A look at the new job-task information in the National Longitudinal Surveys of Youth

O*Net and DOT only supply information on job attributes at the occupational level.⁴ The Princeton Data Improvement Initiative (PDII) is the only dataset we are aware of that has information on job attributes at the worker level. David H. Autor and Michael J. Handel use this dataset to analyze how tasks vary both within and across occupations.⁵ Although their sample size is small (their regressions typically have 1,333 observations), they find meaningful task variation at the worker level, even controlling for occupation.

Recent National Longitudinal Surveys of Youth 1979 (NLSY79) and 1997 (NLSY97) fielded questions like those in the PDII. This article is the first to look at a new source of information on how job attributes vary among workers both within and across occupations. These new questions allow us to evaluate how analytical, routine, and manual job tasks vary within and across jobs as well as how they relate to workers’ characteristics and wage differences.

Upon analyzing the data, we find that the distributions of tasks across occupations conform to one’s expectations. However, we also find a substantial amount of job-task variation within occupations. Wage regressions indicate that at least some of the within-occupation task variation is meaningful. Job-task information also appears to be helpful in explaining some of the wage variation among demographic groups.

In the sections that follow, we describe the NLSY data and provide an overview of the new information on job tasks. We then examine how job tasks vary among several demographic groups and analyze the relationship between job tasks and wages. In the final section, we conclude with our findings.

Data and characteristics of the sample

We analyze data in the NLSY79 and the NLSY97, the two active surveys. The NLSY79 is a dataset of 12,686 individuals who were ages 14 to 22, when first interviewed in 1979. These youth were interviewed annually from 1979 to 1994 and every 2 years since then. The NLSY97 is a dataset of 8,984 individuals who were ages 12 to 17, when first interviewed in 1997. These youth were interviewed annually from 1997 to 2017 and every 2 years since then. In the 2016–17 survey year, the NLSY79 obtained information on job attributes. The NLSY97 did the same in the 2017–18 survey year.

Seven questions from the PDII were added to the NLSY79 and the NLSY97. These questions evaluate (1) how much time is spent on physical tasks, (2) how much time is spent on repetitive tasks, (3) how much time is spent managing or supervising, (4) the frequency of resolving complicated problems, (5) the frequency of using higher level math skills, (6) the typical length of documents read at work, and (7) whether the respondent has much face-to-face contact with people other than coworkers and supervisors.⁶

The NLSY79 and NLSY97 samples used for this article are restricted to respondents who reported their sex and education and who have a valid Armed Forces Qualifying Test (AFQT) score.⁷ Respondents also had to report a wage and occupation, answer all the job-task questions, and be currently working at the time of the interview. To classify respondents by educational attainment, we use their most recent report of highest degree completed. We also our restrict analysis to occupations with at least two observations.⁸ After these restrictions, the sample from the NLSY79 consists of 3,606 workers, representing 291 occupations. The sample from the NLSY97 consists of 3,656 workers, representing 281 occupations.

An overview of the task variables in the National Longitudinal Surveys of Youth

The primary determinant of the tasks that workers perform on the job is the occupation they are in. Table 1 presents summaries of tasks by broad occupation for the NLSY79 and the NLSY97. (See appendix A, tables A-1 and A-2, for more detailed breakdowns of the task variables by broad occupation.) The distributions of tasks across broad occupations conform to one’s expectations. Workers who are managers say that they spend most of their time managing and supervising. Workers who are managers and professionals spend the most time reading long documents and solving problems of more than 30 minutes. Construction, production, and transportation workers spend more time on physical tasks. Workers in clerical occupations spend more than half their time on repetitive tasks. Workers in production, transportation, and service occupations also spend much of their time performing repetitive tasks. As noted in the introduction, routine jobs have the highest risk of being displaced by automation.

Jobs often involve the performance of several related multidimensional tasks. Therefore, looking at how the National Longitudinal Surveys task variables are correlated is important. The correlations in table 2 accord with one’s intuition. Managing, problem solving, using math, and reading documents task variables are positively correlated with each other. The routine and physical-task variables are correlated with each other and negatively correlated with managing, problem-solving, using math, and reading documents task variables. However, apart from the routine and physical-task variables, the correlations are not terribly strong. Unlike Autor and Handel’s analysis, our analysis does not group together variables using principal components.⁹ The relatively weak correlations among the analytical variables show that they are fairly different from each other and that grouping them would result in a loss of information.

