The Growing Importance of Social Skills in the Labor Market*

Abstract

The labor market increasingly rewards social skills. Between 1980 and 2012, jobs requiring high levels of social interaction grew by nearly 12 percentage points as a share of the U.S. labor force. Math-intensive but less social jobs—including many STEM occupations—shrank by 3.3 percentage points over the same period. Employment and wage growth were particularly strong for jobs requiring high levels of both math skill and social skills. To understand these patterns, I develop a model of team production where workers “trade tasks” to exploit their comparative advantage. In the model, social skills reduce coordination costs, allowing workers to specialize and work together more efficiently. The model generates predictions about sorting and the relative returns to skill across occupations, which I investigate using data from the NLSY79 and the NLSY97. Using a comparable set of skill measures and covariates across survey waves, I find that the labor market return to social skills was much greater in the 2000s than in the mid-1980s and 1990s.

We can never survey our own sentiments and motives, we can never form any judgment concerning them; unless we remove ourselves, as it were, from our own natural station, and endeavour to view them as at a certain distance from us. But we can do this in no other way than by endeavouring to view them with the eyes of other people, or as other people are likely to view them.

—Adam Smith, The Theory of Moral Sentiments (1759)

I. Introduction

A vast literature in economics explains increasing returns to skill as a product of the complementarity between technology and high-skilled labor, or skill-biased technological change (SBTC) (e.g., Bound and Johnson 1992; Katz and Murphy 1992; Juhn, Murphy, and Pierce 1993; Acemoglu and Autor 2011). Beginning in the 1990s, the labor market “hollowed out” as computers substituted for labor in middle-skill routine tasks and complemented high-skilled labor, a phenomenon referred to as job polarization (Autor, Levy, and Murnane 2003; Autor, Katz, and Kearney 2006; Goos, Manning, and Salomons 2014; Michaels, Natraj, and Van Reenen 2014).

However, while job polarization implies rising demand for skilled labor, there has been little or no employment growth in high-paying jobs since 2000, and this slow growth predates the Great Recession (Acemoglu and Autor 2011; Beaudry, Green, and Sand 2014, 2016). Beaudry, Green, and Sand (2016) show evidence of slow growth in cognitive skill–intensive occupations in the U.S. labor market during the 2000s, and Castex and Dechter (2014) find smaller returns to cognitive test scores in the 2000s compared to the 1980s. These findings are especially puzzling in light of the rising heterogeneity in worker-specific pay premiums found in studies that use matched employer-employee data (Card, Heining, and Kline 2013; Card, Cardoso, and Kline 2016). If technological change is skill-biased, why have the returns to cognitive skill not increased over the past decade?

One possible explanation is that weak growth in high-skilled jobs is caused by a slowdown in technological progress. Beaudry, Green, and Sand (2016) argue that the slowdown in demand for cognitive skill can be explained as a boom-and-bust cycle caused by the progress of information technology (IT) from adoption to maturation, and Gordon (2012) shows that innovation and U.S. productivity growth slowed down markedly in the early 2000s.

On the other hand, Brynjolfsson and McAfee (2014) discuss advances in computing power that are rapidly expanding the set of tasks that machines can perform. Many of the tasks that they and others highlight—from automated financial management and tax preparation to legal e-discovery to cancer diagnosis and treatment—are performed by highly skilled workers (Levy and Murnane 2012; Brynjolfsson and McAfee 2014; Remus and Levy 2016). Thus another possibility is that computer capital is substituting for labor higher up in the skill distribution, redefining what it means for work to be “routine” (Autor 2014; Lu 2015).

Figure I investigates this possibility by showing relative employment growth between 2000 and 2012 for the set of high-skilled, “cognitive” occupations that are the focus of Beaudry, Green, and Sand (2016).¹ The upper panel of Figure I focuses on science, technology, engineering, and mathematics (STEM) jobs, while the lower panel shows all other cognitive occupations.

Figure I

Change in Relative Employment for Cognitive Occupations, 2000–2012

Each row presents 100 times the change in employment share between 2000 and 2012 for the indicated occupation. Consistent occupation codes for 1980 to 2012 are updated from Autor and Dorn (2013) and Autor and Price (2013) and consolidated to conserve space (see the Online Data Appendix for details). Source: 2000 census and 2011–2013 ACS.

