Macro Roundup

Daron Acemoğlu and Pascual Restrepo, "Tasks, Automation, and the Rise in US Wage Inequality," National Bureau Of Economic Research, June 2021, https://www.nber.org/papers/w28920

Three core findings, "…First, and in contrast to models of SBTC with factor-augmenting technologies, in our framework automation can have a negative effect on workers who are displaced from tasks they used to perform, and such changes can take place with limited increases in total factor productivity (TFP). Hence, real wage declines and slow productivity growth despite rapid automation are not puzzles within this framework. Second, we derive a simple equation linking wage changes of a demographic group to the task displacement it experiences, which forms the basis of our reduced-form analysis. Third, our framework implies that the task displacement experienced by a group can be measured by its employment share in routine tasks in industries undergoing automation. Moreover, the extent of automation in an industry can be inferred from declines in its labor share, thus providing an explanation for the relationship reported in the right panel of Figure 2….The second part of the paper documents a robust and negative (reduced-form) relationship between our measure of task displacement and real wages across groups of workers. Our baseline results use the decline in an industry labor share to infer the overall extent of automation and task displacement taking place in that industry. We show that close to 50% of the variation in labor shares (and measured task displacement) across industries is driven by automation technologies, including the adoption of industrial robots, specialized machinery and software. In line with this result, we find a very similar negative relationship between wage changes and task displacement using an IV strategy that exploits the component of the labor share decline driven by automation. We document that offshoring also contributes to task displacement (and conceptually, its effects on wages should be similar since it also works by displacing workers from their tasks), but quantitatively it accounts for a smaller portion of observed wage changes than automation. These results bolster our confidence that task displacement driven by automation technologies is responsible for the relationships depicted in Figure 2…We find that task displacement—incorporating general equilibrium effects—accounts for about 50% of the observed relative wage changes during this period and explains 80% of the observed increase in the college premium. These sizable distributional effects are accompanied by modest increases in the average wage level, GDP and TFP. For example, technologies causing task displacement only account for a 3.8% increase in TFP between 1980 and 2016. We thus conclude that automation can explain a sizable fraction of changes in wage structure and real wage declines in the data, while having a tiny impact on productivity growth…Relatedly, Figure 9 displays average task displacement and average real wage change by gender and education. It reveals that men without college degree have experienced the highest levels of task displacement as well as substantial real wage declines, while men and women with a post-graduate degree and women with a college degree have been subject to negligible task displacement and have enjoyed robust real wage growth. Once again, these patterns are explained by the component of task displacement driven by automation technologies…”

Impact across demographics groups, “…We now present descriptive statistics for our measure of task displacement at the level of demographic groups and take a preliminary look at its association with real wage changes. Figure 7 shows large differences across demographic groups in terms of task displacement between 1980 and 2016, with some experiencing a 25% reduction in their task shares, while others saw no change in theirs. Importantly, 95% of the variation in task displacement across groups is driven by our index of automation technologies, as can be seen from the left panel of Figure 7, which depicts task displacement by demographic group (computed using equation (12)) against the component driven by automation technologies (computed using equation (13)) The right panel of Figure 7 depicts task displacement by demographic group against a group’s real wage in 1980 and documents that task displacement has been particularly high during this period for groups in the middle of the wage distribution—thus playing both an unequalizing and a polarizing role. Figure 8 provides a first glimpse at the relationship between task displacement and real wage changes across demographic groups. The left panel plots the bivariate correlation between our task displacement measure and real wage changes between 1980 and 2016 (as in the right panel of Figure 2). The figure reveals a powerful negative relationship between task displacement and changes in real wages, with groups experiencing the highest levels of task displacement seeing their real wages fall or stagnate. The right panel displays a falsification exercise demonstrating that the relationship depicted in the left panel is not driven by secular declines in labor market fortunes of some demographic groups. It confirms that there is no correlation between our task displacement measure, which only uses post-1980 information, and real wage changes between 1950 and 1980—a period that predates major advances in automation. All demographic groups, including those who later on experienced adverse task displacement after 1980, enjoyed robust real wage growth, of about 50%, between 1950 and 1980…”

Research from Acemoglu/Restrepo finds that somewhere btw 50-70% of changes in the American wage structure btw 1980 and 2016 were driven by the relative wage declines of worker groups specialized in routine tasks in industries experiencing rapid automation,“…Our reduced-form evidence is based on estimating this equation and reveals a number of striking new facts. Most notably, we documented that between 50% and 70% of the changes in US wage structure between 1980 and 2016 are accounted for by the relative wage declines of worker groups specialized in routine tasks in industries experiencing rapid automation. In our first set of regression models, industry level task displacement is approximated by (the unexplained component of) labor share declines. We also estimate very similar results using explicit measures of industry-level automation and offshoring, confirming that our task displacement variable captures the effects of automation technologies (and to a lesser degree offshoring) rather than increasing markups industry concentration, or import competition. These alternative economic trends themselves do not appear to play a major role in the evolution of the US wage structure….”