While the primary determinant of the tasks that workers perform is the occupation they are in, casual observation suggests that workers in the same job do not all perform the same tasks. We can determine how much tasks vary across occupations by regressing the task variables against the occupation dummies. Table 3 shows the portion of the total variation in the task variables that is explained when the task variables are regressed against the broad occupation dummies.¹⁰ The portion of total variation explained is generally low. The math-task variable has the least explained variation. Only 3 percent of the variation is explained by broad occupation in the NLSY79, which means that 97 percent of the variation in this variable is unexplained. In the NLSY97, only 5 percent of the variation in the math-task variable is explained by broad occupation, which means that 95 percent is unexplained. The percentage of variation explained is greatest for the physical-task variable. Thirty percent of the variation in this variable is explained in the NLSY79. In the NLSY97, this figure is 31 percent. Thus, nearly 70 percent of the variation in physical tasks is unexplained by variation in broad occupation.

In summary, a great deal of variation in the task variables is not explained by the broad occupation variables. This result is not surprising because the various broad occupation groupings are composed of several diverse occupations.¹¹ A more interesting experiment is to regress the task variables against detailed occupations. These results are also presented in table 3. While the detailed occupations explain more task variation than the broad occupations, a perhaps surprising amount of task variation is still within the detailed occupations. In fact, most of the variation in the task variables is not explained by the more detailed four-digit occupations and is thus variation within these occupations.

Table 3 also shows the variation in the task variables that is explained by the demographic variables plus the education and AFQT variables.¹² These variables alone generally explain less variation in the task variables than do the detailed occupation variables, except for reading and routine tasks reported in the NLSY79. Finally, table 3 shows the portion of the variation in the task variables that is explained by the four-digit occupations, the demographic variables, education, and the AFQT score. The addition of the demographic variables, education, and the AFQT score to the regression with the detailed occupation variables explains much more of the variation in the routine task for both cohorts. However, the additional variables explain little more of the variation in the remaining task variables for either cohort.

For both cohorts, the complete regression never explains more than 50 percent of the variation in any of the task variables. And for most task variables, the proportion of the variation explained is substantially less than 50 percent. Of course, some of the unexplained task variation is almost certainly statistical noise. The question that arises is, “Is any of the variation meaningful, and if so, how much?” One way to address this question is to look at whether the variation in the task variables explains variation in wages that is not explained by the detailed occupation variables. We look at this in a later section. First, however, we examine how tasks vary among demographic groups and across the two cohorts.

Variation in tasks among demographic groups

Table 4 summarizes the occupation distributions of our two cohorts by various demographic characteristics.¹³ Men are more likely to be employed in management, construction, production, and transportation occupations, while women are more likely to be employed in professional, sales, clerical, and service occupations. This pattern is consistent for both cohorts. White workers are more likely to be employed in management and professional occupations, and Black workers are more likely to be employed in transportation and service occupations. Hispanic workers are more likely to be employed in clerical occupations, especially for the younger cohort in which nearly 1 in 5 Hispanic workers are in clerical occupations.

As table 4 shows, workers with more education are more likely to be in managerial and professional occupations and less likely to be in production, transportation, service, and construction occupations.¹⁴ The pattern with respect to AFQT scores is similar to that for education. Individuals with higher AFQT scores in the NLSY79 and NLSY97 are more likely to be in managerial and professional occupations and less likely to be in production, transportation, service, and construction occupations.¹⁵

Given that the various demographic groups are distributed unevenly among occupations, one should expect that the task content of jobs will differ among demographic groups. Table 5 summarizes the task content of jobs by sex, race (White, Black, Hispanic), education (less than a high school diploma; high school graduates, no college; some college or associate’s degree; bachelor’s degree and higher), and AFQT score quintile for the 1979 and 1997 cohorts. (See appendix B, tables B-1 and B-2, for more detailed breakdowns of the task variables by demographic group, education, and AFQT score.) For the older cohort, we see that only about 21 percent of wage and salary workers use math at a high school level or above at least weekly, while a little less than 33 percent read documents longer than six pages regularly as part of their jobs. A larger percentage report solving somewhat complicated problems at least weekly (72 percent) and approximately 29 percent report spending at least half their time managing or supervising others. A little more than 48 percent of workers have a lot of face-to-face contact with people other than coworkers and supervisors as part of their job. This contact seems to be mostly with customers and clients: 43 percent of respondents indicate a lot of face-to-face contact with customers and clients, while far fewer have regular face-to-face contact with suppliers or patients (less than 12 percent in both cases).¹⁶ A little less than 43 percent of workers in the NLSY79 report spending more than half their time on short, repetitive tasks and a little more than 48 percent report spending at least half their time doing physical tasks, such as standing, handling objects, or operating equipment.

