Does rising demand for abstract tasks offset declining racial discrimination in wages?
Core argument: Declining discrimination and narrowing racial skill gaps shrank the Black-White wage gap by 8% between 1980 and 2018, while the rising return to Abstract tasks widened it by 7%.
The wage return to cognitive skills is higher for Black men than for White men with the same occupation and education, highlighting task-based discrimination. Between 1980 and 2018, the racial wage gap narrowed by 8% due to declining discrimination and narrowing racial skill gaps. However, the increasing return to Abstract tasks widened the gap by 7%, disadvantaging Black workers who are underrepresented in these tasks. Despite declining race-specific barriers, the rising return to Abstract tasks has favored Whites since 1980, preventing Black-White wage convergence. The data from the U.S. Department of Labor's DOT and O*NET reveal that occupations demanding Abstract tasks, such as those requiring analytical flexibility and complex interpersonal communications, have seen increased returns, underscoring the systemic challenges Black workers face in achieving wage parity.
Ed Comment:That outcome seems at odds with the study which says there is discrimination against blacks in abstract/conceptual jobs. Ben, how does it reconcile? Is there an appendix that shows the wages of black and whites vs their relevant abstract/conceptual credentials? They mention AFQT test scores but don't seem to use them. Do they use them? Does the appendix show wages vs test scores (and other credentials), which they clearly seems to have? Using education as a measure of intelligence is problematic because blacks are admitted to colleges with substantially lower scores. Let's read and summarize the reference Neal (decker) 2006, which accompanies their claim that blacks in abstract jobs are paid more than whites. I glanced at it. It's pretty direct. It says: "Some may argue that what matters is not realized returns to investments but rather the fact that many blacks believe that the labor market will not reward them or their children for investing in human capital. The literature on statistical discrimination began with Arrow (1973) and Phelps (1972). When Arrow and Phelps first developed these ideas, existing cohorts of adult workers had lived most of their lives in labor markets where blacks may well have earned lower returns to education and skills than whites workers.28 However, it is hard to find data that would support this belief today. Blacks who reached adulthood after the Civil Rights Act of 1964 have not earned systematically lower returns to attainment or achievement than their white peers. During the past decade or more, black adult respondents in the NLSY79 have earned relatively high returns to skills measured by the AFQT.29 Further, as I demonstrate below, apparent returns to formal schooling among black young adults have typically been as large or larger than the corresponding returns among whites since at least 1970....In the balance of this section, I present details concerning the relationship between education and labor market outcomes among black and white adult men. Existing panel data sets do not provide information concerning racial differences in the relationship between cognitive skills and adult outcomes for birth cohorts other than those included in the NLSY79.30 However, census data document the relationship between educational attainment and labor market outcomes for a broad range of cohorts. Education differs from reading and math scores in that it is an indirect measure of skill that is easily observed by employers, but the patterns documented below are similar to those in Figures 4(a)-5(b). Correlations between schooling and positive labor market outcomes among blacks are as strong or stronger than the corresponding correlations among whites, especially in recent decades. 30 The National Educational Longitudinal Surveys provide panel data on high school experiences and adult labor market outcomes. However, these samples are drawn from students who are enrolled in specific grades at a point in time, and thus they provide select samples from persons born in a set of birth years. In addition, the follow-up data on earnings and employment are of relatively poor quality in the NELS surveys when compared to NLSY79 on work outcomes...." Has he updated this or related work? Has he written anything else of high interest? Maybe I'll talk to him at U Chicago.