Core idea, “..Figure 1 depicts some salient aspects of US developments: while the real wages of workers with a post-graduate degree rose, the real wages of low-education workers declined significantly. The real earnings of men without a high-school degree are now 15% lower than they were in 1980....One of the most powerful pieces of evidence offered in favor of SBTC (skill-biased technological change) in the 1990s was that increases in the demand for skills were pervasive across industries. This evidence also motivated the interpretation that industry-level differences in technology played a limited role in the transformation of US wage structureThis paper proposes an alternative approach for thinking about wage inequality. We argue that much of the changes in US wage structure are driven by the automation of tasks previously performed by certain types of workers in some industries (e.g., numerically-controlled machinery or industrial robots replacing blue-collar workers in manufacturing or specialized software replacing clerical workers). Workers who are not displaced from the tasks in which they have a comparative advantage, such as those with a postgraduate degree or women with a college degree, enjoyed real wage gains, while those, including low-education men, who used to specialize in tasks and industries undergoing rapid automation, experienced stagnant or declining real wage…”

“According to this framework, the demand for different types of workers comes from an aggregate production function of the form F(A H H,A L L), where H and L are employment levels of high-skill and low-skill workers, and A H and A L represent technologies (or equipment) augmenting these two types of workers. SBTC corresponds to technology becoming more favorable to high-skill workers (e.g., a greater increase in A H than in A L, provided that F has an elasticity of substitution greater than one)” It’s Ed’s smart guys working for smart guys, but formally modeled. Worth highlighting that at least some people are thinking along these lines.They claim declining labor share is a telltale sign of automation but this is not true. Note that a baseline assumption of their modeling is that displaced workers cannot compete with non-displaced workers. This is odd. I think this effect would actually exacerbate their results (since it would deepen the competition for low skill labor). Really useful to compare this paper to the Demings paper. They’re sort of 2 sides of the same coin, right? Demings is saying managerial and non-routine tasks are driving high-end wages up. Here, they’re saying routine/automatable tasks are sending low-end wages down. Good to put those together.”

Ben Comments:Look at figure 1 - 1980 is the highest real wages for working class people. Odd year to pick since there was a big run up in the 10 years before that. In fact, for men, only post-graduate degree is earning more than the 1980 group. It’s impossible to tell, from their graph, if 1980 is the anomaly or if the low (or negative) growth since then is the anomaly. It’s possible that substitutable low-skill labor just earns a low wage that’s relatively constant. This is Ed’s notion of the ratio of high skill to low skill.

I am surprised wages declined but I’m also suspicious for 4 reasons

• Does it include healthcare, pension, etc? if not, that’s cynical or stupid.
• the deflator (which one did they use?)
• the index year, especially given the extreme rise and fall of wage
• and, perhaps more importantly, the changing composition of workers in education categories. Today, only the dredges are dropouts for example.

New paper from Science Advances highlights the replication problem in science, “….We use publicly available data to show that published papers in top psychology, economics, and general interest journals that fail to replicate are cited more than those that replicate. This difference in citation does not change after the publication of the failure to replicate. Only 12% of post replication citations of nonreplicable findings acknowledge the replication failure. Existing evidence also shows that experts predict well which papers will be replicated. Given this prediction, why are nonreplicable papers accepted for publication in the first place? A possible answer is that the review team faces a trade-off. When the results are more “interesting,” they apply lower standards regarding their reproducibility….”

Core result, “…Our main finding is that papers that fail to replicate in (5-7) are cited more than those that are replicable. We find no significant change in citation trends, even after the publication of the failed replication. Notably, only a minority of publications after the failed replications were published acknowledge the failure. The “quality” of citations of papers that failed to replicate is similar to that of papers that were replicated: We do not find a difference in how often the citations of nonreplicable publications are cited by others or in the impact factor of the journals in which citations are published…”

The evidence, "...Figure 1 shows the distribution of total citation counts by the end of 2019 of the papers included in the replication projects, separately depending on replicability for Nature/Science, economics journals, and psychology journals. Replicability is measured according to the criterion that the replication project featured a P value of 0.05 or lower in a two-sided test, with an effect in the same direction as the original test….”

