«IZA DP No. 3067 How General Is Human Capital? A Task-Based Approach Christina Gathmann Uta Schönberg September 2007 Forschungsinstitut zur Zukunft ...»
where qjo is the fraction of workers using task j in occupation o and qjo0 is de…ned analogously. This measure de…nes the distance between two occupations as the cosine angle between their positions in vector space. The measure has been used extensively in the innovation literature to characterize the proximity of …rms’technologies (Ja¤e, 1986).7
measure. The measure varies between zero and one. It is zero for occupations that use identical skill sets and unity if two occupations use completely di¤erent skills sets. The measure will be closer to zero the more two occupations overlap in their skill requirements. To account for changes in task usage over time, we calculated the distance measures separately for each wave. For the years 1975-1982, we use the measures from the 1979 cross-section, for 1983-1988 the task measures from the 1985 wave; for the years 1989-1994, we use the measures based on the 1991/2 wave; and the 1997/8 wave for the years 1995-2001. While there have been changes over time in the distance measures, they are with 0.7 highly correlated. Our results are robust to assigning di¤erent time windows to the measures.
The mean distance between occupations in our data is 0.24 with a standard deviation of 0.22 (see Table 2). The most similar occupational move is between paper and pulp processing and a printer or typesetter with a distance of 0.002. The most distant move is between a banker and an unskilled Unlike the Euclidean distance, the angular separation measure is not sensitive to the length of the vector, i.e. whether an occupation only uses some tasks but not others. For example, two occupations using all tasks moderately (and thus have a position close to the origin of the coordinate system) will be similar according to the angular separation measure even if their task vectors are orthogonal, and therefore distant according to the Euclidean distance measure. If all vectors have the same length (i.e. if all tasks are used by at least some workers in all occupations), our measure is proportional to the Euclidean distance measure.
construction worker. Table 2 also shows at the bottom the distance measure for the three most common occupational switches separately by education group. The most popular move for low-skilled worker is between a truck driver and a warehouse keeper, while for the high skilled, it is between an engineering occupation and a chemist or physicist.8
3.3 The German Employee Panel Our second data set is a two percent sample of administrative social security records in Germany from 1975 to 2001 with complete job histories and wage information for more than 100,000 employees. The data has at least three advantages over household surveys commonly used in the literature to study mobility in the United States. First, its administrative nature ensures that we observe the exact date of a job change and the wage associated with each job. Second, measurement error in earnings and occupational titles are much less of a problem than in typical survey data as misreporting is subject to severe penalties. Finally, occupational titles are consistent across …rms as they form the basis for wage bargaining between unions and employers.
The data is representative of all individuals covered by the social security system, roughly 80 percent of the German workforce. It excludes the self-employed, civil servants, and individuals currently doing their compulsory military service. As in many administrative data sets, our data is right-censored at the highest level of earnings that are subject to social security contributions. Top-coding is negligible for unskilled workers and those with an apprenticeship, but reaches almost 25 percent for university graduates. For the high-skilled, we use tobit or semiparametric methods to account for censoring.
Since the level and structure of wages di¤ers substantially between East and West Germany, we drop from our sample all workers who were ever employed in East Germany. We also drop all those working in agriculture. In addition, we restrict the sample to men who entered the labor market in or after 1975.
This allows us to construct precise measures of actual experience, …rm, task, and occupation tenure Our distance measure treats all tasks symmetrically. It may, however, be argued that some tasks are more similar than others. For instance, the task ‘equipping machines’ may be more similar to ‘ repairing’ than to ‘teaching’ In order.
to account for this, we also de…ned the angular separation measure using information on the 3 aggregate task groups (analytical, manual and interactive tasks). The results based on this alternative distance measures are qualitatively very similar to the ones reported in the paper.
from labor market entry onwards. Labor market experience and our tenure variables are all measured in years and exclude periods of unemployment and apprenticeship training.
Since the concept is novel, we now explain how we calculate our measure of task human capital.
Each individual starts with zero task tenure at the beginning of his career. Task tenure increases by the duration of the spell if a worker remains in the same occupation. If he switches occupations, we calculate task tenure in the new occupation as the weighted sum of time spent in all previous occupations where the weights are the distance between the current and all past occupations.9 As an example, consider a person who starts out in occupation A, then switches to occupation B after one year, and switches to occupation C again after one year. Suppose the distance between occupation A and B is 0.5, between occupation A and C 0.2 and 0.8 between occupation B and C.
Before moving to occupation B, he has accumulated 1 unit of task tenure. Since he can only transfer 50 percent of his task human capital to occupation B, his task tenure declines to 0.5. After working one year in occupation B, he accumulates another unit of task human capital, so task tenure increases to 1.5 (0:5 1 + 1). Switching to occupation C after the second year, the worker can transfer 20 percent of his task human capital he accumulated in occupation A in the …rst period and 80 percent of the human capital accumulated in occupation B in the second period. His task tenure variables is thus 1 = 0:2 1 + 0:8 1.