As shown in table 5, the overall figures reported by the NLSY97 cohort are generally similar to those reported by the NLSY79 cohort. The largest difference concerns the use of math on the job. In the NLSY97, approximately 29 percent use math compared with 21 percent in the NLSY79. Some other differences exist but are smaller. Workers in the NLSY79 were ages 52–59 and entering the latter stage of their career when they were asked questions about their job tasks. In contrast, workers in the NLSY97 were ages 32–38 and entering the prime stage of their career when they were asked questions about their job tasks. It is interesting that job tasks in the two cohorts are distributed so similarly.

We have seen that both job tasks and worker demographics vary across jobs. We would therefore expect job tasks to be related to workers’ demographic characteristics. We highlight a few interesting results from table 5 and follow this up later with a more detailed discussion based on regression analysis. In the 1979 cohort, women are much more likely than men (49 percent versus 37 percent) to spend more than half their time on repetitive tasks. For the younger 1997 cohort, women are only slightly more likely than men (45 percent versus 42 percent) to spend more than half their time on repetitive tasks. In both surveys, women engage in fewer tasks that involve managing, reading, and using math tasks.

Blacks and Hispanics are more likely than Whites to spend at least half their time on physical job tasks (46 percent, 60 percent, and 51 percent for Whites, Blacks, and Hispanics, respectively, in the 1979 cohort; and 45 percent, 57 percent, and 47 percent, respectively, in the 1997 cohort). Blacks and Hispanics also spend more time on repetitive tasks and less time reading documents and problem solving.

As expected, the time spent on repetitive tasks decreases as a worker attains more education. In the NLSY79, 58 percent of workers with less than a high school education and 20 percent of those with a college degree or more education spent at least half their time on repetitive tasks. In the NLSY97, the percentages were 73 percent and 26 percent for each group of workers, respectively. The same decreasing trend is true for AFQT scores. Of the NLSY79 workers, 61 percent in the first quintile and 17 percent in the fifth quintile spent at least half their time on short repetitive tasks. However, in the NLSY97, the percentage of workers were 68 percent and 18 percent, respectively, for each quintile. Similarly, workers with more education and higher AFQT scores are less likely to engage in physical tasks. The opposite is true of managing, problem-solving, and reading tasks. Workers with more education and higher AFQT scores are more likely to spend at least half their time on managing or supervising, to solve problems of more than 30 minutes weekly, and to typically read documents of more than six pages.

Regression analysis

To complete our analysis of how tasks vary among demographic groups, we also estimate regressions. Unlike the means reported in table 5, the resulting regression coefficients show comparisons holding all other variables in the equation constant. The regression results also provide a convenient way to test whether observed differences among demographic groups are statistically significant. The dependent variables in these regressions are “standardized task variables” that factor in the entire range of responses to the task questions.¹⁷ Table 6 presents estimates when only demographic characteristics are included in the regression equations. The regressions tell a similar story as the means reported in table 5. In the NLSY79, women spend less time managing, reading long documents, problem solving, using math, and doing physical tasks, and they spend more time doing repetitive tasks. In the NLSY79, women have more face-to-face contact with individuals who are not supervisors or coworkers than do men. However, in the NLSY97, the women have less face-to-face contact. The sex discrepancies are generally smaller in the NLSY97, most notably for repetitive tasks. And the sex discrepancy for math is greater in the NLSY97. As reported in the NLSY79, compared with Whites, Blacks spend less time reading documents and problem solving and more time doing repetitive and physical tasks. This discrepancy is generally attenuated in the NLSY97 and considerably so for problem solving and physical tasks. As do Blacks, though to a lesser degree, Hispanics also spend less time reading documents and problem solving and more time on repetitive and physical tasks. However, contrary to the discrepancies for Blacks, these discrepancies all increased by notable amounts in the NLSY97.

Comparing the regressions for the two cohorts, one sees that from the NLSY97, individuals report spending substantially more time using math than do individuals in the NLSY79.¹⁸ Other task differences exist but are smaller. Individuals in the NLSY97 spend somewhat more time in management (the constant and the coefficient for women are statistically larger in the NLSY97). They also spend somewhat more time reading and problem solving (all coefficients except those for Hispanics are statistically larger in the NLSY97). Women in the NLSY97 spend somewhat less time in routine jobs, and Blacks in the NLSY97 spend less time in physical jobs than do their counterparts in the NLSY79. Women in the NLSY97 spend somewhat more time having face-to-face contact with people other than coworkers and supervisors.¹⁹

The regressions in table 7 include education and AFQT. Adding education and AFQT has little effect on the coefficients for women in the task equations. In contrast, adding the education and AFQT substantially affects the coefficients for Blacks. The discrepancy for Blacks in problem solving and reading is substantially reduced in the NLSY79 and eliminated in the NLSY97. For Blacks performing repetitive tasks, the discrepancy is reduced in both cohorts. However, in physical tasks, the discrepancy for Blacks is eliminated in the NLSY79 and becomes negative in the NLSY97 once one controls for education and the AFQT score.

