in occupation j, based on their skill vector, according to a simple additive form, using the notation of Yamaguchi (2012):
(1)The production structure of occupation j is characterized by yj. The variation in yj allows the relative productive value of tasks to differ across occupations.
Assuming workers are paid their marginal product, the log wage in occupation j for a worker with skill vector s is given by:
(2)where lnpjis the output price of occupation j.5 Equation 2 captures the more general “bundling” idea for skills or tasks that creates variation in wages across occupations for the same skill vector.6 Without this feature, the individual skills in s would be priced out, and a worker moving the same skills across occupations would not experience any wage change. In these models the dimension of s is low, and hence the number of distinct tasks is small. By contrast, the number of ways the tasks can be bundled to produce (occupation) output is large so there can bea large number (J ) of occupations. The concepts of distance and direction put some structure on how to describe observed moves across this large number of occupations. How many of them are very different? In what sense can the direction of a move between different occupations be characterized? In what sense is human capital specific to skill or occupation?
B. Specific Human Capital, Distance, and Direction
In the literature on occupation-specific human capital, for example, Kambourov and Manovskii (2009), a switch across detailed (three-digit) occupation leads to the loss of the occupation-specific capital. Poletaev and Robinson (2008) argue that human capital is not specific to detailed occupation. They argued that provided the occupations were similar in terms of the basic skills used, there is no loss of specific human capital. The above framework incorporates these ideas. Truly occupation-specific human capital in this framework requires that specific components of s only produced output in specific occupations—that is, the associated component in y would be positive in one occupation and zero in the others. A shift of this skill from the occupation where it produced context of occupations includes Yamaguchi (2012); Autor and Handel (2013); Firpo, Fortin, and Lemieux (2013); Gathmann and Schonberg (2010); Speer (2017); and Guvenen et al. (2015). In an earlier paper Heckman and Sedlacek (1985) specified a Roy model of comparative sector advantage in which a worker's skills generated certain sector-specific tasks via sector-specific task functions in the context of industry sectors. 5. In essence, the occupation-specific markets are thick enough, in the sense of Lazear (2009), to warrant the assumption that workers are paid their marginal products. positive output would represent a loss of occupation-specific capital. More generally, with a large number of occupations, many of which are “close” in terms of y vectors, the sense in which there can be significant “loss” of specific human capital is when a worker's skill vector is moved from the worker's optimal occupation match to an occupation with avery different y thatyields much less output for a unit of the worker's abundant skill.7
In Autor and Handel (2013), where the focus is a one period assignment problem, workers choose the occupation that maximizes earnings given their skill vector. Yamaguchi (2012) considers a dynamic framework where workers choose the occupation in each period that maximizes lifetime utility. In either case this optimizing behavior links skill vectors to occupations in equilibrium. A major contribution of Yamaguchi (2012) is the novel use ofwhat he regards as task information that makes a clear distinction between worker skills and job tasks. For the purposes of this paper, however, an important result in Yamaguchi (2012) is that, despite the importance of allowing for meaningful differences in skills and tasks to understand occupational choice and wage evolution over the life cycle, the “derived policy function for occupational choice suggests that observed tasks can be interpreted as a noisy signal of unobserved skills.”8 Given this link, the approach taken here, as in Poletaev and Robinson (2008), is to characterize occupations as similar if skill vectors of the typical worker in the occupations are similar.9
Assume that at any point in time, most workers in the labor force have found their way into their optimal occupation.10 Workers with the same skill vector have the same optimal occupation. Forsimplicitysupposethatthereareenoughdistinctoccupationsto make the variation in the individual worker skill vectors si within occupation small. Denote the typical worker's skill vector in occupation j by sj, so that for all individuals in any occupation j, si ≠sj. The primary distance measure used in this paper between occupation j and j' is given by the distance, d(j, j') between the empirical counterparts for the skill vectors Si and Sj'. The analysis of direction is based on the sign of the differences in the elements of sj and sj'.
A complementary distance measure also defined in this paper, especially for use in contrasting distance and direction is based on the relative skill mix. Consider two workers, i and i', with the same relative skill vector in the sense that Si'= αSi. Will their ranking of occupations, and hence preferred occupation, be the same? Or, are occupations distinct in ways that depend on more than the relative bundle of tasks, so that offered wages are not simply scaled up or down? Worker i', for example, may be a “senior surgeon” with more experience than worker i, the “junior surgeon,” which resultedin ahigherlevelof, say, cognitive and motor skill, thoughstillusedin the same proportion. Does a “senior surgeon” do the same thing as a “junior surgeon”? The primary distance measure, as defined above, assumes that they are different occupations.
However, an alternative is to define a distance measure based on the relative skill vector. This emphasizes the similarity of task combinations or bundles across occupations. It allows for there to be different levels within an occupation, that is, better or worse carpenters or surgeons, but focuses on specificity residing in the combination rather than the level of the bundle.11
C. Implications for Different Types of Mobility
The empirical analysis deals with two types of occupational mobility: occupational mobility following involuntary job loss, and total occupational mobility. What happens to the allocation of workers to occupations following involuntary job displacement? Assume, as above, that prior to the displacement most workers in the labor force have found their way into their optimal occupation. In the absence of frictions, a worker who experienced involuntarily displacement from their job would not change occupation unless they were marginal in that occupation, in which case the distance between the occupations would be very small. Given mobility costs or search frictions, however, occupation changes involving larger distances could occur. The interpretation of these larger distances isthattheworkerisnowinaninferiorallocationorhasaninferiormatchbetweentheirown individual skill vector, si, and the skill vector that has an optimal match with their postdisplacement occupation, j; that is, siss'j and lnπj' + y1ys1+ &2fS2 + ... + 0„j-s„ • lnπj + y1js1+y2js2+. + ynjsn. For displaced workers, therefore, the wage consequences for these larger distance moves is negative.
Moscarini (2001) proposed a search-frictional Roy model, and Moscarini and Vella (2008) present evidence based on this model that the sorting of workers across occupations based on comparative advantage is noisier when unemployment is high. More generally, evidence from the displacement literature suggests that mobility costs or search frictions may be substantial.12 Especially for large plant closures, while there may be jobs available in the same occupation in the national labor market, so that eventually many workers may find their way back to the same (optimal) match, there may be a large enough impact on the local market for specific occupations so that those that have higher moving costs or lower search skills may not.
The assumption for displaced workers is that their predisplacement allocation was optimal, and that the displacement does not coincide with any change in their own skill vector, si. One interpretation of direction for displaced workers is in terms of whether they slipped backwards “down” a career ladder to an occupation that packages skills in a way that has at least one important element in sj that is lower than the corresponding element in sj. While the individual's own si is unchanged, the wage consequence is negative because of the suboptimal allocation to an occupation that is optimal for workers with a skill vector that is lower in at least one important element. The interpretation of a significant specific capital loss in this case is that the worker's skill vector has been moved from the worker's optimal occupation match to an occupation with a different y that yields much less output for a unit of the worker's abundant skill.
Total occupational mobility has different implications. Some mobility, especially for younger workers, will involve experimentation to find an optimal occupation match. However, a large part of the mobility will reflect life-cycle human capital accumulation where many of the occupation switches are promotions in the sense that the worker's individual skill vector, si, has evolved, making the choice (promotion) to a different (higher level) occupation optimal. For these cases, denoting the worker's new skill vector by s i, the worker now chooses occupation j such that s ixsj , and lnpj +y1j s 1 + y2j s 2 + . + ynj s n >lnpj + y1js 1 + y2js 2 + . + ynjs n >lnpj +y1js1 + y2js2 + . + ynjsn. Distance measures for total occupational mobility will thus reflect the “step size” of the skill vector for typical promotions. While it is unclear how the magnitude of this distance might compare with the distance for mobility following involuntary job displacement, for direction there is a more clear implication: mobility following involuntary job displacement is expected to be negative, whereas total mobility is expected to be mainly positive, reflecting life-cycle accumulation of human capital.
In a standard Ben-Porath style human capital model, as in Heckman, Lochner, and Taber (1998) and Bowlus and Robinson (2012), life-cycle accumulation of homogeneous (within education group) human capital takes place outside the framework of occupations. The accumulation is inferred through observed increases in wages. In the recent literature using multiple tasks and skills, as in Gathmann and Schonberg (2010), Yamaguchi (2012), and Bowlus, Mori, and Robinson (2016), human capital accumulation is associated with changes in occupations that represent “promotions” into occupations with higher levels of tasks or skills. Accumulation in this literature is inferred through observed changes in occupations and their associated skill vectors.