Ben Comment:From the paper: "The differences in pre-labor market skills and discrimination facing Black workers give rise to differential sorting into tasks between Black and White individuals in the spirit of Roy (1951). These sorting differences need to be accounted for in order to parse out the effects of race-neutral driving forces (such as time trends in task specific returns) and race-specific driving forces (such as a narrowing of racial skill gaps and/or declining discrimination) when explaining changes in racial wage gaps over time. Using the model structure and empirical moments on the differential occupational sorting of Black and White men, the changing returns to various tasks, and the evolution of the aggregate racial wage gap, we estimate the key driving forces of the model.” Most black people are kept out of the abstract jobs because of discrimination/skill differences (below I pull a quote where they say these skill differences are the result of discrimination, too). The only black people who move into the abstract job categories are the ones who have overcome these. If you think in terms of diminishing marginal productivity, only the absolutely most productive black workers are in the “abstract” category, so they’re making higher wages than the average. If more black people entered abstract jobs, they would bring down the average productivity and (presumably) the wage. The issue of discrimination has to do with the number (or proportion) of black people who enter the field, not with what they make conditional on entering the field. From the discussion of figure 4: "Most of the racial gaps in the average task returns are slightly negative suggesting that the task return to Black men are systematically smaller than White men of similar age and education.” So it’s only in “abstract” tasks that we see a “premium” paid to black workers. And according to this graph it’s very small. Additionally, in the model section: "First, the average skill level may differ across race groups stemming from current or past discrimination.” The assumption is that skill differences are because of discrimination and are not “genetic”. I didn’t see any graphs in the appendix like the ones you asked about.
Ed Comment:I’m still confused (about the math). (I don’t care what the paper says about the causes of racial difference in test scores. The authors are not experts in that. That’s obligatory BS to cover their asses. …and let’s not debate it. It would help my understanding of the paper if you ignored your hypotheses and focused on what the data and the math in the paper actually measure and the limitation of those measures. My own view, from a lot of experience hiring people into highly analytical jobs, is that there is great demand and competition between employers for scarce analytical blacks because of societal demands. Like colleges, we greatly lowered the bar. And their chances of success were far greater as a result of these allowances. So I’m not surprised blacks are paid more than whites in analytical jobs, with the premium growing larger the more precisely they can measure the true analytical capacity of the person and the extent to which the job demands analytical capability.) I don’t see how they measure the extent to which selection bias improves the pool of workers relative to their credentials. I would think they merely have their wage, some measure of the pedigree of their analytical credentials, and measures of the extent to which the job they perform is analytical. If not, please explain. What measures do they use for the pedigree of their analytical credentials? How do they measure the extent to which the job is analytical? My guess is that unless they are using their AFQT score—something the most analytical people are unlikely to have—they don’t have a measure of the analytical skill beyond yes/no. What measure of analytical prowess do they use? Similarly, is the analyticalness of the job measured or is it just yes/no? If it’s measured, how so? Are they merely saying, black people in analytical positions get paid more than white people in analytical positions, on average, without regards to measures of the person’s or the job’s anlysiticalness? Or are they making some adjustment for the extent to which the job is analytical and for the pedigree of the person’s analytical credentials? If black people are paid more than whites on average in analytics jobs, how do they measure discrimination in analytical jobs? I don’t see how the math works or could work unless the measurements are quite precise. For example, they could say black people in analytical jobs have test scores of 125, whereas white people in analytical jobs only have test scores of 115. Blacks are paid more than whites, but, given their higher test scores, they should be paid even more than they are being paid. Show me if and how they do that. If they have all these measures and wages, I’d love to see a graph of x - measure of analyticalness (e.g., test score), and y - wages for blacks vs whites. Again, let’s not speculate about what the data means. Let’s focus on the data they have and how they use it to calculate how people who are paid more are discriminated against.