Figure I shows clearly that the relative decline in high-skilled employment over the past decade is driven by STEM jobs. STEM jobs shrank by a total of 0.12 percentage points as a share of the U.S. labor force between 2000 and 2012, after growing by 1.33 percentage points over the previous two decades. By comparison, all other cognitive occupations grew by 2.87 percentage points between 2000 and 2012, which surpasses the growth rate of 1.99 percentage points in the previous decade. Most important, the fastest growing cognitive occupations—managers, teachers, nurses and therapists, physicians, lawyers, even economists —all require significant interpersonal interaction.

In this article, I show that high-paying jobs increasingly require social skills. Technological change provides one possible explanation. The skills and tasks that cannot be substituted away by automation are generally complemented by it, and social interaction has—at least so far—proven difficult to automate (Autor 2015). Our ability to read and react to others is based on tacit knowledge, and computers are still very poor substitutes for tasks where programmers don’t know “the rules” (Autor 2015). Human interaction requires a capacity that psychologists call theory of mind—the ability to attribute mental states to others based on their behavior, or more colloquially to “put oneself into another’s shoes” (Premack and Woodruff 1978; Baron-Cohen 2000; Camerer, Loewenstein, and Prelec 2005).

I begin by presenting a simple model of team production between workers. Workers perform a variety of tasks on the job, and variation in productivity generates comparative advantage that can be exploited through specialization and “task trade.” I model cognitive skills as the mean of a worker’s productivity distribution and social skills as a reduction in trading costs. Workers with higher social skills can specialize and “trade tasks” with other workers more efficiently. This takes on the structure of a Ricardian trade model, with workers as countries and social skills as inverse “iceberg” trade costs as in Dornbusch, Fischer, and Samuelson (1977) and Eaton and Kortum (2002).²

The model generates several predictions, which I investigate using data from the National Longitudinal Survey of Youth 1979 (NLSY79). I first demonstrate that there is a positive return to social skills in the labor market and that cognitive skill and social skill are complements in a Mincerian wage equation. This follows recent evidence from Weinberger (2014), who finds growing complementarity over time between cognitive skills and social skills using different data sources. Complementarity emerges naturally in the model, because the value of lower trade costs increases in a worker’s average productivity (i.e., cognitive skill).³ Importantly, I do not find complementarity between cognitive skill and widely used measures of “noncognitive” skills (e.g., Heckman, Stixrud, and Urzua 2006).

The model provides a key link between social skills and routineness through the variance of productivity over workplace tasks. Some high-skilled occupations (such as a computer programmer or engineer) require the repeated execution of explicit rules, whereas others are less structured and require a diverse range of tasks (such as a manager or consultant). I model this as an increase in the variance of productivity over the tasks that workers perform on the job. Higher variance in productivity broadens the scope for gains from “task trade” and thus increases the return to social skills.

Although I cannot directly measure the variance of workplace tasks, I use two empirical analogs. First, I compare the returns to social skills across occupations that vary in their routineness, as measured by data from the Occupational Information Network (O*NET). I find that workers with higher social skills self-select into nonroutine occupations, and this sorting leads to within-worker wage gains that are increasing in social skills.⁴ These empirical patterns are consistent with the predictions of the model. Notably, I find no evidence of greater returns to social skills in math-intensive occupations.

Next, I draw on a large literature in organizational economics which shows that all occupations are becoming less routine over time. Information and communication technology (ICT) has shifted job design away from rigid categorization and toward increased job rotation and worker “multitasking” (Bresnahan 1999; Lindbeck and Snower 2000; Caroli and Van Reenen 2001; Bloom and Van Reenen 2011). Case studies of ICT implementation show that computerization leads to the reallocation of skilled workers into flexible, team-based settings that facilitate adaptive responses and group problem solving (e.g., Autor, Levy, and Murnane 2002; Bresnahan, Brynjolfsson, and Hitt 2002; Bartel, Ichniowski, and Shaw 2007). This literature shows a clear link between the computerization of the labor market and the decline of routine work. Yet the link between the increased variability of workplace tasks, team production, and social skills has not previously been explored.

I investigate the growing importance of social skills in two ways. First, I present evidence of increasing relative demand for social skills in the U.S. labor market. Between 1980 and 2012, social skill–intensive occupations grew by 11.8 percentage points as a share of all jobs in the U.S. economy. Wages also grew more rapidly for social skill–intensive occupations over this period. I find that employment and wage growth has been particularly strong in occupations with high math and social skill requirements. In contrast, employment has declined in occupations with high math but low social skill requirements, including many of the STEM jobs shown in Figure I. Contemporaneous trends in the labor market over this period such as offshoring, trade, and shifts toward the service sector can partially—but not completely—explain these patterns.⁵

Second, I test directly for the growing importance of social skills by comparing the returns to skills in the NLSY79 and the NLSY97 surveys. Comparing cohorts between the ages of 25 and 33 who entered the labor market in the mid-1980s versus the mid-2000s, I find that social skills are a significantly more important predictor of full-time employment and wages in the NLSY97 cohort. Cognitive skills, social skills, and other covariates are similarly defined across survey waves, and the results are robust to accounting for other contemporaneous trends such as increasing educational attainment and women’s labor force participation. Finally, I show that the within-worker wage gain from sorting into social skill-intensive occupations is much greater in the NLSY97 cohort.