III. Construction of the Distance and Direction Measures
The first step is a skill vector characterization, sj, of each of the J threedigit occupations. This requires a source of data on skills or tasks that goes beyond the usual information gathered for standard occupational or industry coding. The literature has taken two approaches to constructing s. One approach uses a relatively small number of underlying skill or task measures to characterize each occupation directly in terms of a vector of these measures.13 A more indirect approach starts with a much larger dimension vector of underlying measures and uses some form of factor analysis to extract a smaller number of “skills” or factors. This is the approach taken here.14 The underlying information on skills or tasks for the occupations is taken from the fourth edition of the DOT. These data represent the ratings or scores given for a large number of “characteristics” of the 12,741 DOT jobs by the DOT job analysts. While the units vary, the ratings all have a clear hierarchy as required for the elements of sj. They all refer to levels of skills or levels of task performance associated with skills of a typical worker or required of a worker in the job. The basic rationale for using a factor analysis is the assumption that occupations can be distinguished on the basis of a relatively small number of skills and their associated tasks (“factors”) and that the large number of characteristic ratings in the DOT are reflections of the amounts these underlying skills and their associated task units for the typical worker in the occupation.
The past literature on heterogeneous human capital has made some distinction between skills and tasks, though often they have been treated interchangeably. A recent discussion of the distinction in the context of the DOT job-based measures used in this paper is provided in Yamaguchi (2012). Yamaguchi (2012) treats the underlying skills as unobserved and uses a principle components analysis on a small number of DOT“job characteristics” to derive what are interpreted as direct measures of the tasks or “task complexity.” Some of the DOT measures are clearly interpretable as tasks, but a much larger number of them used in this paper, are more naturally interpreted as skills. Three of the DOT measures (“data,” “people,” and “things”) come from the three digits that are part of the DOT code itself. According to the DOT codebook description, these “represent the worker function ratings of tasks performed in each occupation” for the worker's interaction with data, people, and things. It is reasonable to interpret the hierarchical levels of, for example, “people,” such as mentoring, supervising, orserving as different hierarchical level tasks, or task complexities as in Yamaguchi (2012). The remaining 46 DOT measures (excluding environmental conditions variables) are in the DOT trailer. The 11 “temperaments” are referred to as “personal traits required of a worker” in the job. The 11 “aptitudes,” such as finger dexterity, are measured on a scale expressed with reference to how much of the aptitude the worker “possesses.” For a score of 1, for example, the worker's aptitude for finger dexterity is equivalent to the finger dexterity in “The top 10 percent ofthe population. Thissegmentofthepopulation possesses an extremely high degree of the aptitude.” The language of “personal traits” and “possession” of an aptitude suggests that these measures could reasonably be interpreted as direct skill measures rather than task measures.
The analysis of occupational mobility requires a skill vector characterization at the levelofthree-digitoccupationcodingusedinstandarddatasets. First,theinformationon the12741 DOT jobs was used to create mean ratings for 49 DOTjobcharacteristicsjobs for each three-digit occupation using primarily a “weighted crosswalk” approach. The DOT crosswalk files assign census occupation codes to DOT jobs. The weighted crosswalk approach uses the employment weighting information by sex at the level of the DOT job from a special April 1971 CPS file and applies this to the DOT crosswalk files. This allows DOT job-level employment weighting within three-digit occupation to be used in calculating mean values of the DOT characteristics scores by census occupation code for the different coding schemes. Second, values for 49 DOT job characteristics scores are assigned to all employees in the January 1992 Current Population Survey on the basis of their sex and three-digit occupation code, and a standard factor analysis was applied to extract factors from 49 DOT job characteristics ratings. Identification is achieved in the standard factor analysis by normalizing the factors to be mean zero, with a standard deviation of one and by assuming that the factors are orthogonal. It results in four factors, each expressed as alinearfunction ofthe 49 DOT job characteristics ratings.
The factor that explains most of the variance (about 40 percent) appears to capture some kind ofgeneral intelligence or analytic skill; the second factor (about 20 percent) emphasizes fine motor skills, while a third factor (about 12 percent) is more related to physical strength. After these three, the remaining factors contribute relatively little. One additional factor appears to pick up visual skills—the factor loadings emphasize color discrimination, color vision, far acuity, and field of vision. These factors are very similartothoseinPoletaevandRobinson(2008).15However,thesefactors,unlikethose used in Poletaev and Robinson (2008), have different values for males and females within occupation. In principle, the factor analysis could be done separately for males and females and a potentially different set of factors for males and females derived. The assumptionusedinthispaperisthatmalesandfemaleshavethesameunderlyingskills, but that the amounts may differ. This requires a single factor analysis to impose the same factors or skills, and allows different levels of the skills for males and females within the same occupation to be generated by different mean DOT characteristic ratings within three-digit occupation due to different DOT-level job weights for males and females within occupation.16
The four-dimensional vector of factor scores provides the estimated skill vectors, sj, that characterize each three-digit occupation. The distance measures are alternative descriptions of the distance between these vectors. The starting point is euclidean distance. A standard practice in factor analysis is to “rotate” the factors as a way to find more easily interpretable factors. An important advantage of the euclidean distance measure is invariance to factor rotation, so that no a priori information need be imposed. In particular, it is not necessary to interpret the factors as the “true” underlying skills. The euclidean based distance between occupations j and j employed here is defined as:
(3)where Δk = skj - skj is the difference in the kth component of sj and sj , and the wk are weights that sum to one. The benchmark measure used in this paper is an equally weighted continuous measure, where the weights are all 0.25. In terms of the units of the underlying factors, this measure is equal to l when there is a change of l standard deviations in each of the factors. In Poletaev and Robinson (2008), a distinction was made between the “main” skill and less important skills in an occupation, based on the individual factor scores. To capture this idea theweights can be adjusted according to the importance of the factor in the occupation as reflected in the factor scores.17
A complementary relative distance measure is based on the same skill vectors but assumes that distinct occupations are mainly characterized by their relative skill vectors, as in the examples of surgeons and carpenters discussed in Section II.B. In the normalization from the factor analysis each factor has a mean zero and standard deviation of one in the population used for the estimation. This presents a problem for computing skill proportions, so the factors are first renormalized by adding a constant so they are always positive to construct the relative skill vectors, pj. The distance measure based on the skill proportions, reledist4, is then defined as:
(4)where (ΔPk) is the difference in the kth component of pj and pj , and the wk are each 0.25. While the absolute and relative distance measures, edist4(w) and reledist4 are based on the same underlying skill vectors, they emphasize different features of mobility. Suppose, for example, that job changes for most workers typically involve scaling up or scaling down a given skill vector, that is, performing the same job at a higher or lower level. In this case, reledist4 will be insensitive to this mobility relative to edist4(w), which will capture the scaling effects. Alternatively, if job changes frequently involve changes in the skill mix, both reledist4 and edist4(w) will capture these effects, though reledist4 will put more weight on the deviation from the overall mix.
The paper uses two approaches to measuring the direction of occupational mobility. The most direct is an examination of the differences in the individual components, Δk, ofsjandsj . A characterization of direction is straightforward when all components have the same sign, but is more complicated otherwise. The benchmark direction measure is a simple sum of the components. An alternative direction measure uses weights based on the importance of the skill component in determining wages. The direction measures are used together with the distance measures to distinguish between increases or decreases in a multidimensional skill vector used on the job for a given magnitude of change in the distance.
IV. Occupational Mobility: Patterns of Distance and Direction
The framework of Section II suggests that both distance and direction should be interpreted differently for occupational mobility following involuntary job loss compared to total mobility. For total mobility, provided most workers accumulate human capital over the life cycle, the expected direction of mobility is positive, reflecting this accumulation process. In contrast, for occupational mobility following involuntary job loss, unless all workers are able to find jobs in their previously optimally chosen occupation, the expected direction of mobility is negative, reflecting specific capital losses. The distance estimates for the two types of mobility should also be interpreted differently. The estimated distance for occupational mobility following involuntary job loss reflects a combination of the magnitude of specific capital losses and the level of frictions in the labor market. For total mobility the estimated distance reflects primarily the size of the promotion or human capital accumulation “steps” over a labor market career.
A. The Data
The data for occupational mobility arising out of involuntary job loss is the DWS. The DWS is a supplement questionnaire applied every two years since starting in 1984 to the monthly CPS. From 1984 to 1992 respondents are asked to provide information on job separations in the previous five years; from 1994 onward, the period is reduced to three years. The respondents areaskedwhetherthey had been displaced fromajob, and why. The supplement is designed to focus on the loss of specific jobs that result from firms'business decisions, unrelated to the performance of particular workers. Two types of “involuntary” job loss are considered in the analysis. The first is displacement due only to plant closures. In Neal (1995) and Poletaev and Robinson (2008), attention was restricted to plant closures, which has always been the first listed reason for displacement in the DWS. This is commonly used as an approximation for exogenous involuntary job loss (Gibbons and Katz 1991). An alternative, broader definition used by the BLS adds the additional reasons “slack work” and “position or shift abolished,” providing a larger sample.