Ben Comment:They do not create the graph “ x - measure of analyticalness (e.g., test score), and y - wages for blacks vs whites”. Looking at the BLS website it seems like that might be something we can do (it’ll take time, I don’t know these data at all). Other questions below, without my commentary - all language is drawn from the paper. ED: How do they measure the extent to which the job is analytical? "we measure the skill demands in each occupation using the following data: (i) the U.S. Department of Labor’s Dictionary of Occupational Titles (DOT) and (ii) the Occupational Information Network (O*NET) sponsored by the U.S. Department of Labor/Employment and Training Administration (USDOL/ETA) … The first three measures are taken exactly from Autor and Dorn (2013) and Deming (2017) using the DOT data. Below, we provide a brief summary of these measures. The last task measure is new and was created specifically for this paper to help get at the concept of taste-based discrimination. Building on the work in Deming (2017), Contact measures the extent to which an occupation requires interaction and communication with others within the organization (co-workers) or outside the organization (customers/clients) …
Abstract: indicates the degree to which the occupation (i) demands analytical flexibility, creativity, reasoning, and generalized problem-solving and (ii) requires complex interpersonal communications such as persuading, selling, and managing others. Occupations with high measures of Abstract tasks include accountants, software developers, high school teachers, college professors, judges, various medical professionals, engineers, and managers. Routine: measures the degree to which the task requires the precise attainment of set standards and/or repetitive manual tasks. Occupations with high measures of Routine tasks include secretaries, dental hygienists, bank tellers, machinists, textile sewing machine operators, dressmakers, x-ray technology specialists, meter readers, pilots, drafters, auto mechanics, and various manufacturing occupations. Manual: measures the degree to which the task demands eye, hand, and foot coordi- nation. Occupations with high measures of Manual tasks include athletes, police and fire fighters, drivers (taxi, bus, truck), skilled construction (e.g, electricians, painters, carpenters) and landscapers/groundskeepers. Contact: measures the extent that the job requires the worker to interact and communi- cate with others (i) within the organization or (ii) with external customers/clients or potential customers/clients. To create our measure of Contact tasks we use two 1998 O*NET work activity variables taken from Deming (2017). Specifically, we use the variables Job-Required SocialInteraction(Interact)andDealWithExternalCustomers(Customer).4 Interactmea- sures how much workers are required to be in contact with others in order to perform the job. Customer measures how much workers have to deal with either external customers (e.g., retail sales) or the public in general (e.g., police work). To make our measure of the Contact task content of an occupation, we take the simple average of Interact and Customer for each occupation. Occupations with high measures of Contact tasks include various health care workers, waiter/waitress, sales clerks, lawyers, various teachers, and various managers." Appendix A.3 has even more detail and Autor and Dorn (2013) and Deming (2017) have even more detail if you’d like us to find that for you. ED: What measures do they use for the pedigree of their analytical credentials/What measure of analytical prowess do they use? From the ACS: "In most specifications, we control for the worker’s age and accumulated years of schooling.” From the NYSL: "The NLSY data allows us to link a worker’s subsequent occu- pational choice with pre-labor market measures of cognitive, non-cognitive, and social skills… The key reason we use the NLSY data is to have measures of racial differences in pre- labor market skills. We use measures of performance on cognitive test and psychometric assessments to generate a set of unified proxies for cognitive, non-cognitive and social traits across the two NLSY waves. These skill measures were primarily collected before the indi- viduals entered the labor market. Again, our goal is to take these skill measures directly from the existing literature. We summarize these measures briefly here and include a more detailed discussion in the online appendix. Cognitive Skills (COG): We follow the literature and use the respondent’s standard- ized scores on the Armed Forces Qualifying Test (AFQT) as our measure of cognitive skills. The AFQT is a standardized test which is designed to measure an individual’s math, verbal and analytical aptitude. The test score was collected from all respondents in their initial year of the survey and was measured in both the 1979 and 1997 waves.7 Non-cognitive Skills (NCOG): We use the measures of non-cognitive skills created by Deming (2017). Deming (2017) uses questions pertaining to the Rotter Locus of Control Scale and the Rosenberg Self-Esteem Scale for the NLSY79 cohort to make a measure of non- cognitive skills.8 Likewise, for the NLSY97 cohort Deming (2017) uses respondent answers (provided prior to entering the labor market) to the question “How much do you feel that conscientious describes you as a person?” to approximate respondents’ non-cognitive skill. Deming (2017)’s non-cognitive skill measures are expressed in z-score units. Social Skills (SOC): We again follow Deming (2017) to generate a unified measure of social skills using a standardized composite of two variables that measure extroversion in both waves. Specifically, for the NLSY79, we use self-reported measures of sociability in childhood and sociability in adulthood. Individuals were asked to assess their current sociability (extremely shy, somewhat shy, somewhat outgoing, or extremely outgoing) and to retrospectively report their sociability when they were age 6. For the NLSY97, we proxy for social skills using the two questions that were asked to capture the extroversion factor from the commonly-used Big 5 personality inventory. For each wave, we normalize the two questions so they have the same scale and then average them together. We then convert the measures into z-score units. Deming (2017) shows that these measures of social skills positively predict individual wages when they are adults even conditional on controlling for individual measures of cognitive skills (AFQT).” ED: are they making some adjustment for the extent to which the job is analytical and for the pedigree of the person’s analytical credentials?