Scott Winship uses PSID data to track families across generations, core finding is that 20% of blacks versus 1% of whites have at least three consecutive generations living in poverty, “…We find that Black families are over 16 times more likely than White families are to experience three generations of poverty (defined as the bottom fifth of the income distribution). Three-generation poverty occurs among one in 100 Whites, but it describes the experience of one in five Black adults. Black Americans are 41 percent more likely to be in third-generation poverty than White adults are to be poor. The grandparents of Black adults had much lower incomes than the grandparents of their White counterparts had; this initial inequality has been compounded by lower rates of subsequent Black upward mobility out of poverty and by greater Black downward mobility. These patterns mean that poor Black and White adults today have dramatically different family poverty trajectories. Half of Blacks in the bottom fifth of the income distribution have parents and grandparents who were also poor, compared to just 8 percent of poor Whites. We show that the longer the time frame, the starker the racial gaps…”

Scott Winship, Christopher Pulliam, Ariel Gelrud Shiro, Richard V. Reeves, and Santiago Deambrosi, "Long Shadows: The Black White Gap In Multigenerational Poverty," American Enterprise Institute, June 2021, https://www.aei.org/wp-content/uploads/2021/06/Long-Shadows.pdf

The Data:

Shane Frederick, “Cognitive Reflection and Decision Making,” Journal Of Economic Perspectives, Fall 2005
https://pubs.aeaweb.org/doi/pdfplus/10.1257/089533005775196732

Gender differences, “…Men scored significantly higher than women on the CRT, as shown in Table 6.The difference is not likely due to a biased sampling procedure,because there were no significant sex differences for any other cognitive measure, except SATmath scores, for which there was a modest difference corresponding to national averages. Nor can it be readily attributed to differences in the attention or effort expended on the survey, since women scored slightly higher on the Wonderlic test, which was given under identical circumstances (included as part of a 45-minute survey that recruited respondents were paid to complete)…Four facts are noteworthy. First, CRT scores are more highly correlated with time preferences for women than for men; the low and high groups differ more. Second, as suggested by most prior research (Byrnes, Miller and Schafer, 1999 present an overview), women were considerably more risk averse than men, and this remains true even after controlling for CRT score. Third, for the selected risk items, CRT is as important as sex. In other words, high-scoring women behave almost identically to low-scoring men (compare the upper left and lower right cells within each of the five items in the lower panel). Fourth, in contrast to the pattern observed for the time preference items, CRT scores are more highly correlated with risk preferences for men than for women.The curious finding that CRT scores are more tightly linked with time preferences for women than for men, but are more tightly linked with risk preferences for men than for women held for the other tests of cognitive ability, as well. Expressed loosely, being smart makes women patient and makes men take more risks.14 This result was unanticipated and suggests no obvious explanation. The only related finding of which I am aware is in a study by Shoda, Mischel and Peake (1990), who found that the patience of preschool girls was strongly related to their subsequent SAT scores, but the patience of preschool boys was not….”