The data for total occupational mobility come from the MCPS and CPS. Occupation coding inallthreedatasetsis done at the level of three-digit census occupation codes. Almost all original 1980 and 1990 census occupation codes are retained by constructing a modified 1990 occupation coding scheme with 494 codes, which are consistent across all the DWS surveys from 1984–2002 and for the MCPS and CPS survey years 1983–2002. There is a significant coding break with the introduction of the 2000 census occupation codes in 2003. Additional surveys up to 2010 are included in the analysis using dual 1990 and 2000 occupation coded CPS files.18Distance, di, for individual i observed in the data is constructed using a comparison of the occupation codes in the pre- and postdisplacement jobs for individuals in the DWS. In the MCPS the comparison is between the occupation codes in the occupation of the longest job held last year with that of the current job. In the monthly CPS, individuals are matched in the files for adjacent months using a unique Integrated Public Use MicroData Series (IPUMS) person identifier, and the comparison is between the occupation codes in the current and preceding month.19
There is a well-known problem of measurement error in occupation coding in most datasets. These potential coding errors have been extensively discussed in connection with accurate measures ofoccupational mobility. It is wellknown, for example, that the amount of spurious occupation switching in the CPS is very high unless a form of dependent coding is used that uses direct questions on whether the individual is still doing the same job.20 The MCPS has dependent coding throughout the 1983–2010 period. The monthly CPS introduced dependent coding in 1994.
The January 1977 CPS Validation Survey provides two coded sources of what should be the same 1970 three-digit occupation code, one based on an employee response in the basic CPS survey and one based on an employer response in the validation supplement. Comparison of the responses for the matched sample show strong evidence that miscoding ofindependently coded occupations is largein terms of the frequency of producing “switches” when there should be none, as has commonly been reported. However, in addition there is also strong evidence that this miscoding is largely between “close” occupations in terms of the distance measure used in this paper. The measurement error problem is thus much smaller for distance analysis. A more detailed analysis is given in the Appendix.
B. Distance Patterns
Figure 1 presents the cumulative distribution of the observed distances, di, for fulltime males age 20–61 displaced from the private sector by plant closing from the pooled DWS sample for the 1984–2002 period of consistent (modified 1990) occupation coding using the equally weighted euclidean distance measure, edist4. The center-hand panel presents the unconditional distribution, and the right-hand panel showsthedistributionconditionalonswitchingoccupation.Thesearecomparedwith reference random mobility benchmarks showing the simulated distributions that follow from randomly displacing workers from a predisplacement occupation and randomly assigning them to postdisplacement occupations using the observed occupation employment weights.21
Less than 1.5% stay in the same occupation under random mobility, but about 40% stay in the same occupation among males displaced from the private sector by plant closing, based on the independent occupation coding in the DWS. Only 29% of the mobility from the random experiment benchmark have d • 1, compared to 69% of the mobility from plant displacement. (For females the comparison is 31% vs. 73%.) Since, as discussed in the Appendix, coding error tends to impart an upward bias to the estimated distance, this is a lower bound on the fraction in the DWS with true d • 1.
A large part of the difference from random mobility in the unconditional distributions is due to the difference in the fraction switching occupation. However, the distances are alsosmallerintheDWSdatacomparedwithrandommobilityconditionalonswitching occupation shown in the right-hand panel. In the random mobility benchmarkfor males, only 28% haved • 1 while 51% of males displaced fromaplant closure in the DWS have di• 1. For females, the difference is even more pronounced: 30% compared to 61%.22
The mean distances, conditional on switching occupation, for workers displaced from plant closures, 1.05 for males and 0.91 for females, are much lower than the means from random mobility, but should they be considered large or small themselves? The measure edist4 is constructed such that it is equal to l when there is a change of l standard deviations for each of the individual components of s in the 1992 population of employees. This provides some interpretation of the magnitude. An alternative is to compare these means with the average distance of switches within three-digit occupation at the DOT job level. Given 12,741 DOT jobs in the revised fourth edition DOT, compared with about 500 three-digit occupations, there are about 25 DOT jobs per three-digit occupation. Unfortunately, the DWS only contains three-digit occupation codes, not DOT jobs. However, it is possible to compare the distribution of distance measures for random mobility across three-digit occupations and across DOT jobs within three-digit occupation. The four factors estimated from the factor analysis and used to describe distance across occupations are simple linear functions of the DOT characteristics. The same linear functions may be used to estimate the s at the level of the DOT job. The same distance measure may then be used to describe distance across DOT jobs.
Table 1 compares the distribution of distance measures for random mobility across three-digit occupations and across DOT jobs within three-digit occupation. The top half of Table 1 reports theresults for edist4. The distances for moves within occupations are, as expected, much smaller than those across occupations. Approximately 95% of random mobility within occupation has adistancebelow one, compared to only29% (31%) for random mobility for males (females) across occupations. However, the mean (and median) distance within occupation is still quite large at almost half of that across occupations. The lower half of Table 1 repeats the comparison for the distance measure based on skill proportions, reledist4, with similar results. Since the (DPi) are proportions, rather than the levels in (Δi), the scales for the distance measures are different, but they both show a similar relation for the “within” and “across” comparison.
Based on the mean(and median)distance within occupations reported in Table 1, only occupation switches with a distance in excess of 0.61 are larger distance moves relative to random moves within three-digit occupations. For displaced workers in the DWS that switch occupations, 26% of males and 29% of females have distances less than 0.61. Within the framework of Section II, this evidence is consistent with the presence of some frictions and potential specific capital losses. However, it also shows that despite the frictions, displaced workers are seeking, with reasonable success, to find their way back to jobs in occupations that are “close” to their previously achieved optimal assignment.
Figure 2 repeats the analysis for total mobility comparing the cumulative distribution of d from random mobility with the distribution of di for males and females age 20–61 that switch jobs in the MCPS. The center panel shows that while less than 1.5% stay in thesameoccupationunderrandommobility,over70%ofjobchangersintheMCPS stay in the same occupation.23 As with the DWS, while a large part of the difference is due to the difference in the fraction coded as switching occupation, there is also a significant difference due to smaller distances conditional on switching occupation shown in the right panel of Figure 2. For males under random mobility, only 28% have d • 1 while 47% of observed male job switchers in the MCPS have di • 1. For females, the difference is more pronounced: 30% compared to 59%. In terms of the simple framework in Section II, these results for total mobility are consistent with a combination of displaced workers seeking to return to their optimal assignment, and for voluntary job switches reflecting career moves similar to those in Gathmann and Schonberg (2010) and Yamaguchi (2012) where workers move between relatively close occupations.
The results are robust to alternative distance measures based on relative factor importance. In the random mobility benchmark across occupations, the two measures, edist4 and reledist4, are highly correlated (0.84). Across DOT jobs within occupations it is a little higher (0.87). It is also higher in the DWS (0.89) and MCPS (0.87). Not surprisingly given this high correlation, the pattern of results in Figures 1 and 2 is replicated when reledist4 is substituted for edist4. This suggests that a large part of mobility across occupations in the DWS and MCPS involves changes in the skill mix and is not simply a scaling up or down of a given skill vector.
C. Direction of Mobility
The framework of Section II has very different implications for the direction of total mobility compared to mobility following involuntary job loss. Table 2 presents evidence on the direction of change of s for workers switching occupation in the DWS (19842010) following involuntary job loss and contrasts this with the direction of change for total occupational mobility in the MCPS (1983–2010) and the matched CPS (19942010).24 The first four columns of Table 2 report the average change, Dsi, in each of the components of s. For the purposes of discussion, the four components of s may be given similar shorthand names to Poletaev and Robinson (2008), “analytic-general intelligence skills” (s1), “fine motor skills” (s2), “strength and gross motor skills” (s3), and “visual skills related” (s4), based on inspection of the characteristics that loaded most heavily in the factor analysis.
As noted earlier, defining overall direction on the basis of s is complicated when not all the components of s move in the same direction. In most cases in Table 2 the component changes are significant in only one direction, so the difference of direction in the two forms of mobility is relatively clear. However, to get an overall picture of the direction patterns, the remaining two columns report results for two alternative summary direction measures. The first is a simple unweighted sum of the four components. The second weights the four components according to their importance in determining logwagesinthepopulationoffull-timeworkersintheMCPS. DWSresults arereported forthesampleofworkersdisplacedbyplantclosurewithfull-timeworkatbothjobsand positive wage observations in both jobs, similar to Poletaev and Robinson (2008). The MCPS and CPS samples are also restricted to full-time employees.