"wijt is the log wage of individual i working in occupation j during year t. Our coefficients of interest are again the βg ’s, the Mincerian wage premium of task k in year t kt for group g. For this regression, we use are sample of full-time workers… the demographically adjusted Black-White gaps in the wage premium by task requirement (Panel B) … Finally, the racial gaps in the wage premiums to tasks are relatively small and roughly constant over time. Most of the racial gaps in the average task returns are slightly negative suggesting that the task return to Black men are systematically smaller than White men of similar age and education”. ED: If black people are paid more than whites on average in analytics jobs, how do they measure discrimination in analytical jobs? "New to our model is differential returns to performing each type of tasks across different race groups. We allow for three forces that can cause the returns to tasks to differ across races. First, the average skill level may differ across race groups stemming from current or past discrimination.16 We capture this by writing the skill endowment vector as Φgi = {ηgk +φi1.., ηgK +φiK }, where ηgk represents the skill (or human capital) level of group g associated with task k. Second, workers of a particular race group may face taste-based discrimination in performing a task. Taste-based discrimination may exist if for example customers do not like interacting with Black employees or White workers do not like interacting with their Black co-workers. In the presence of taste-based discrimination, we assume that employers perceive the efficiency of the discriminated workers to be δtaste+ηgk+φik rather than ηgk+φik, gk where δtaste is the race-specific taste-based discrimination coefficient in task k. Thus, the gk potential log wages of a worker belonging to race g becomes:
Lastly, workers may face statistical (rather than taste-based) discrimination if their employers do not perfectly observe individual workers’ skills. If skills are observed with noise, employers form expectations about a worker’s marginal product by using information about the individual’s group. The statistical discrimination coefficient, δstat, will therefore differ gk by task based on underlying gaps in group mean of skills (ηgkt’s) used to perform the task. We formally incorporate the notion of statistical discrimination into the model by intro- ducing noise to skill measurement. Suppose employers cannot observe worker’s true efficiency, ηgk + φik, and instead only observe a noisy skill measure given by where the noise εik is drawn from a normal distribution with mean zero and variance σ2 (common to all race groups). Employers, however, observe a worker’s group affiliation and know the underlying distributions of ηgk + φik and εik. In this environment, employers set the wage of each worker at the worker’s expected marginal revenue product conditional on observed skills (sˆi1..,sˆik) and the worker’s group affiliation… In sum, our task-based model of discrimination predicts differential sorting patterns across race groups as group-specific forces such as discrimination and racial skill differences make the task returns differ by group.”
“…Contact: measures the extent that the job requires the worker to interact and communicate with others (i) within the organization or (ii) with external customers/clients or potential customers/clients. To create our measure of Contact tasks we use two 1998 O*NET work activity variables taken from Deming (2017). Specifically, we use the variables Job-Required Social Interaction (Interact) and Deal With External Customers (Customer). 4 Interact measures how much workers are required to be in contact with others in order to perform the job. Customer measures how much workers have to deal with either external customers (e.g., retail sales) or the public in general (e.g., police work). To make our measure of the Contact task content of an occupation, we take the simple average of Interact and Customer for each occupation. Occupations with high measures of Contact tasks include various health care workers, waiter/waitress, sales clerks, lawyers, various teachers, and various managers….”
"... Figure 1 shows the difference in log wages between Black and White workers conditional on age and education using our sample of Census/ACS individuals. In particular, for each year, we regress an individual’s log hourly wage on a race dummy and controls of age (fiveyear age dummies) and series of dummies indicating the individual’s accumulated level of education. Consistent with other findings in the literature, the demographically adjusted racial wage gap narrowed substantially between 1960 and 1980 but has been constant at a gap of roughly 20 log points since 1980.5 Our goal is to explain both the wage convergence in Figure 1 between 1960 and 1980 as well as its stagnation post-1980..."