I am aware of few other papers that study social skills. In Borghans, Ter Weel, and Weinberg (2014), there are “people” jobs and “nonpeople” jobs and the same for skills, with workers sorting into jobs based on skills and relative wages. Kambourov, Siow, and Turner (2013) develop a model where high levels of “relationship skill” (as measured by a worker’s occupation) are associated with stable marriage and employment outcomes. McCann et al. (2015) develop a multisector matching model with teams of workers who specialize in production tasks and a manager who specializes completely in communication tasks. In contrast, there are no communication tasks in my model, nor are there formal teams.⁶ This is consistent with case studies of modern teamwork, where workers are organized into temporary, fluid and self-managed groups to perform customized sets of tasks (e.g., Lindbeck and Snower 2000; Hackman 2002; Bartel, Ichniowski, and Shaw 2007; Edmondson 2012).

While the model considers teamwork in production, one can view many customer-oriented occupations—consulting, health care, teaching, legal services—as requiring joint production between worker and customer. Katz (2014) discusses growing demand for artisanal workers who can provide a creative, personal touch and customize production to the needs of clients. Social skills in production will be important for customer service occupations to the extent that the final product is uncertain and crafted specifically for the needs of the client.

Are social skills distinct from cognitive skills, or are they simply another measure of the same underlying ability? When surveyed, employers routinely list teamwork, collaboration, and oral communication skills as among the most valuable yet hard to find qualities of workers (e.g., Casner-Lotto and Barrington 2006; Jerald 2009). In 2015, employers surveyed by the National Association of Colleges and Employers (NACE) listed “ability to work in a team” as the most desirable attribute of new college graduates, ahead of problem solving and analytical/quantitative skills (NACE 2015). Tests of emotional and social intelligence have been developed and validated by psychologists (Salovey and Mayer 1990; Mayer, Caruso, and Salovey 1999; Baron-Cohen et al. 2001; Goleman 2006). Woolley et al. (2010) show that a test designed to measure social intelligence predicts team productivity even after controlling for the average intelligence of team members.⁷

A growing body of work in economics documents the labor market return to “noncognitive” skills, including social skills and leadership skills (Kuhn and Weinberger 2005; Heckman, Stixrud, and Urzua 2006; Lindqvist and Vestman 2011; Borghans, Ter Weel, and Weinberg 2014). This article builds on the seminal observation of Heckman (1995) that since measured cognitive ability (i.e., g) explains only a small fraction of the variation in earnings, productivity is likely influenced by multiple dimensions of skill. Subsequent work, summarized in Heckman and Kautz (2012), finds that “noncognitive” or “soft” skills explain important variation in adult outcomes. This article should be viewed as an attempt to extend and formalize the definition of one particular dimension of “soft” skills—the ability to work with others.

The remainder of the article proceeds as follows. Section II presents the model and develops specific empirical predictions. Section III describes the data. Section IV explores the predictions concerning the returns to social skill across occupations using the NLSY79. Section V studies the growing importance of social skills over time, using both census/ACS data and a comparison of the returns to skills in the NLSY79 and NLSY97. Section VI concludes. All appendix material, including supplementary tables and figures, a more detailed data description, and proofs for the model—can be found in the Online Appendix.

II. Model

In a standard human capital model, worker skill takes a simple factor-augmenting form, where the output of worker j is increasing in some measure of skill (such as cognitive ability or education) A_j times L_j, the quantity of labor supplied:

\begin{equation} y_{j}=A_{j}L_{j}. \end{equation}

(1)

Recent work has enriched the standard model by drawing a distinction between skills and job tasks (e.g., Autor, Levy, and Murnane 2003; Acemoglu and Autor 2011; Autor and Handel 2013). In the spirit of this “task framework,” consider the following modification of the human capital model:

\begin{equation} y_{j}\left(i\right)=A_{j}\alpha _{j}\left(i\right)l_{j}\left(i\right), \end{equation}

(2)

where y_j(i) specifies the production function for task i as worker j’s cognitive skill A_j (still taking the factor-augmenting form) times a task-specific productivity parameter α_j(i) times labor supplied to task i.