The first two panels of Table 2 report the results for the full age range, 20–61, separately for males and females. Using the summary measures, the direction pattern, consistent with the discussion in Section II, shows a generally highly significant positive direction for total mobility, in contrast to a strong and negative pattern in the DWS. In terms of the individual components, for males, a strong contrast is the significantly negative change for “analytic” in the DWS (−0.1248) compared with the significantly positive change in the CPS(0.0193) and MCPS(0.0113). For females, there is also ahighly significant contrast for “analytic”: the positive effects in the CPS (0.0269) and MCPS (0.0419) are higher than for males and highly significant; the negative effect in the DWS is smaller than for the males and not statistically significant. However, the negative effect for females in the DWS is reflected in all components, with statistically significant decreases in both “fine motor” and “strength.”
The discussion in Section II suggests that total mobility reflects in part the life-cycle accumulation of skills, and hence, the strength of positive change will depend on age, as in standard human capital investment models. The remaining panels in Table 2 report results separately for younger (20–40) and older (40–61) workers. A large literature documents a slowdown in the rate of human capital accumulation with age and even some net depreciation. The age pattern of the estimates for total mobility are consistent with this. Both weighted and unweighted (positive) direction are larger for the younger workers in the CPS and MCPS. Despite this decline in accumulation with age in total mobility, the results for older workers show even stronger contrasts between involuntary and total mobility in the overall measures mainly because of the strong negative effect in DWS.
The significant exception when not all the components of S move in the same direction is for the sample ofyoung male displaced workers. A significantly negative change for “analytic” skill in the DWS is accompanied by a significant shift towards “strength.” This is strong enough to produce the same pattern in the pooled 20–61 male sample, even though it is absent (insignificant) for older males. A possible interpretation ofthis pattern is as follows. Prior to involuntary job displacement, some youngerto middle career workers who in an earlier occupation jused mainly strength orgross motor skills then acquired increased levels offine motor skills or supervisory skills, leading to a new optimal occupation match j'(promotion) where sj'has higher values of the components s1 and s2 and a lower value of “strength” factor s3 than occupation j. Following displacement, these workers still have occupation j'as their optimal match, but not are able to find it and, at least temporarily, move back down their own career ladder.25
V. Wage Losses and Displacement: The Importance of Direction
Table 2 shows strong evidence that the direction of change of the skill vector is negative for displaced workers in contrast to the positive direction in overall occupational mobility. This is very important for understanding the consequences of a given distance move for different types of occupational mobility. Poletaev and Robinson (2008) show a strong link between wages losses and discrete measures of occupation distance. Large wage losses were experienced only by displaced workers that significantly switched their skill portfolio. This section explores and documents the key role for direction in this result.
Table3reportsregressionresultsfortherelationbetweenlogwagechangesfollowing displacement and distance (edist4), with and without an interaction term (edist4 * neg) that interacts the distance measure with a negative direction indicator. The sample is the sameas for Table 2: workers displaced due to plant closure in the DWS 1984–2010. The dependent variable is the difference in the log usual weekly earnings on the current job, from those of the predisplacement job. To control for general year effects, including variation in unemployment rates, all the specifications reported includes a full set of survey year dummy variables. The remaining independent variables are dummy variables for age and education groups.26
For this exercise in distinguishing distance and direction, the distance measure, edist4, together with the direction indicator, neg, which is based on the unweighted direction measure, are used for Table 3, but the results are not sensitive to the choice of distance or direction measure. The use of the alternative relative distance measure, reledist4, produces almost identical results. The results are also insensitive to the inclusion of the number of weeks without work after displacement as a control.27
The results are quite striking. In all cases the estimated negative effect of distance for displaced workers is overwhelmingly due to the predominantly negative direction of occupational mobility following involuntary job loss. The effect of distance on the estimated wage loss is negative without an interaction term for direction. However, including an interaction term for direction shows a strongly negative effect for distance only when the direction is negative, and an insignificant effect otherwise. For example, forthepooledsampleofmales and females inthefirst column, theestimateforeffectof distance is −0.0450 with a standard error of 0.0092. When direction is taken into account, this drops to apoint estimate of −0.0066, which is insignificantly different from zero, while the interaction term shows a highly significant negative effect of −0.0771 for distance in a negative direction. Similarly, in the third column, the statistically significant effect of distance for full-time workers that switched occupation is −0.0291, while the fourth column shows that this overall negative effect is only due to the highly significant negative effect of −0.0756 for mobility in a negative direction. This pattern is the same for both males and females separately shown in the lower panels of the table.
The importance of direction for the relative magnitude of wage losses from displacement for occupation switchers compared to occupation stayers following plant closure is documented in Table 4. Dummy variables are defined for negative direction occupation switchers, nonnegative direction occupation switchers, and occupation stayers. The omitted group in the regression are occupation stayers.
Results are presented for the unweighted sum direction measure and the weighted measure using the relative importance of the skills in explaining wages. In all cases, individuals switching in a negative direction have significantly larger losses than those staying in their occupation. By contrast, when the direction is nonnegative, the losses for switchers are insignificantly different from the losses for those that stay in the same occupation. Overall, the results suggest that the specific human capital loss as reflected in mobility between occupations that are further apart in terms of their skill mix is important when the direction is negative, but not otherwise.
VI. Declining CoStS of DiSplacement?
In documenting an increase in occupational mobility in the United States, especially over the period 1969–85, Kambourov and Manovskii (2008) remark: “This rise in mobility may imply a substantial increase in the destruction rate of human capital.”28 More broadly, the previous literature links the potential loss of specific capital from occupational mobility to a number of factors: the magnitude of displacement, the amount of occupational switching following displacement as well as total occupational mobility, and the distance of the switches. The argument in this paper is that, given the evidence on the importance of direction for specific capital losses, and the difference in direction for displacement compared to total mobility, the most important source of potential losses is from displacement. This section uses estimates of displacement, occupational mobility, and distance over the period covered by the 19842010 DWS surveys to shed light on the path ofthe potential costs to the economy from specific capital losses following displacement.
The overall potential loss of specific capital to the economy due to involuntary job mobility depends on the fraction of workers displaced, the fraction of these displaced 28. Kambourov and Manovskii (2008, p. 56).
workers that switch occupation, and on the distance of the displaced workers with occupational mobility in a negative direction. Over the 1984–2010 period, the fraction of negative direction moves following displacement shows no trend. Thus, the potential for specific human capital losses generated by mobility following involuntary job loss is given by d(t)a(t)Dc(t) = d(t)D(t), where d(t) is the fraction of workers displaced in period t, a(t)is thefractionofdisplacedworkers that switchoccupationinperiodt, and Dc(t)is the average distance of the displaced worker conditional on switching occupation. (D(t) is the average distance for all displaced workers, including those who stay in the same occupation.)
The previous sections used consistent occupation coding with 494 modified 1990 census occupation codes for the DWS surveys of 1984–2002, but changed to 2000 coding for 2004–2010. For the cross-section ,or pooled cross-section analysis, the break in occupation coding is a minor problem. However, for a secular analysis to examine changes over time in mean distance, the break is a significant problem. There is a clear break in the series using these data for distance starting with the 2004 DWS. To reduce the problems arising from this break, this section uses an occupational coding scheme from IPUMS, denoted occ1990, that is consistent over the full period. These occupation codes are based on the 1990 census codes, but are reduced in number to provide better consistency for the original census 1980–2000 coding periods. The occ1990 codes provide the major benefit of a series for distance that shows no break, at the modest cost of a 20% reduction in the maximum number of occupations (384 vs. the 494 modified 1990 occupations used in the previous sections).29
Table 5 reports estimates of d(t)D(t) and its separate components: the fraction of workers displaced in period t [d(t)], the fraction of displaced workers that switch occupation in period t [a(t)], and the average distance of the involuntary move of the displaced worker conditional on switching occupation [Dc(t)] for displacements following plant closure for the DWS surveys for 1984–2010 using the consistent occ1990 coding.
Three-year displacement rates for exogenous involuntary job loss due to plant closings are estimated from the DWS data. There are several problems ofcomparability over time discussed in detail in Farber (2004). The respondents in the DWS are asked if they were displaced from a job in the last three years (surveys 1994–2002) or five years (surveys 1984–1992). One way of dealing with this is to restrict observations to displacements in the last three years for all survey years. However, as Farber (2004) notes, there remains a comparability issue due to the fact that respondents in the 1984–1992 surveys were asked to report the displacement from the longest job over the last five years. In the case, for example, where there was a displacement three years ago and another five years ago, dropping this observation would undercount displacements in the last three years for the earlier surveys. Based on evidence from the Panel Study of Income Dynamics, Farber (2004) suggests an upward adjustment ofabout 11% in the estimated rates for these surveys.30
The pattern of job loss in Table 5 shows a decline from job losses up to the 1990 survey (representing displacements in 1987–89). The job loss rate jumps up for the 1992 survey, representing displacements in 1989–91, and thus affected by the onset of the recession in 1990. After the effects of this recession, job loss rates begin a long decline to the recession in the early 2000s. With the onset of the recession in the early 2000s, job loss rates again rise.31 Thus, the time path of involuntary job loss has a cyclical pattern, similar to that noted in Farber (2005) for the period up to the 2004 survey, though no clear secular pattern.32 The cyclical pattern continues after the recovery from the early 2000s recession, with a large increase corresponding to the 2008 recession.