Skills measure
“..Cognitive Skills (COG): We follow the literature and use the respondent’s standardized scores on the Armed Forces Qualifying Test (AFQT) as our measure of cognitive skills. The AFQT is a standardized test which is designed to measure an individual’s math, verbal and analytical aptitude. The test score was collected from all respondents in their initial year of the survey and was measured in both the 1979 and 1997 waves….”
“…Non-cognitive Skills (NCOG): We use the measures of non-cognitive skills created by Deming (2017). Deming (2017) uses questions pertaining to the Rotter Locus of Control Scale and the Rosenberg Self-Esteem Scale for the NLSY79 cohort to make a measure of noncognitive skills.8 Likewise, for the NLSY97 cohort Deming (2017) uses respondent answers (provided prior to entering the labor market) to the question “How much do you feel that conscientious describes you as a person?” to approximate respondents’ non-cognitive skill. Deming (2017)’s non-cognitive skill measures are expressed in z-score units….”
“…Social Skills (SOC): We again follow Deming (2017) to generate a unified measure of social skills using a standardized composite of two variables that measure extroversion in both waves. Specifically, for the NLSY79, we use self-reported measures of sociability in childhood and sociability in adulthood. Individuals were asked to assess their current sociability (extremely shy, somewhat shy, somewhat outgoing, or extremely outgoing) and to retrospectively report their sociability when they were age 6. For the NLSY97, we proxy for social skills using the two questions that were asked to capture the extroversion factor from the commonly-used Big 5 personality inventory. For each wave, we normalize the two questions so they have the same scale and then average them together. We then convert the measures into z-score units. Deming (2017) shows that these measures of social skills positively predict individual wages when they are adults even conditional on controlling for individual measures of cognitive skills (AFQT)….”
“…in Figure 2. Panels A and B, respectively, show the patterns excluding and including the vector… of demographic controls. The main take-away from the figure is that both the level difference in racial task gaps in 1960 and the subsequent time series trend differ markedly by task. The differences are especially pronounced when we compare the racial gaps in Abstract and Contact tasks.In the early 1960s, Black workers were systematically underrepresented both in occupations that required a high intensity of Abstract tasks and in occupations that required a high intensity of Contact tasks. In terms of magnitudes, in 1960 a one-standard deviation increase in the Abstract task contents of an occupation reduced the probability that an individual working in that occupation was Black by about 3 percentage points, and a one-standard deviation increase in the Contact task contents reduced the probability that the individual was Black by about 4 percentage points, both conditional on education. Over the last half a century, however, Black men have made significant progress relative to White men with respect to sorting into occupations that require Contact tasks, while they made no progress relative to White men in the extent to which they sort into occupations that require Abstract tasks.Whereas the racial gap in Abstract tasks remained essentially constant through 2000 and widened slightly after 2000, the large racial gap in Contact tasks that existed in 1960 has all but disappeared by 2018. This finding persists whether or not we control for individual age and education (Panel A vs. Panel B), although the level of the Abstract task gap narrows once we control for them. If discrimination took the form of co-workers and customers not wanting to interact with Black workers in 1960, the patterns in Figure 2 are consistent with that form of discrimination abating over time.10 Although we focus on such differences to a lesser extent, there were only small racial differences in the propensity for Black and White men to work in occupations that require Manual tasks. Likewise, there was some convergence in the propensity to work in occupations that require Routine tasks between 1960 and 1980, but large racial gaps still remain post-1980….”
“…Figure 4 reports estimates of the raw wage premium by task requirement for White men (Panel A) and the demographically adjusted Black-White gaps in the wage premium by task requirement (Panel B). Three main findings emerge from this figure. First, the average wage premium of Abstract tasks for White men was between 10 and 15 percent higher than the return to the other tasks in 1960. Moreover, the relative return of Abstract tasks has been increasing since 1980. This increase in the return to Abstract tasks has received lots of attention in the literature (Autor and Dorn (2013), Deming (2017)). Second, in contrast, the wage premium associated with Contact tasks was notably lower in the early 1960s and has not changed much since then. Finally, the racial gaps in the wage premiums to tasks are relatively small and roughly constant over time. Most of the racial gaps in the average task returns are slightly negative suggesting that the task return to Black men are systematically smaller than White men of similar age and education….”