The outcomes, “…The proposition that the three CRT problems generate an incorrect “intuitive” answer is supported by several facts. First, among all the possible wrong answers people could give, the posited intuitive answers (10, 100 and 24) dominate. Second, even among those responding correctly, the wrong answer was often considered first, as is apparent from introspection, verbal reports and scribbles in the margin (for example, 10 cents was often crossed out next to 5 cents, but never the other way around). Third, when asked to judge problem difficulty (by estimating the proportion of other respondents who would correctly solve them), respondents who missed the problems thought they were easier than the respondents who solved them. For example, those who answered 10 cents to the “bat and ball” problem estimated that 92 percent of people would correctly solve it, whereas those who answered “5 cents” estimated that “only” 62 percent would. (Both were considerable overestimates.) Presumably, the “5 cents” people had mentally crossed out 10 cents and knew that not everyone would do this, whereas the “10 cents” people thought the problem was too easy to miss. Fourth, respondents do much better on analogous problems that invite more computation. For example, respondents miss the “bat and ball” problem far more often than they miss the “banana and bagel” problem: “A banana and a bagel cost 37 cents. The banana costs 13 cents more than the bagel. How much does the bagel cost?”…Table 2 shows the responses of the low and high CRT groups for each of the 17 items. The reported value is either the percentage choosing the patient option or the mean response. The subscripts are the total number of respondents in the low and high CRT groups who answered that item. The rightmost column reports the level of statistical significance of group differences—the p-values from a chisquare test (for dichotomous responses) or a t-test (for continuous responses). Those who scored higher on the CRT were generally more “patient”; their decisions implied lower discount rates. For short-term choices between monetary rewards, the high CRT group was much more inclined to choose the later larger reward (see items a and b). However, for choices involving longer horizons (items c through h), temporal preferences were weakly related or unrelated to CRT scores…. It appears that greater cognitive reflection fosters the recognition or appreciation of considerations favoring the later larger reward (like the degree to which the implied interest rate exceeds the rate offered by the market). However, it remains unclear whether cognitive reflection also influences other determinants of intertemporal choices (like pure time preference). CRT scores were unrelated to preferences for the massage and tooth-pull items, which were intended as measures of pure time preference. On the other hand, those in the low CRT group (the “cognitively impulsive”) were willing to pay significantly more for the overnight shipping of a chosen book (item l), which does seem like an expression of an aspect of pure time preference (the psychological “pain” of waiting for something desired)”…”

Evidence, “The CRT was administered to 3,428 respondents in 35 separate studies over a 26-month period beginning in January 2003. Most respondents were undergraduates at various universities in the midwest and northeast who were paid $8 to complete a 45-minute questionnaire that included the CRT and measures of various decision-making characteristics, like time and risk preferences.3 On the page on which the CRT appeared, respondents were told only: “Below are several problems that vary in difficulty. Try to answer as many as you can.”Table 1 shows the mean scores at each location and the percentage answering 0, 1, 2 or 3 items correctly.Most of the analyses that follow compare the “low” group (those who scored 0 out of 3) with the “high” group (those who scored 3 out of 3). The two “intermediate” groups (those who scored a 1 or 2) typically fell between the two extreme groups on whatever dependent measure was analyzed. Thus, focusing attention on the two “extreme” groups simplifies the exposition and analysis without affecting the conclusions. Since more of the respondents were college students from selective schools, the two “extreme” groups that formed the basis for most statistical comparisons were far more similar in cognitive abilities than two extreme groups formed from the general population. Thus, the group differences reported here likely understate the differences that would have been observed if a more representative sample had been used…”

Table 4 shows the relationship btw people who took the test and other scores (when available) for context.

Results

Shane Frederick devised the cognitive reflection test to measure as a short measure of cognitive ability, "...there is no good reason for ignoring the possibility that general intelligence or various more specific cognitive abilities are important causal determinants of decision making. To provoke interest in this neglected topic, this paper introduces a three-item “Cognitive Reflection Test” (CRT) as a simple measure of one type of cognitive ability. I will show that CRT scores are predictive of the types of choices that feature prominently in tests of decision-making theories, like expected utility theory and prospect theory. Indeed, the relation is sometimes so strong that the preferences themselves effectively function as expressions of cognitive ability—an empirical fact begging for a theoretical explanation”

More evidence that minimum wage increases reduced low skilled workers net welfare. In this case via several different channels, hour reduction, lower benefits eligibility and less consistent schedules based on evidence from a retailer in California.

Overall impact on workers, "...We further show that increasing the minimum wage can diminish worker welfare due to the changes in firms’ scheduling practices, even when it does not reduce the overall employment. For an average worker in a California store in our data,we estimate the net loss of welfare due to their reduction of hours, lower eligibility for benefits, and less consistent schedules (that resulted from a $1 increase in the minimum wage) to be at least $1,590 annually or 11.6% of the worker’s total wage compensation... loss in worker welfare due to the reduction in weekly hours, eligibility of benefits, and consistency of their work schedules is estimated to be $4,710 (= $1, 872 + $467 + $2, 371) annually, while the value that workers may derive from the reduced hours may be worth up to $3,120. As a result, the net loss in worker welfare is estimated to be at least $1,590 annually or 11.6% of the worker’s total wage compensation. We note that this net loss in worker welfare is entirely due to workers’ reduced eligibility for benefits and the consistency of theirs schedules not directly due to the reduced hours, when assuming workers can use all the reduced hours in their second jobs.