The time pattern for the average distance of the involuntary move of the displaced worker conditional on switching occupation [Dc(t)] for displacements following plant closure reported in Table 5 shows a clear decline until the onset of the early 2000s recession for both males and females.33 For males, there is a large jump in the 2002 survey, and thereafter there is no clear trend. The distance falls inthemid2000s after the recession, but increases again with the onset of the Great Recession. For females, there is no fall in the mid 2000s and a large effect for the Great Recession. These patterns are explored in more detail in a regression analysis including controls for age and education.34
Distance was regressed on a full set of survey year dummies, as well as controls for age and education at the level of the individual displaced worker using samples restricted to displacement through plant closureas well as samplesusingthebroaderBLS definition of displacement. Regressions were estimated separately for males and females. The results provide strong evidence of a secular decline in the expected distance between pre- and postdisplacement occupations for male and female workers displaced by either plant closure or using the broader BLS measure up to the onset of the early 2000s recession (1984–2000 DWS surveys). The omitted survey year dummy in the regressions is 1984, and the dummy variable for the 2000 survey is always significantly negative, at a high level of statistical significance. The full regression results are reported in Appendix Table A1.
The predicted expected distance from the regression analysis, evaluated at overall sample mean values for the age and education controls, are plotted in Figure 3. Males experience higher distance following plant closings on average than females, but both show similar percentage declines overtheperiod of the 1984–2000 DWS surveys and a reversal thereafter. There is also evidence of cyclical effects, indicating higher distances in the surveys influenced by recessionary periods.
The strong decline in distance over the 1984–2000 period is largely confined to occupational mobility following displacement. A comparable analysis for total mobility, reported in Robinson (201 1), shows that while the patterns for education and age, and the differences between males and females are similar to the patterns for mobility following displacement, there is no evidence of a significant downward trend in distance for total mobility. This contrast with theDWS suggests that over the 1984–2000 period displaced workers were able to find their way back to increasingly similar jobs, but promotions, job experimentation, etc. were occurring at similar rates and in similar magnitude across the period.
One possibility forthedeclinefordisplacedworkers is that the overall mean distance between jobs has declined over time, so that there are more “close” jobs for workers to move to after displacement. To the extent that workers increasingly expect to have to change jobs more frequently in their lifetimes than in the past, they would have an incentiveto choose skill portfolios orjobs that are not too far away fromotherjobs than workers in the past who could specialize more in very specific jobs. The distance distribution from random mobility among the 494 modified 1990 occupations uses the occupation employment weights from 1992. Robinson (2011, Table 5) shows limited evidencethatthereisashifttocloseroccupationsonaverage,butthemagnitudeissmall. Thus, almost all of the expected distance decline is due to displaced workers on average finding new jobs that areclosertotheiroldjobsthan theycouldbefore, whilethedensity of “close” jobs to the ones they lost remained approximately the same. Moreover, there is little indication of the strong decline in mean distance over the period that can be picked up in the path of major group switching, particularly for males (Robinson 2011).
The reason for the decline in distance in the 1980s and 1990s is unclear. The share of college graduates and older workers in the workforce, and in the displaced worker samples, have increased over time, and these groups, particularly for males, show lower average distances. However, the decline is robust to the inclusion of these individual characteristics and to industry controls. Thus, it does not appear to be due to changes in the type of workers being displaced or the industry from which they are displaced, suggesting that the labor market may have been more efficiently reemploying workers following displacement during this period.35 Additional support for this interpretation is that, coincident with the decline in distance for 1984–2002, there was also a decline in weeks without work following displacement over the same period.36
With the onset of the early 2000s recession, mean distance increases from its 2000 low. For males, it declines somewhat during the recovery periods of the mid 2000s but increases again, especially with the onset of the 2008 recession. Females show no evidence of recovery in the mid 2000s, but like the males, they experience increased distance during the 2008 recession. The onset of the recession in the early 2000s and the Great Recession starting in 2008 reverses the pattern of decline of the earlier period in both distance and weeks without work following displacement. It is too early to tell if the decline seen in the 1980s and 1990s will continue after a full recovery from the Great Recession.
Section V established the link between expected distance and wage losses for displaced workers provided the direction was negative. Actual wage changes in the DWS, adjusting for changes in age and education composition, are plotted in Figure 4.37 While specific human capital losses are only one component of wage losses following displacement, the actual losses up to the early 2000s recession do show evidence of a secular decline, mirroring the decline in distance. They then reverse, coincident with the distance pattern, and show particularly large losses at the same time that distance increases substantially with the onset of the 2008 recession.
The drop in expected distance up to the 2000 recession for males and females relative to their respective means is around a 13% drop over a relatively short period. The standard deviation of the displaced worker distances is about 0.5, so this represents about a quarter of a standard deviation point. The fraction of negative moves among displaced workers did not increase over the 1984–2002 surveys; thus, other things the same, the decline in expected distance may have reduced the costs of displacement due to specific human capital losses.38 The results from Table 3 suggest that this distance drop would reduce the average wage losses for displaced workers.
How important was this effect? Applying the estimates in Table 3 implies that the observed decline in distance, conditionalonswitchingoccupation, accounted for 4% of the drop in average wage losses for full-time workers and 10% for all workers. This implies a 7% lower average wage loss for full-time workers and an 8% lower average wagelossforallworkers.Unfortunately,thisisreversedaftertheearly2000srecessions, and especially in the period of the Great Recession; and it is unclear whether average distance will return to a secular decline again after a full recovery from the Great Recession.
VII. Summary and Conclusions
Recent research has linked occupational mobility and specific human capital loss. This paper develops occupation distance and direction measures that are used to provide a better understanding of this link. Wage losses from occupational mobility following involuntary job displacement are strongly related to distance. However, the negative wage consequences of increasing the distance between the skill mix in the pre- and postdisplacement occupations are only important when the direction is negative. A comparison with total mobility shows quite similar distances, but markedly different direction. On average, displaced worker mobility is in a significantly downward direction. In contrast, on average, total mobility is in an upward direction, and the difference between the two types of mobility is strongly significant. This has important consequences for the link between occupational mobility and specific human capital loss.
For occupational mobility across occupations that are “close,” there is little potential for loss of specific human capital.For occupational mobility following involuntary job loss, there is potential for specific human capital loss, but it is only important for mobility in a negative direction in terms of the skill vector. A large fraction of total occupational mobility is voluntary. The average upward direction of this mobility suggests that specific capital losses here are less than for involuntary mobility. Total mobility reflects to a large extent the life-cycle evolution of the skill vector. This may for certain career paths involve changes in the relative importance of the individual components and possible specific capital losses due to this, but also has a generally upward direction reflecting the life-cycle accumulation of various types of human capital or skills rather than losses.
In the 1980s and 1990s, displaced workers increasingly found new jobs that were close to their previous job, suggesting that the labor market may have been more efficiently reemploying workers following displacement over this period. The 2000s wereheavilyinfluencedbytherecessionoftheearly2000sandtheGreatRecession,and declining distance was reversed. The distance measure for these workers may be useful in assessing how well the labor market matches workers and firms both over cycles as well as over time. Distance measures for voluntary occupational mobility where on average direction is positive may reflect, instead, the typical “step size” in career paths in terms of skill acquisition and promotions. The evidence on the strong difference in direction of total occupational mobility compared to mobility following involuntary job loss and the important role of direction in determining losses for a given distance move for displaced workers indicates the need for future research to better understand the nature and measurement of the underlying skill vectors and their differential evolution through different forms of occupational mobility.
Appendix
This appendix describes the construction of the skill vectors at thelevel of three-digit census occupation codes, random mobility benchmarks for the distance distributions, and consistency across occupation coding schemes. It also details the sample restrictions for the DWS and MCPS and discusses measurement error.
A Construction of the Occupation Skill Vectors
The strong similarity of the 1980 and 1990 census occupation codes allowed the con- structionofamodified 1990 census occupation coding scheme that provides consistent occupation coding for the DWS for all the surveys from 1984 to 2002. The correspondence between the original 1980 and 1990 codes and the modified 1990 coding is given in Robinson (2011, Appendix Table A1). The occupation skill vectors are constructed using a factor analysis to identify four basic “skills” or factors from the DOT data on 49 job characteristics. The observations are workers in the January 1992 CPS file. Each worker is assigned a value for 49 DOT characteristics based on their sex and modified 1990 occupation code.39 This requires calculation of means for each DOT characteristic for each of the modified 1990 occupation codes.