Adding education and AFQT to the regression also affects some of the coefficients for Hispanics. The reading and problem-solving discrepancies for Hispanics are eliminated in the NLSY79 and greatly reduced in the NLSY97. The discrepancy for Hispanics in repetitive tasks is eliminated in the NLSY79 and substantially reduced in the NLSY97. When one controls for education and AFQT, in the NLSY79, Hispanics spend less time on physical tasks, and in the NLSY97, they spend about the same amount of time as non-Hispanics.

Controlling for AFQT scores, we find that workers with more education, in the NLSY79, spend more time on math, but in the NLSY97, they spend less time on math. The higher the level of an individual’s education, the greater the time spent on reading documents as reported in the NLSY79. In the NLSY97, only individuals with 4 or more years of college spend more time reading documents.²⁰ Workers with more education in the NLSY79 spend more time problem solving. The effect of education on time spent problem solving is weaker in the NLSY97 and only positive for 4 or more years of college. In both surveys, individuals with more education spend less time on repetitive tasks. Individuals with some college or more education have more face-to-face contact with nonworkers in the NLSY79 but not in the NLSY97.

Finally, looking at the coefficients on AFQT, one sees that in both surveys, individuals with higher AFQT scores spend more time reading and less time doing routine and physical tasks. The greater one’s AFQT score, the more time is spent on problem solving in the NLSY79. In the NLSY97, individuals in the lowest AFQT quintile spend less time on problem-solving tasks than others, but no discernible difference in time is spent on problem solving among the four upper quintiles.

The regressions in table 8 include indicators for four-digit occupations. Adding the occupation variables primarily affects the coefficients for women, education, and AFQT score. Adding the occupation controls substantially reduces the coefficients for women in the management, problem-solving, and math tasks for both cohorts, with the problem-solving discrepancy eliminated entirely in the NLSY97. The coefficient for women in the physical-task equation is reduced to zero in both cohorts, and the coefficient for women in the face-to-face contact equation becomes negative. The education and AFQT effects generally become smaller when the occupation variables are added to the job-task regression equations.

Wage regressions

As shown in table 8, a large amount of variation remains in the task variables, even after adding controls at the detailed occupation level. The question that arises is whether this variation is meaningful or simply statistical noise. To explore this question, we now estimate wage equations.²¹

We begin by examining how much wage variation across occupations can be explained by the job-task variables. Toward this end, we calculate the mean wage and the mean values of the job-task values in the various four-digit occupations. The results of regressing the logarithm (log) of the mean occupational wage against the mean task values are shown in the first two columns of table 9. The task variables are powerful predictors of occupational wages, especially in the NLSY79. Occupations in which workers spend more time on tasks that involve managing, problem solving, using math, and reading documents pay a higher wage (although the math and management tasks are not statistically significant in the NLSY97 occupational wage regression). Occupations in which workers spend more time on routine tasks and physical tasks pay a lower wage (time spent on physical tasks is not statistically significant in the NLSY97 regression). The task variables explain 68 percent of the occupational log-wage variation in the NLSY79 and 57 percent of the occupational log-wage variation in the NLSY79.

We next examine whether the task variables can explain wage variation within occupations. To do this, we calculate the difference between each individual’s log wage and the mean log wage of all workers in the individual’s occupation. We also calculate the differences between the individual’s job-task variables and the mean job-task values of all workers in the occupation. The first two columns of table 10 show the results of regressing the log-wage difference against the job-task differences. The coefficients on the job-task differences are all statistically significant and have signs that are consistent with the across-occupation results: individuals who spend more time on managing tasks, problem-solving tasks, using math, and reading documents than their occupational counterparts receive a higher wage; and individuals who spend more time on routine tasks and physical tasks than their occupational counterparts receive a lower wage. However, the adjusted R-squared is only 0.12 for the NLSY79 and 0.07 for the NLSY97. We can conclude that some of the within-occupation variation in the task variation is meaningful, but the task variables explain only a relatively small percentage of within-occupation wage variation.²² This result is likely due to noise in both the task and the wage variables.