“…The results are shown in Table 1. Each column comes from a separate regression projecting an individual’s cognitive, non-cognitive or social skills on the relative task content of the occupation in which they work...Table 1 highlights that individual skill supplies respond differentially to relative task demands. Workers with higher cognitive skills are more likely to match with jobs that require higher Abstract tasks and workers with higher social skills are more likely to match with jobs that require higher Contact tasks. For example, occupations where their relative Abstract task requirement is one-standard deviation higher attract workers who score approximately 0.2 standard deviations higher on cognitive tests. Occupations where their Contact tasks requirement is one-standard deviation higher attract workers whose social score is 0.08 standard deviation higher. The results in this table highlight that the NLSY skill measures are informative about the types of jobs into which individuals sort. Having shown that the occupational task measures are associated with particular NLSY skill measures within the sample of White men, we now explore differences in pre-labor market skills between Black and White men within the NLSY cohorts.13 Table 2 reports the racial gap in cognitive, non-cognitive, and social skills with various controls for the two separate NLSY samples. The first column for each sample includes all NLSY respondents in the sample without conditioning on employment; each of these samples has only one NLSY respondent per regression. The remaining columns pool over all years and only include individuals that were working. The second column within each sample adds no further controls, while the third column controls for the individual’s maximum level of education and the last column controls also for their occupation The main takeaway from this table is that the racial gap in cognitive skills is large and narrows over time (especially for working men), whereas the gaps in non-cognitive and social skills are relatively small and constant over time. As seen from Table 2, Black men from the NLSY79 have AFQT scores that are 1.20 standard deviations lower than White men from the NLSY79 cohort. The gap declines to 0.93 when we additionally control for racial education and occupation differences. The race gap in AFQT scores narrowed somewhat between the NLSY79 and NLSY97 cohorts but was still large at -0.58 standard deviations conditional on education and occupation in the later period. On the other hand, the racial skill gaps in non-cognitive and social skills conditional on education and occupation were close to zero for both NLSY cohorts….”
"... Once we control for the rising return to cognitive skills over time, however, we find a strong convergence in racial wage gaps post 1980. Specifically, in column 3, we control for time-varying return to just cognitive skills. In this column, we find a narrowing of the racial wage gap relative to the 1980s of about 4 log points in the 1990s, about 9 log points in the 2000s, and about 10 log points in the 2010s. The results are nearly identical when we additionally control for time-varying returns to non-cognitive and social skills (column 4).27 As suggested by our model, conditioning out the effects of time-varying skill returns - the rising return to cognitive skills in particular - unveils the convergence in the racial wage gap due to changing race-specific factors. Strikingly, the magnitude of the convergence we estimate in the NLSY between 1980 and 2018 once properly controlling for the changing returns to skills (column 4 of Table 4) is close to the magnitude we estimate from our structural model(Panel B of Figure 7)..."
Good factoid in table four, "... the wage return to cognitive skills is higher for Black men than for White men with the same occupation and education...." implies higher skilled black men see a wage premium.
"... First, in Table 4 above, we exploit the panel structure of the NLSY and show that controlling for unmeasured traits by including individual fixed effects hardly affects the estimated changes in the racial wage gap over time (compare columns 1 and 2 of Table 4). This suggests that omitted skills play little role in the evolution of the racial wage gap over the last forty years. Second, in the appendix, we examine whether the labor market returns to skills differ between Black and White men in the NLSY. We find that the labor market returns to social skills are similar between Black and White men. This finding is consistent with there being no differential bias between Black and White men with respect to predicting Contact task efficiency from measured traits. On the other hand, consistent with the findings in Neal (2006), the wage return to cognitive skills is higher for Black men than for White men with the same occupation and education. This is suggestive of the possibility that missing traits associated with Abstract tasks differ systematically between Black and White men...."