The final DOT master file (version 4.3) has 12741 unique DOT codes. Each DOT job has characteristics associated with it. Three characteristics come from the fourth, fifth, and sixth digits of the DOT code itself: the complexity of the interaction with data, people, and things. This provides a numerical ranking of the complexity. All other characteristics are provided in separate fields in the DOT master file. The master file also acts as a crosswalk between the 12,741 unique DOT codes and 1990 census three-digit occupation codes by assigning an occupation code to each DOT job code. Averaging the DOT characteristic scores over the DOTjobs in each 1990census three-digit occupation code in the crosswalk to compute the mean characteristic scores by occupation implicitly assumes that employment within a three-digit occupation is equally divided across the DOT jobs for that occupation, which is not in fact the case. An alternative approach is to use DOT level employment information made available in a special CPS monthly file for April 1971 to construct weights for DOT job employment. This special dual DOT and three-digit occupation coded file is the DOT augmented CPS monthly file (April 1971), available at ICPSR as project 7845. The study description is as follows:
The April 1971 CPS was coded routinely with 1970 occupation codes by census personnel. The occupational descriptions from this CPS were also coded with the 9 digit 3rd edition (1965) DOT codes and added to the tape by the staff of the Division of Occupational Analysis of the U.S. Employment Service. (Lloyd) Temme cleaned and edited the data. Using a map which relates third and fourth edition DOT codes (created by the Division of Occupational Analysis), he then added third edition GED, SVP and worker trait group scores as well as all fourth edition (1977) DOT occupational characteristics to the tape. In several instances there was no equivalence between the third and fourth edition codes, either because the title had been deleted between the two editions, or multiple fourth edition codes were created from a single third edition code. Temme reviewed the non matches on a case by case basis, assigning fourth edition codes by hand where possible. If no match could be made, 9's were assigned to the DOT codes. On the computer file there are 6984 persons with no information on fourth edition DOT codes.
The special CPS April 1971 file contains 1970 occupation codes, and the DOT information in the file is for the original fourth edition (1977) DOT characteristics scores. However, the revised fourth edition (1991) scores are most appropriate for the period of the analysis. Unfortunately, the procedure for using the revised characteristic scores together with the original fourth edition codes available in the April 1971 file is not straightforward because of a complicated relation between the original and revised fourth edition DOT jobs, especially the fact that the three middle digits of the job code are, in fact, ratings on the three characteristics, “data,” “people,” or “things,” so the same job gets a new code if the rating changes. The special file in fact has 3,885 unique DOT (1977) codes assigned to the 53,457 individuals in the file. This is a subset of the total number of unique original revised fourth edition DOT codes, which is over 12,000. However, since the CPS special file is a large representative sample, the unused codes from the 12,000 must have very low employment. These 3,885 original (1977) DOT job codes have original characteristic scores assigned to them by those responsible for creating the special file. (These are in almost all cases the same as the values in the DOT fourth edition original 1977 crosswalk.) The task is to trace these 3885 1977 DOT jobs into the 1991 DOT edition, which records the revised characteristic scores.
Unfortunately, the jobs cannot be traced by the codes because the codes can change for two reasons: (i) the middle three digits of the code have to change if there is an updating to the rating on any of the three characteristics, “data,” “people,” or “things,” and (ii) some codes were changed or merged or divided or dropped. Thus, it is necessary to find the DOT jobs (codes) in the revised edition that correspond to the 3885 DOT jobs in the April 1971 file when the job is the same, but the code may have changed. The revised fourth edition DOT master file containing the final updating for all the DOT characteristics contains the variable dlu, which indicates the last date on which the code or characteristics were updated. The file was examined for duplicates on the DOT code alone and on the DOT code and DOT title together, before construction of a file with unique codes withrevised (1991) characteristic scores. Similarly, theApril1971 filewas examined for duplicates on the DOT code alone and on the DOT code and DOT title together, before construction of a file with unique codes with original (1977) characteristic values that appear in the special 1971 CPS file.
Examination of dlu shows 10,289 of the 12,741 unique DOT jobs in the revised fourth edition had not been updated since 1977. These should therefore be codes that are the same in both the original and revised editions with characteristic values that are unchanged between the original and revised editions. The remaining 2,452 DOT jobs intherevisedfourtheditionmaybeneworrevisedcodesorthesamecodeswithtrailer characteristics updated.40 The next step was to identify which of the original 3,885 codes in the April 1971 file needed updating to the equivalent job code in the revised edition, and those that do not. There were 2,715 codes that represented DOT jobs that were unchanged between 1977 and 1991. The characteristic values for these in the April 1971 file should be the same as in the 1991 master file.41 The titles from both original and revised editions were compared for these 2,715 codes. For 2,563 of the 2,715, the titles are identical. Inspection of the remaining 152 shows that almost all titles differ only in the use ornonuse of hyphens or roman numerals, confirming that they are the same job.
The1,170codes that appearonlyin the April 1971 file andnotinthemasterfilehave to betraced into the revised 1991 edition. Autor et al. (2003) used the DOT materials in the ICPSR study (6100) that describe the relation between the original DOT fourth edition and the revised fourth edition. The ICPSR project 6100 documents the changes between the original fourth edition and revised fourth edition. The relevant section is “Conversion Tables of Code Changes Fourth to Revised Fourth Edition Dictionary of Occupational Titles.” This provides tables containing “occupations with a DOT code and/or title which changed or was deleted between 1977and 1991 or appeared since the publication of the Fourth Edition in 1977.” The occupational analysts revised 646, combined 136, and deleted 75 occupational codes and titles, so these 875 DOT codes and/ortitles donotappearin the revised fourth edition, butareidentifiedin thesetables. In addition, between 1977 and 1991 analysts reviewed, updated and/or added 2,452 occupational elements to the dictionary. These were too extensive to tabulate, but can be identified by the “Date of Last Update” (dlu) variable in the DOT data files. For these cases,dluwillhaveavalueotherthan“77.”(IfollowAutoretal.[2003]inassumingthat any job that was not revised experienced no task change.) While there were 2452 “reviewed, updatedand/oradded2452occupational elements,”itwasonlynecessaryto trace the relevant 1,170 1977 DOT job codes in the April 1971 file.
First, “Conversion Tables of Code Changes Fourth to Revised Fourth Edition Dictionary of Occupational Titles” was used to identify which of the relevant 1,170 codes hadatitle orcode change between the editions. Of the 1,170 codes, 265 were identified in Table 1 “Occupational Code and/or Title Changes,” with 181 having code changes and 85 having only title changes. This center 905 codes still to be traced into the revised edition. Of the 905 remaining codes to be traced, 25 codes were identified from Table 2
“Occupations Deleted From or Combined with Another in the Revised Fourth Edition DOT” as deleted and 52 as combined. This center 829 original codes to be traced into the revised edition. The variable dlu identifies that they are codes that were reviewed and possibly updated after 1977. Most of these may have identical codes in the revised edition, where the “data,” “people,” and “things” characteristic scores stayed the same (otherwise the code would have changed), but some of the “trailer” values changed. To identify these cases, the 829 codes were merged with a file of the 2,452 unique revised 1991 DOT codes with dlu values indicating an update after 1977.
The merge shows 827 common codes. Of these, 784 (94.8%) also have the identical title. The remaining 43 do not have the same title, but examination shows that the only differences in the titles are in spelling, hyphens, etc. Thus, all these original fourth editioncodes are the same jobs as the revised fourth edition codes, where the code is the same because the revision did not affect the “data,” “people,” or “things” rating, and hence the three middle digits stayed the same. In summary, of the 3,885 original codes from the special 1971 CPS file, almost all can be traced into the revised edition. 2,715 are the same job with the same code and title and so are all directly traced. Of the remainder, all but 25 (deleted) can be traced.
Examination of the April 1971 file shows that these 25 codes make up less than 0.2% of the population. The revised DOT characteristic score means by 1990 occupation are constructed using two approaches. The main approach is a “weighted crosswalk” approach. First, from the final version of the DOT master file (4.3) a file was created with 12,741 unique DOT codes, the 1990 census occupation code and the revised DOT characteristics scores. This provides a single 1990 census occupation code for each DOT code.42 Second, revised DOT job level employment weights (separately by sex where possible) were assigned to this file from the information on the original DOT codes in theApril1971 file that were traced into the revised codes. Third, this augmented master file was used to calculate weighted mean revised DOT characteristic ratings by 1990 or modified 1990 occupation. The total number of 1990 occupations for which means can be calculated using this approach is a little less than the total of 501. The missing occupations are primarily postsecondary teaching occupations by specialty that do not have DOT codes in the master file. Most of these, however, are available from a “dual occupation codes files” approach.