Any job can be separated into an infinite number of discrete tasks that must be performed jointly to produce some final good Y. Following Acemoglu and Autor (2011), I assume that workers perform a continuum of tasks indexed over the unit interval according to a Cobb-Douglas technology:

\begin{equation} Y_{j}=\mathrm{exp}[\int _{0}^{1}\mathrm{ln}y_{j}(i)di]. \end{equation}

(3)

For simplicity, I assume that each worker supplies one unit of labor inelastically:

\begin{equation} \int _{0}^{1}l_{j}(i)di=L_{j}=1. \end{equation}

(4)

Equation (2) allows two workers with the same cognitive skill level A_j to vary in their productivity over individual tasks. This suggests that workers can specialize in the production of tasks in which they have a comparative advantage.

To think about how the productivity gains from specialization can be realized, I develop a model in the spirit of Ricardo (1891). In Ricardo (1891), countries specialize in the production of goods and trade with each other for mutual benefit. In this model, workers can increase their total output Y_j by producing tasks in which they have comparative advantage and then trading for mutual benefit, just as countries trade goods in Ricardo’s classic formulation. Thus I conceive of teamwork as “trading tasks.”

Applying the Ricardian framework to task trade between workers yields two important benefits. First, it provides an explanation for why social skills matter that is grounded in economic theory. I argue that social skills are valuable because they reduce the cost of “trading tasks” with other workers.

Specifically, let S_i,n ∈ (0, 1) be a depreciation factor that is applied proportionately to any trade in tasks between workers—S_{i, n} = S_i * S_n for i ≠ n. Moreover let S_{i, i} = 1, |$\mathcal {\forall }i$| so workers can trade costlessly with themselves. Workers with higher social skill pay a lower coordination cost to trade with other workers. This allows them to earn higher wages by specializing in their most productive tasks and trading their output with others.⁸ Workers with high cognitive skill A_j but low social skill S_j have high average productivity, but will perform “too many” tasks themselves rather than working in a team.

The second important feature of the model is that it generates intuitive predictions about when social skills will matter. The return to social skills and the benefits of task trade will be increasing in the variance of productivity over tasks (the α_j’s), because higher productivity dispersion increases the scope for gains from trade. To see this, consider the limiting case where α_j(i) takes the same value for all tasks i. In this case, equation (2) collapses to equation (1) and becomes the standard human capital model. With zero variance in productivity over tasks, cognitive skill A_j is the sole determinant of relative productivity and there are no gains from trade.

If a worker has very low social skills, she will produce the same combination of tasks regardless of her comparative advantage relative to others. On the other hand, the task mix of a worker with high social skills will be quite sensitive to changes in the relative productivities of her co-workers. Thus another sensible interpretation of S_j is that it represents flexibility.

Here I develop the case with bilateral task trade between two workers. This two-worker model is isomorphic to the two-country Ricardian trade model of Dornbusch, Fischer, and Samuelson (1977). Thus I keep the presentation brief and refer the reader to the Online Appendix for proofs and more detailed exposition.⁹ With only two workers, two dimensions of skill, and one final good, the model developed here is highly stylized. However, it yields a set of intuitive predictions that help guide the empirical work below.

II.A. Setup

Consider a competitive market where Y is the unique final good—produced according to equation (3)—and labor is the only factor of production. Identical firms hire pairs of workers and pay market wages that are equal to output Y_j times an exogenous output price P*. Thus workers maximize output Y_j, subject to the labor supply constraint in equation (4). Firms maximize profits, defined as total revenue [P * (Y₁ + Y₂)] minus the wages paid to workers (w₁ + w₂).

Because the order of tasks over the unit interval is arbitary, it is convenient to index tasks in order of decreasing comparative advantage for worker 1 (i.e., |$\frac{\alpha _{1}(0)}{\alpha _{2}(0)}>\cdots >\frac{\alpha _{1}(i)}{\alpha _{2}(i)}>\cdots >\frac{\alpha _{1}(1)}{\alpha _{2}(1)}$|⁠). Define the comparative advantage schedule over tasks as:

\begin{equation} \gamma (i)\equiv \frac{A_{1}\alpha _{1}(i)}{A_{2}\alpha _{2}(i)}, \end{equation}

(5)

with γ΄(i) < 0 by assumption.