Bottomline"...There are two important results highlighted in the paper. First, our paper provides an explanation for both why the Black-White wage gap narrowed between the 1960s and 1970s and why the racial wage gap has remained constant thereafter. We find that the constant wage gap between Black and White men post-1980 is due to two offsetting effects. Both the racial skill gap narrowed and taste-based discrimination fell post-1980 resulting in the wages of Black men converging to those of White men, all else equal. However, during the same period, the return to Abstract skills rose disadvantaging Blacks relative to Whites. This latter effect resulted in increasing racial wage gaps during the 1980-2018 period. The magnitude of these two effects were roughly similar resulting in a relatively constant racial wage gap post-1980. On the other hand, we show that the relative wage gains of Black men relative to White men during the 1960 to 1980 period stemmed solely from declining discrimination and a narrowing of racial skill gaps; changing task prices did not undermine any of these gains during this earlier period. We also provide a road map to empirical researchers looking to uncover changing race specific factors in micro data. Second, our paper establishes that the declining racial gap in Contact tasks between 1960 and 2018 is a good proxy for declining taste-based discrimination during this period.We motivated the introduction of this novel task measure by conjecturing ex-ante that occupations which require many interactions with others are more likely to be susceptible to taste-based discrimination; our model and data work confirm this conjecture ex-post. Specifically, the fact that there are very small racial gaps in social skills - combined with the fact that measures of pre-labor market social skills in the NLSY are highly predictive of subsequent entry into occupations that require Contact tasks - implies that racial gaps in Contact tasks must stem almost entirely from taste-based discrimination and very little from racial skill differences or statistical discrimination. Our model thus implies that the changes in racial gaps in Contact tasks over time is a good proxy for taste-based discrimination. To further provide evidence for this conclusion, we document that state-level racial gaps in Contact tasks correlate strongly with state-level survey measures of taste-based discrimination, while state-level racial gaps in Abstract tasks correlate with them only weakly….”
Ed Comment:I read the paper, so I read the definitions. I was looking for your explanation of what they were measuring after you read the paper, not theirs. after reading the paper, (and these snippets) it was unclear to me, and I was uncertain about what data they were truly using/measuring. Experience has taught me that people often say they did (or measured) X when they did something other than what they said they did. Sometimes they exaggerate. Sometimes they don’t fully understand what they did. Often, they assume they did X because they believe they did X or hoped to do X. In their equation, which variable is the AFQT test score? …or years of schooling? Is years of schooling their measure of analytical capability? I still don’t understand how blacks are discriminated against in analytical tasks when they are paid more on average of analytical tasks. (Are they paid more on average? That was unclear.) Blacks would have to have higher scores/more years of schooling or perform a mix of more analytical (more higher paid tasks). Which is it? My life experience is that the scores are lower on average. They are paid more for lower scores because high scoring blacks are highly sought after. But they may be paid less than higher scorers performing the same job. So I wouldn’t think the above would not be the case, i.e., blacks do not have higher scores on average or more analytical jobs. So I’m looking for your description for what the math in the study says in this regard. How does the study account for this?
Ben Comment:Ok, let me try again: ED: How do they measure the extent to which the job is analytical? First they take an occupation’s title (economist, researcher, scientist, burger flipper, etc.) and look up that occupation in DOT/O*NET. DOT/ONET matches an occupation’s title with the types of tasks that occupation performs and what skills are needed to perform them Firms are surveyed to see how much of a particular task is required of a particular occupation (the results have different scales so they’re harmonized). Then these tasks are classified into the 4 categories - by Autor & Dorn in the cases of Abstract, Routine, Manual and by Demings in the case of Contact. They determine which jobs are analytical by asking businesses which jobs require analytical tasks. NOTE - occupation is a control. The paper is about *tasks* and *skills*. So they merge DOT/ONET with ACS on occupation and regress wages onto the tasks/skills necessary. ED: What measures do they use for the pedigree of their analytical credentials/What measure of analytical prowess do they use? In the sections that examine the ACS, they’re using years of schooling. That’s section 3.1. In equation (1) on pg 12, the years of schooling is in X_(it) and is a control variable. They then move onto the NLSY data in section 3.2 (pg 16). These data are more robust in terms of demographic detail but also have a smaller sample. There are 2 waves