The dual occupation coded files approach starts with the means by the 1970 occupation coding directly from the special April 1971 file. (The means were calculated separately for males and females where possible.) There is no simple correspondence betweenthe1970and 1980 occupation codes, so a simple occupation code crosswalk is not available to convert these to means by 1980 codes. Instead, the special sample from the 1970 census that was retrospectively dual coded with 1980 codes, known as the “Treiman File,” was used. First, the predicted revised characteristic scores were assigned to the individuals in this file on the basis of their sex and their 1970 occupation code. Next, modified 1990 codes were assigned to individuals in this file on the basis of their 1980 code. Revised characteristic score means by modified 1990 code were then estimated by averaging (separately by sex) over individuals in each modified 1990 code.
B Random Mobility Benchmarks
Assume a fixed distribution of occupied jobs (employees) by occupation at time t and t + 1. Displace a person at random from a job in this distribution in t , and assign the displaced person at random to another job in t + 1. The expected distance for a displacement from occupation j in t is a weighted sum of the distance between occupation j and the occupation picked at random in t + 1, where the weights are the employment shares of the occupations. The overall expected distance is a weighted sum of these expected distances from occupation j, where where the weights are, again, the employment shares of the occupations.
Denote j as an occupation in t and j in t + 1. Define d( j, j ) as the distance between occupations j and j , and o( j) as the fraction of jobs (employees) in occupation j. The overall expected distance is:
(5)where J is the total number of occupations. The number of possible pairs, d( j, j) is J · J. In a population with equal employment in all occupations, the expected distance is the simple mean of the J · J distance terms d(j, j). If employment is unequal across occupations, it is a weighted mean, where the weights are the product o( j)o(j ).
The primary analysis in this paper uses 494 modified 1990 occupation codes in total. From the factor analysis, most of these (406) were assigned skill vectors separately for males and females, so distances between the same pair of occupations may be different for males and females. The expected values of the distance measure for a population of random movers between each of the possible 494 occupations (including moving to the same occupation) using the actual employment weights from the 1992 sample used in the factor analysis are 1.29 for males and 1.27 for females. The maximum distance is 3.46 for males and 3.40 for females. Overall, the frequency distribution of distances for males and females are very similar. Less than 1.5% of the random mobility results in staying in the same occupation and only 6.8% ofthe mobility formales and 7.3%for females involves distances less than 0.5.43
C. Empirical Evidence on Measurement Error
The mean distance from random mobility using the 1970 codes reported in Robinson (2011, Appendix Table A2) shows a maximum of3.59 and mean of1.28 for males and a maximum of 5.25 and a mean of 1.26 for females. The conditional and unconditional (on switching) distributions are very similar since there are very few stayers in random mobility. If miscoding in observed data was random with respect to closeness of the occupations, the estimated distance distributions should show something similar. Alternatively, if miscoding leading to switches primarily occurs between close occupations, the mean distance should be smaller. Evidence from the January 1977 CPS Validation Survey suggests that this alternative is true. This survey provides two coded sources of what should be the same 1970 three-digit occupation code, one based on an employee response in the basic CPS survey and one based on an employer response in the validation supplement. Comparison oftheresponses for the matched sample show a lot of false differences between what should be the same occupation coded from the employer and employee responses, consistent with previous evidence on large overestimate of switching due to miscoding.For males 42.74% are coded differently, andfor females 41.73%, when there should be zero. Clearly, for analysis of switching, this miscoding is a major measurement error problem. However, for distance analysis, there is much less of a problem. In contrast to the unconditional random mobility mean distance of 1.28 for males, the validation data comparison shows an unconditional mean ofonly0.348 and even aconditionalmean of only 0.813. For females, in contrast to the unconditional random mobility mean of 1.26, the validation data shows an unconditional mean of only 0.292 and even a conditional mean of only 0.701.44 To put this in perspective, as reported in Table 1 the random mobility mean distance within three-digit occupations across DOT jobs is about 0.6, so in the validation data the miscoding is largely between occupations that are as close as jobs within three-digit occupations.
D. Sample Restrictions for the DWS, MCPS, and CPS
The distance distributions for the DWS are estimated for workers displaced from the private sector. The only other sample restrictions were the requirement for valid occupation codes for pre- and postdisplacement jobs and displacements restricted to the last threeyears.TheDWS analysis of distance and direction in Sections IV and V restricts the sample to full-time private sector workers with positive current and past weekly earnings observations and excludes displacements from construction and agriculture. It also drops workers 62 and older. This approximates the sample in Poletaev and Robinson (2008).
The data for total occupational mobility come from the MCPS and matched individual records across adjacent months in the CPS. The MCPS records a three-digit occupation codeforemployed workers in their current job in (the third week of) March of the survey year and a three-digit occupation code for the longest job held in the previous year (“earnings year”). The main sample restrictions follow Moscarini and Vella (2003). This restricts the sample to past and current employees and excludes observations after the 1988 survey in which all of the supplement responses were allocated, including occupation code. The allocation “ag” for current occupation shows afraction around 1% or less up to 1988. The 1989 survey year was the first with anew system of treating allocated values, especially for the annual supplement, with the introduction of the variable suprec indicating allocated supplementresponses.From this pointthefractionallocatedforcurrentoccupationincreases,rangingbetween36%upto earnings year 1997. For the earnings years 1998 to 2001 the fraction increases to over 11%. The allocation ag for occupation in the longest job last year is not available until the survey year 1989. After that, the allocation fraction is extremely low, but this is because this only refers to allocated values in supplementary records where the whole record was not allocated. The suprec variable shows that starting in 1989 nearly 10% of supplementary records were allocated, including occupation. The only additional restrictions are the age range 20–64 and that the longest job last year was in the private sector, for comparison with the DWS. The sample from the matched files in the monthly CPS is restricted to records for which there are occupation codes for the current and past month. The remaining restrictions match those for the MCPS.
E. Estimates of the Time Path of Conditional Distance
The time path for the expected value of the distance between the pre- and post- displacementoccupations,conditionalonswitchingoccupationplottedinFigure3,was estimated by regressing the distance measure for the sample of displaced workers who switched occupation on a full set of survey year dummies, a dummy variable for female, dummy variables representing three of four education groups (dropouts, high school graduates, some college, and college graduates) and two of three age groups (20–29, 3039, and 40–64). Appendix Table A1 reports the regression results.
The secular declines in distance between pre- and postdisplacement occupations for male and female workers displaced by either plant closure or using the broader BLS measure up to the onset of the early 2000s recession (1984–2000 DWS surveys) are highly significant. The omitted survey year dummy is 1984, and the dummy variable for the 2000 survey is always significantly negative, at a high level of statistical significance.
F. Time Path of Weeks without Work after Displacement
Appendix Table A2 reports the value of the cumulative distribution (%) of weeks without work for workers age 20–64 displaced due to plant closure at four points: 0 weeks, 12 weeks, 26 weeks, and 52 weeks.
Acknowledgments
Special thanks are also due to David Autor for making available a copy of the Treiman file with dual 1970 and 1980 occupation coding. This work is supported by the Human Capital and Productivity Centre at the University of Western Ontario. The author notes that data used in this paper are almost entirely from readily available public sources, and he will assist with any data access inquiries
Footnotes
Chris Robinson is a professor of economics at the University of Western Ontario. He wishes to thank participants at workshops at the University of Michigan, Yale University, and Michigan State University, as well as Craig Riddell, discussant at the Canadian Economic Association Meetings in Quebec City, May 2010, and participants at the RCEA Conference in Rimini, Italy, June 2010 for comments on related preliminary working papers
↵1 See, for example, Poletaev and Robinson (2008) and Gathmann and Schonberg (2010).
↵2 By the nature of occupation coding, occupational mobility is a discrete measure: all occupation moves are equal. The coding classifications do have some grouping levels, so that a distinction can be made between moves within groups and across groups. However, the analyses in this paper using U.S. data, and in Robinson (2009), using data from the 2006 U.K. Skills Survey, show that grouping occupations under the higher level occupational classifications is very far from the grouping that is produced using a distance measure based on skill or task vectors.
↵3 This literature is related to the broader literature on human capital specificity and occupational mobility that goes back for many years. The empirical work in this broader literature is based on datasets that record standard occupation and industry codes. Changing firm, industry, or occupation meant some kind of change in the skill set used or tasks performed. However, this literature bypassed the issue at the level of skills or tasks and instead focused on firm, industry, and occupation tenure, as measured by the length of time a worker spent with a firm, industry, or occupation. It assumes that firm-, industry-, or occupation-specific capital is an increasing function of tenure. Prominent examples are Neal (1995), Parent (2000), and Kambourov and Manovskii (2009).