We conclude this section by examining to what extent the job-task variables can explain demographic wage differentials. The columns under regression 2 of table 9 show the results of regressing mean occupational wages against the mean occupation values of the demographic variables, education, and AFQT. As expected, occupations in which workers have more education and higher AFQT scores pay higher wages. It is also noticeable how much lower wages are in occupations with more female workers. In addition, in the NLSY97, wages are substantially lower in occupations that employ more Hispanics. Adding the mean occupational task values to this regression equation yields the results shown in the columns under regression 2 of table 10. The routine, physical, management, and problem-solving task variables are statistically significant in the NLSY79 regression, but only the routine variable is significant in the NLSY97 regression. The coefficients on education, the AFQT score, and the demographic variables fall. The lower wage in occupations with a greater female presence is partly explained by the job-task variables, because adding these variables reduces the coefficient for women by 17 percent and 14 percent in the NLSY79 and NLSY97, respectively.

The last two columns of table 9 show the results of regressing the difference between an individual’s log wage and the mean log wage in the individual’s occupation against differences between the individual’s demographic characteristics and the mean values of the demographic characteristics in the occupation. Individuals with more education and a higher AFQT score earn higher wages than their counterparts in the same occupation. And women receive lower wages than men.²³ Table 10’s last two columns add differences between the individual’s job-task variables and the mean job-task values of all workers in the occupation to this regression equation. The coefficients on the task differences are statistically significant with the expected signs (except for the math task, which is statistically insignificant in the NLSY79 regression). Some of the coefficients on the education and AFQT variables fall, but the coefficients on the demographic variables are affected only marginally.

Conclusion

Recent NLSY79 and NLSY97 obtain information on the tasks that workers perform on the job. In this article, we have provided an initial look at these data. The data show that substantial differences exist among demographic groups in the task content of their jobs. However, nearly all the racial and ethnic differences disappear once one controls for education and AFQT scores. In contrast, controlling for education and AFQT scores has little effect on the discrepancy between the task content of the jobs held by men and women. Adding four-digit occupation controls reduces but does not eliminate these discrepancies. Within occupations, women spend less time on tasks involving managing, problem solving, reading long documents, and using math and more time on tasks that are repetitive.

Workers in the NLSY79 were ages 52–59 and entering the latter stage of their career when they were asked questions about their job tasks. In contrast, workers in the NLSY97 were ages 32–38 and entering the prime stage of their career when they were asked questions about their job tasks. Workers in the NLSY97 report using more math on the job than do workers in the NLSY79. However, somewhat surprisingly, the job tasks in the two cohorts are mostly distributed similarly.

As one would expect, some of the variation in task content among workers is explained by their occupation. However, even after we added controls at the detailed occupation level, a great amount of task variation still exists within the detailed occupations. In fact, most of the variation in the task variables is not explained by detailed four-digit occupations.

The job-task variables explain a substantial portion of the variation in wages across occupations. Regressing the difference between an individual’s wage and the mean occupational wage against the differences between the individual’s job-task variables and the mean job-task values of all workers in the occupation provides some insight into whether any of the within-task variation that we observe is meaningful or whether it is nearly all simply statistical noise. The coefficients on the task variables are statistically significant, indicating that some of the within-occupation variation in job tasks is meaningful. However, the portion of within-occupation wage variation explained by the task variables is relatively small, because the adjusted R-squared is only 0.12 for the NLSY79 and 0.07 for the NLSY97.

We could not use the longitudinal information in the National Longitudinal Surveys because the job-task information has only been asked once for each cohort. Job-task information in future surveys could be helpful in examining how the content of individuals’ jobs changes as they gain more labor market experience, which may be helpful in analyzing wage growth. Job-task information in future surveys could also be helpful in analyzing voluntary and involuntary job mobility as well as the consequences of automation. Researchers can also study comparisons across occupations by merging O*NET data into the NLSY79 and NLSY97. Perhaps the main advantage of the National Longitudinal Surveys job-task variables is that, unlike O*NET, they allow for comparisons within occupations. However, given that the NLSY job-task variables explain only a relatively small part of wage variation within occupations, the amount of value they add is not clear. A comparison of the NLSY job-task and the O*NET variables is a topic for future work.

Appendix A. Princeton Data Improvement Initiative task measures of employed workers, ages 52 to 59 and ages 32 to 38, by occupation

Appendix B. Task measures of employed workers, ages 52 to 59 and ages 32 to 38, by major demographic group