Table 3 reports summary statistics for the main variables. In our sample, about 16 percent are low-skilled workers with no vocational degree. The largest fraction (68.3 percent) are medium-skilled workers with a vocational degree (apprenticeship). The remaining 15.4 percent are high-skilled workers with a tertiary degree from a technical college or university. Wages are measured per day and de‡ated to 1995 German Marks. Mean task tenure in our sample is between 4.6 years for the low-skilled and
4.8 years for the medium-skilled. Total labor market experience is on average a year higher (since the general skills captured by time spent in the labor market do not depreciate) and about one year lower In principle, our separation measure takes values between 0 and 1. However, the maximum distance observed in our data (across all occupation pairs) is 0.93. In our calculation, we assume that a worker cannot transfer any skills if he makes the most distant occupation switch, and de…ne the relative distance between two occupations A and B as the di¤erence between the maximum distance in our data and the distance between occupations A and B, divided by the maximum distance. Our results do not change if we use the actual distance instead of this relative distance measure.
for occupation tenure (since this measure assumes that these skills fully depreciate with an occupational switch).
Occupational mobility is important in our sample: 19 percent of the low-skilled switch occupations each year and with 11 percent somewhat lower for the high skilled. To see how occupational mobility varies over the career, Figure 1 plots quarterly mobility rates over the …rst ten years in the labor market, separately by education group. Occupational mobility rates are very high in the …rst year (particularly in the …rst quarter) of a career, and highest for the low-skilled. Ten years into the labor market, quarterly mobility rates drop to 2 percent. The next section uses our distance measure to analyze in more detail the type of occupational mobility we observe in the data.
4 Patterns in Occupational Mobility and Wages We now use the sample of occupational movers to demonstrate that skills are partially transferable across occupations. Section 4.1 studies mobility behavior, while Section 4.2 analyzes wages before and after an occupational move.
4.1 Occupational Moves are Similar Our framework predicts that workers are more likely to move to occupations with similar tasks requirements. In contrast, if skills are either fully general or fully speci…c to an occupation, they do not in‡uence the direction of occupational mobility: in the …rst case, human capital can be equally transferred to all occupations, while in the second case, human capital fully depreciates irrespective of the target occupation.
To test this hypothesis, we compare the distance of observed moves to the distribution of occupational moves we would observe if the direction of occupational moves was random. In particular, we assume that under random mobility the decision to move to a particular occupation is solely determined by its relative size. For example, if occupation A employs twice as many workers as occupation B, the probability that a worker joins occupation A would then be twice as high as the probability that he joins occupation B.
Observed moves are calculated as the percentage of moves for each value of the distance measure. To compare this to expected distance under random mobility, we calculate the fraction of individuals leaving an occupation that would end up in any of the 63 occupations in proportion to their relative size. Each random source-target occupation combination is then multiplied with the appropriate distance measure.
The way we calculate random mobility ensures that we account for shifts in the occupational structure over time, i.e. the fact that employment shares may be increasing or decreasing for some occupations.
Figure 2 plots the density of the distance measure under observed and random mobility. The horizontal axis is the distance measure where larger values are associated with movements to more distant occupations. The distribution under both random and observed mobility is bimodal, with many occupation switches concentrated at the distance measure of about 0.1 and 0.65. The peak at the distance measure of 0.1 is considerably lower, while the share of distant occupation switches is considerably higher under random than under observed mobility. Therefore, observed moves are much more similar than we would expect if occupational mobility was determined by relative size alone. The two distributions are statistically di¤erent at the 1 percent level based on a Kolmogoro¤-Smirnov test.
To allow a more detailed comparison, Table 2 compares selected moments of the distribution of our distance measure under observed and random mobility. The observed mean and the 10th, 25th, 50th, 75th and 90th percentile of the distance distribution are much lower than what we would observe under random mobility.
Our framework also predicts that distant moves occur early in the labor market career and moves become increasingly similar with time in the labor market. One reason is that the accumulation of task-speci…c skills makes distant occupational switches increasingly costly. A second reason is that with time in the labor market, workers gradually locate better and better occupational matches. It therefore becomes less and less likely that they accept o¤ers from very distant occupations.
Table 5 provides empirical support for these predictions. It shows the results from a linear regression where the dependent variable is the distance of an observed move separately by education group. Column (1) controls for experience and experience squared and year and occupation dummies. For all education groups, the distance of an occupational move declines with time spent in the labor market though at a decreasing rate. The declining e¤ect is strongest for the high-skilled, who also make more similar moves on average (see last row). For the high-skilled, 10 years in the labor market decrease the distance of a move by 0.16 or about 70 percent of the standard deviation. For the medium-skilled, the decline is only about 0.030 or 14 percent of a standard deviation.
Column (2) adds the time spent in the last occupation, while column (3) reports the results from an …xed-e¤ects estimator to account for heterogeneity in mobility behavior across individuals. More time spent in the previous occupation decreases the distance of an occupational move in addition to labor market experience. The within estimator shows that occupational moves become more similar even for the same individual. The results are therefore not driven di¤erences between low- and high-experience workers. In fact, the decline in the distance becomes even more pronounced for all education groups in the …xed e¤ects estimation.
Table 4 imposed a quadratic relationship between actual labor market experience and the distance of moves. In Figure 3, we relax this restriction. The …gure displays the average distance of a move by actual experience, separately for the three education groups. The average distance is obtained from a least-squares regression of the distance on dummies for actual experience as well as occupation and year dummies, similar to Column (1) in Table 4. The …gure shows that occupational moves become more similar with time in the labor market for all education groups, but particularly so for the high-skilled.
For this education group, the decline is particularly pronounced in the …rst 5 years in the labor market.
The decline between the …rst and 15th year of actual labor market experience is statistically signi…cant at the 1 percent level for all education groups.
In sum, individuals are more likely to move to occupations in which similar tasks are performed as in their source occupation, particularly so later in their career. Our framework proposes a simple explanation for this pattern. The basic mechanism is that human capital is more transferable between occupations with similar skill requirements.