Impact on benefits, "This decrease in the average number of hours worked not only reduces total wages but also impacts workers’ eligibility for benefits. We show that the percentage of workers with weekly hours larger than 20 (who may be eligible for retirement benefits according to ERISA (1974)) and those with weekly hours larger than 30 (who may be eligible for healthcare insurance according to the Affordable Care Act (ACA)) decreases by 23.0% and 14.9%, respectively. These results suggest that, as the minimum wage increases, firms may strategically adjust their scheduling practice to reduce the number of workers who are eligible for benefits.This is consistent with the results in Clemens et al. (2018) which shows using survey data from the US that minimum wage increase reduces workers’ likelihood to receive healthcare insurance especially in the low wage sectors. To demonstrate the significant financial incentive for firms to do so, we provide a rough estimate of a store’s savings associated with reducing workers’ eligibility for benefits in Section 6.2.1."..."

Impact on change in consistency of workers schedule, "...Besides the direct reduction in total wage compensation and eligibility for benefits,we also show that increasing the minimum wage leads to less consistent worker schedules both in terms of the number of hours they work from one week to another and in terms of the timing of their shifts. In particular, for each $1 increase in the minimum wage, the absolute (relative) deviation in the number of weekly hours worked by each worker increases by up to 33.0% (6.7%). In addition, we show that the absolute (relative) deviation in the number of daily hours increases by up to 9.5% (2.0%). The absolute deviation in the starting time of the work shift increases by up to 9.8%..."
Implication for firms costs, "...In this subsection, we first provide an approximation of the savings in benefit costs for an average store in California in our data through the scheduling adjustments associated with the minimum increase of $1. This will help demonstrate the strong financial motivations for the stores to adopt the schedule adjustments. We then provide a rough quantification for the resultant total welfare loss of an average worker in a California store in our data from the scheduling adjustments. This will help demonstrate the importance of the scheduling implications when analyzing the effect of the minimum wage on worker welfare.... To this end, we consider the average annual cost of retirement benefits per worker for the retail store to be $1,228. The total annual cost of healthcare insurance and retirement benefits for the store is thus estimated to be $14,638 (= 15∗ (3, 800∗ 30.3%∗ 71.2% + 1, 228∗ 56.0%∗ 22.7%)). Based on our results in Section 5.1, when increasing the minimum wage by $1, the percentage of workers with weekly hours longer than 20 and 30 decreases by 23.0% and 14.9%, respectively. Meanwhile, the number of workers for an average store in California increases from 15 to 19. With such schedule adjustments, the total annual cost of healthcare insurance and retirement benefits for this average store is thus estimated to be $9,664 (= 19∗(3, 800∗15.4%∗71.2% + 1, 228∗33.0%∗22.7%)). Therefore, the annual reduction in healthcare and retirement benefit costs is $4,974 (= $14, 638 − $9, 664). Given that the total labor hours per week is 348 for an average store in California, the total increase in the annual labor cost as a result of $1 increase in the minimum wage is $18,096 (= 1∗52∗348), and the reduction in the benefit costs represents 27.5% of the total wage cost increase..."
Qiuping Yu, Shawn Mankad and Masha Shunko, "Evidence of The Unintended Labor Scheduling Implications of The Minimum Wage," Social Science Research Network, June 2021, https://papers.ssrn.com/sol3/papers.cfm

The sample, "...highly granular dataset of worker schedules from a medium-sized chain of fashion retail stores in the United States. Specifically, we study worker scheduling and minimum wage data from 2015 to 2018 for5,760 workers at 45 stores in California and 17 stores in Texas. All stores share the same brand. Our data include all workers employed at the stores. They are all paid by the hour, and most of them are paid with the minimum wage. The key advantage of our dataset is that it allows the precise measurements of labor hours for a store as a whole and for all individual workers within a store, as well as the timings of the shifts for each worker..."
The results, "...Our results show that the minimum wage has no statistically significant impact on the total labor hours employed at the stores, which is consistent with the literature on the employment effect of minimum wage, especially in the non-tradable sectors (e.g., retail and service)…However, we show that the way in which the stores allocate these hours among their workers does change. Specifically, when the minimum wage increases by $1, the number of workers scheduled to work each week goes up by 27.7%, while the average hours per worker per week decrease by 20.8%. For an average store in California, for example, these changes translate into four extra workers per week and five fewer hours per worker per week. This means, for an average worker in California paid the minimum wage, her total wage compensation is reduced by 13.6% when increasing the minimum wage from $11 to $12…”