↵4 This is very similar to the firm-specific concept of human capital in Lazear (2009) where skills are transferable across firms, but firms use them in firm-specific bundles. The recent literature taking this approach in the context of occupations includes Yamaguchi (2012); Autor and Handel (2013); Firpo, Fortin, and Lemieux (2013); Gathmann and Schonberg (2010); Speer (2017); and Guvenen et al. (2015). In an earlier paper Heckman and Sedlacek (1985) specified a Roy model of comparative sector advantage in which a worker's skills generated certain sector-specific tasks via sector-specific task functions in the context of industry sectors.
↵5 In essence, the occupation-specific markets are thick enough, in the sense of Lazear (2009), to warrant the assumption that workers are paid their marginal products.
↵6 Firpo, Fortin, and Lemieux (2013) specify a similar wage equation and provide a more detailed discussion of “bundling.”
↵7 As Yamaguchi (2012) notes, in his use of this framework with a continuum of task complexities and a large number of occupations, there is no truly occupation-specific human capital.
↵8 Yamaguchi (2012), p. 5.
↵9 An alternative approach would be to consider occupations as similar if they have similar y. However, Autor and Handel (2013) discuss the problems that arise in estimating y, citing Heckman and Scheinkman (1986) and Heckman and Honore (1990), and instead focus on measuring the skill vector, s.
↵10 Younger workers may still be searching or experimenting, but these are a minority in the data.
↵11 This permits a more flexible application and extension of a related approach in Poletaev and Robinson (2008), where the similarity was defined primarily on the basis of whether a worker switched the “main” skill.
↵12 See, for example, Davis and von Wachter (2012).
↵13 Gathman and Schonberg (2010) take this direct approach.
↵14 Ingram and Neumann (2006) and Poletaev and Robinson (2008) take this indirect approach. See Robinson (2009) for more discussion of the alternative approaches.
↵15 The factor loadings from the factor analysis are almost the same. The factors are also similar to those computed by Ingram and Neumann (2006) in their study of skill pricing.
↵16 Since the occupational distributions for males and females have become more similar over time, with females entering previously exclusively male or male-dominated occupations, the same may have happened at the DOT job level within occupation. Given that the different characteristic score means within occupations can only be calculated using DOT level employment information by sex from the special 1971 file, caution is needed in interpreting the female factor scores applied to more recent data since the true underlying characteristic score means for females may have moved closer to the score means for males. However, the main results in the paper are robust to using the same task values for both as in Poletaev and Robinson (2008).
↵17 This is no longer invariant to rotation, though for the range of weights used in this paper the measures are in fact quite insensitive to rotation.
↵18 See the Appendix for details.
↵19 The IPUMS data were downloaded from IPUMS-USA, University of Minnesota, www.ipums.org (accessed 17 Aug. 2017). The IPUMS-USA is described in Ruggles et al. (2015).
↵20 See Kambourov and Manovskii (2008, 2010) and Moscarini and Thomsson (2007) for a full discussion, including the distinction between dependent and independent coding and further references.
↵21 Construction of the random mobility benchmarks is described in detail in the Appendix.
↵22 The random mobility benchmark allows mobility across all occupations, many of which are extremely unlikely to occur in the data. An alternative experiment that compares random mobility within higher and lower skill occupation groups to moves in the data similarly confined within these broader occupation groups produces the same contrast.
↵23 An assumption is required on the (unknown) fraction of occupation and industry stayers that changed jobs in the MCPS to estimate the unconditional distribution. It is not required for the conditional distribution. See Robinson (2011) for more discussion and sensitivity of the unconditional distribution alternative assumptions.
↵24 The matched CPS data are restricted to the period of dependent coding.
↵25 Some support for this is that common current occupations for above the mean increases in s3 for males under 40 are truck drivers, driver sales workers, freight stock and material handlers, and laborers of all types. The most common previous occupation for these switchers is miscellaneous machine operators, but other common previous occupations are supervisors in sales occupations and in production occupations.
↵26 Age is included as a set of age dummies, 20–29, 30–39, 40–61; education is represented by dummy variables for the four education groups: dropouts, high school graduates, some college, and college graduates. The specification and restriction to plant closures is similar to Neal (1995) and Poletaev and Robinson (2008). This is commonly used as an approximation for exogenous involuntary job loss. However, the basic results are similar for the larger BLS definition. Education itself is usually insignificant, but there is a strong negative effect for older workers, reflecting a common result in the previous displacement literature. The attempts in Neal (1995) and Poletaev and Robinson (2008) to deal with endogeneity issues in the postdisplacement job did not alter their OLS estimates much. However, since the selection correction techniques employed in Neal (1995) and Poletaev and Robinson (2008) are far from perfect, the results in Table 3 should be interpreted with some caution.
↵27 Table 3 excludes this control as it is not available for the 1994 survey which reduces the sample size and precision.
↵28 Kambourov and Manovskii (2008, p. 56).
↵29 Crosswalks between the consistent IPUMS variable, occ1990, and each of the 1980, 1990 and 2000 census codes were used to assign values for occ1990 to individuals in all the DWS surveys. The skills by occ1990 code were constructed by first assigning skills to all the individuals in the CPS files covered by modified 1990 coding in the same way as for the previous analysis and then collapsing the files, separately for males and females, by the using occ1990 codes.
↵30 See Farber (2004) for more details and a full discussion of the comparability issues.
↵31 Farber (2004) uses a broader definition than the BLS, and reports the patterns for different age and education groups. The same basic pattern is seen in Figures 1–3 of Farber (2005) across age and education groups. (There is a minor difference between the 1992 and 1996 surveys where the figures in Farber (2005) show a small upward movement between 1994 and 1996.)
↵32 Farber (2005) argued that there was no clear evidence of a secular pattern up to the 2004 survey, though noted that while job loss rates among older and more educated workers did decline after 1995, they remained higher up to the 2004 survey than they were at the peak of the 1980s expansion, possibly due to restructuring.
↵33 This decline is also clearly apparent from a comparison of the cumulative distributions for 1984 and 2000. Figures 5 and 6 of Robinson (2011) show that the cumulative distribution for 2000 for displaced males who switched occupation is well to the center of the distribution for 1984, such that while 41% of the occupation switchers in 1984 have di < 1, this increases to 53% in 2000.
↵34 There is a significantly lower distance for older workers and for male college graduates.
↵35 Given the evidence of increased polarization and obsolescence of some skills, one concern is that some displaced workers facing a high probability of a large distance move may have center the labor market. Overall, there is a small increase in nonparticipation in the CPS data for all groups, but the changes are very similar for displaced and nondisplaced workers.
↵36 Appendix Table A2 describes the cumulative distribution of weeks without work for workers displaced by plant closure in the DWS 1984–2010. For both males and females, there is a clear shift to the center over the 1984– 2002 period. For example, for displacements in the 1984 survey about 50% of both males and females were without work for t12 weeks or less, while for the 2002 survey this increases to 83% for males and 74% for females. For a small fraction of displaced workers in the DWS a distance measure cannot be computed. These are observations for workers that are currently unemployed. The displacement may have been recent or up to three years before, and the worker may never have worked since displacement or may have worked again after displacement but be currently unemployed for reasons other than plant closure or BLS displacement reasons. This fraction is constant over the period.
↵37 The log wage difference following displacement is regressed on a full set of dummy variables for survey year and controls age and education using the same specification as for Figure 3.
↵38 This assumes no increase in the magnitude of the relation between (negative) distance and wage loss over the same period. In fact, allowing for change in the effect of distance over time shows no significant differences.
↵39 See Poletaev and Robinson (2008) for more detailed description of the 49 characteristics. Over time there were changes in analyst ratings for the DOT level jobs. The final ratings appear in the revised fourth edition of the DOT dated 1991. These are more appropriate for the period of analysis in this paper than the earlier ratings from the original fourth edition of the DOT, dated 1977.
↵40 The trailer characteristics are all the characteristics except for the three complexity scores for “data,” “people,” and “things”.
↵41 In fact, using the GED characteristics scores to test this, 98.9% are the same, but 30 cases are different by 1.
↵42 The crosswalk assignment of the 1990 census codes is done by the crosswalk center on the basis of the common underlying SOC codes.
↵43 The frequency distribution of distances are given in Figures 1a and 1b of Robinson (2011). The employment weights are for total employment in the occupation. Thus, the random experiment allows males and females to move with probabilities equal to total employment shares, not sex specific shares.
↵44 There is also considerable skewness with lower medians and a long thin upper tail.
- Received August 2014.
- Accepted January 2017.
References
In this issue
Jump to section
- Article
- Abstract
- I. Introduction
- II. Occupations, Skills, and Specific Human Capital
- III. Construction of the Distance and Direction Measures
- IV. Occupational Mobility: Patterns of Distance and Direction
- V. Wage Losses and Displacement: The Importance of Direction
- VI. Declining CoStS of DiSplacement?
- VII. Summary and Conclusions
- Appendix
- Acknowledgments
- Footnotes
- References
- Figures & Data
- Info & Metrics
- References