of the NLSY - 79 and 97. In Appendix A.2, they explain both waves include the AFQT score. In general, you’re right, highly analytical people are unlikely to have taken this. In this particular case, however, everyone in the survey took it. Participants are chosen at random to participate in the survey. As far as I understand, once in the survey, everyone is given this test. Now, it’s true, surveys are likely to miss top top performers but this methodology does not suffer from the same selection bias using just military enrollment would. They also exclude observations with missing scores and reweight the entries. Section 6.3 is where they start using the NYSL for wage regressions: Equation 9 on pg 34. In this equation, one of the S^(NYLS)_(ki)’s is cognitive ability, and one of the variables in X_(it) is education. This regression specification does NOT interact race with cognitive ability so we are looking at the difference in wages of blacks/whites controlling for cognitive ability. Table A1 in appendix E shows the coefficient for black * Cognitive - and it’s positive. They also show the model imputed skills black*abstract (panel B) and that’s positive as well. The abstract skill variable, however, is a model implied variable constructed in section 7. To construct the variable, they take the average of the tasks done by each race. This is a complicated procedure to describe in words but the equations are on page 36+37. Essentially, they regress the occupational-average of observed task specific skills for whites (model generated) onto the average skill levels for whites observed in the NYSL. They then use the estimated white-black skills gap (table 1 pg 17) to impute the racial gaps in the task-specific skills in each occupation. Note: They do not show an explicit regression equation for this table but do explain the controls in the footnotes. I think I’ve covered the different measures of skills and how they’re used. ED: are they making some adjustment for the extent to which the job is analytical and for the pedigree of the person’s analytical credentials? They are NOT making adjustments for pedigree of institutions. They do control for years of education. However - they’re using actual pre-college test scores in their regressions or using those test scores to impute skills, so it’s unclear what the pedigree of the institution adds - especially if you believe those test scores are highly correlated with the pedigree of the institutions. ED: If black people are paid more than whites on average in analytics jobs, how do they measure discrimination in analytical jobs? They are interested in 2 types of discrimination, taste based and statistical. Taste based means - we don’t like black people. Statistical means - if a boss cannot perfectly observe a person’s analytical ability, they default to that person’s group’s mean of ability. This means, for smart blacks, if ability is noisy, they are assumed to have the average level of black intelligence and therefore earn lower wages. This type of discrimination is laid out in section 4, pg 22, the paragraph towards the bottom starting with “Conceptually, it would be useful to see the statical discrimination term ….” Essentially, if there is noise in determining someone’s analytical skill, they’re assumed to have the average and this discriminates against smart blacks. ED: In their equation, which variable is the AFQT test score? …or years of schooling? Is years of schooling their measure of analytical capability? I think I answered which variables are which above, but let me know if it’s still unclear. Their measure of analytical capabilities is the actual AFQT test score in the data. They ALSO include years of schooling as a control. ED: I still don’t understand how blacks are discriminated against in analytical tasks when they are paid more on average of analytical tasks. (Are they paid more on average? That was unclear.) Blacks would have to have higher scores/more years of schooling or perform a mix of more analytical (more higher paid tasks). Which is it? Blacks, on average, are paid less than whites (figure 1). Figure 1 was created with the ACS data so does NOT have the intelligence controls. Figure 4 breaks out the racial wage gap in terms of tasks. This graph, again, is created using the ACS data and does not include controls for tests scores. Figure 5 Panel A shows the model predicted racial wage gap. It’s model generated (not data) but also shows blacks, even smart ones, earning less. They do not have a break out of wages vs test scores from the NYSL like you’re describing. (This is where editorializing has to come in, sorry): As for the discrimination - remember regressions are holding all else equal. So what those positive coefficients mean is: Holding all else equal (occupation, wage, education, and ALL OTHER SKILLS), and increase in intelligence for a black guy will raise his wages more than the same sized increase in intelligence for a white guy. But the discrimination can come from all sorts of other things - like the occupations they have access to or the schooling they have access to or any number of other things. Additionally, if a white and a black guy have similar wages, conclusion of the paper says it’s likely the black guy is more skillful in Contact/Routine (tasks) or in Social/non-cognitive skills than the white guy; so raising his intelligence makes him *more* skillful than his white counterpart.