For concreteness, I assume that the comparative advantage schedule takes the form:

\begin{equation} \gamma (i)=\bar{A}\mathrm{exp}(\theta (1-2i)), \end{equation}

(6)

with |$\bar{A}=\frac{A_{1}}{A_{2}}$|⁠. This functional form for γ(i) can be derived from an underlying process where worker productivity in task i is drawn from a log-normal distribution with a mean that is increasing in cognitive skill A_j, and a variance that is increasing in θ.¹⁰

A_j indexes worker j’s mean productivity across tasks, while θ indexes the variance of task productivities and the steepness of the comparative advantage schedule. When θ = 0, workers with higher cognitive skill are more productive in all tasks by the same ratio |$\bar{A}$|⁠. In that case, there is no comparative advantage and thus no possibility for gains from trade. Thus this model nests the standard human capital model as a special case when θ = 0. As θ increases, productivity over individual tasks is more dispersed.

II.B. Equilibrium with Costless Trade

Each worker maximizes output by obtaining tasks from the lowest-cost producer, including herself. Workers trade tasks with each other at “prices” defined by efficiency units of labor, with a budget equal to each worker’s labor supply constraint in equation (4). The worker-specific price of task i is:

\begin{equation} p_{j}(i)=\frac{w_{j}}{A_{j}\alpha _{j}(i)}, \end{equation}

(7)

where w_j is the endogenously determined wage paid to worker j for a unit of labor. The equilibrium price for each task is the lower of the two offered prices: p(i*) = min {[p₁(i), p₂(i)]}. Since γ΄(i) < 0 and there is a continuum of tasks, it is clear that in equilibrium there will be a marginal task i* such that

\begin{equation} \omega =\gamma (i^{*}), \end{equation}

(8)

where |$\omega =\frac{w_{1}}{w_{2}}$|⁠. Worker 1 will perform all tasks in the interval [0, i*] and worker 2 will perform all tasks in the interval [i*, 1].

The equilibrium wage w_j is also determined by the demand for tasks, which comes out of the production function for the final good Y in equation (3). In equilibrium, the price-adjusted quantity of output for the marginal task i* must be the same for both workers. This, combined with the fact that the Cobb-Douglas production function implies that the same share of output is paid to each task, yields the following equilibrium condition for the demand for tasks:

\begin{equation} \omega =\frac{i^{*}}{1-i^{*}}. \end{equation}

(9)

Equilibrium is found by setting the downward-sloping comparative advantage condition in equation (8) equal to the upward-sloping labor demand condition in equation (9), which yields a unique marginal task as a function of worker skills and the variance of productivity θ.¹¹

The relative wage ω is clearly increasing in the task threshold—for example, if A₁ = A₂, then |$i^{*}=\frac{1}{2}$| and ω = 1. Equilibrium wages for worker 1 are given by:

\begin{equation} w_{1}=P^{*}A_{1}^{i^{*}}(A_{2}\omega )^{1-i^{*}}\mathrm{exp}\left[\int _{0}^{i^{*}}\mathrm{ln}\alpha _{1}(i)di+\int _{i^{*}}^{1}\mathrm{ln}\alpha _{2}(i)di \right]. \end{equation}

(10)

The expression for worker 2 is very similar. Thus wages are increasing in a worker’s own skill A_j as well as the skill of her co-worker. Moreover, the gains from trade are also priced into absolute wages and are increasing in θ.¹²

II.C. Equilibrium with Social Skills

With only two workers, we can define S* = S₁ * S₂ as the (symmetric) cost of trading tasks between the two workers, with self-trade normalized to 1 as above. Thus worker 1 will produce her own tasks rather than trading if:

\begin{eqnarray} p_{1}(i) & < & p_{2}^{S}(i)\nonumber \\ \frac{w_{1}}{A_{1}\alpha _{1}(i)} & < & \frac{w_{2}}{S^{*}A_{2}\alpha _{2}(i)}\nonumber \\ \omega & < & \frac{\gamma (i)}{S^{*}}. \end{eqnarray}

(11)

Likewise, worker 2 will produce her own tasks if ω > S*γ(i). Thus in equilibrium there will be two task thresholds, defined by:

\begin{eqnarray} \gamma (i^{H}) & = & S^{*}\omega \end{eqnarray}

(12)

\begin{eqnarray} \gamma (i^{L}) & = & \frac{\omega }{S^{*}}. \end{eqnarray}

(13)

Since γ΄(i) < 0, it is clear that i^H > i* > i^L when S* < 1.

Tasks in the interval [0, i^L] will be produced exclusively by worker 1, tasks in the interval [i^H, 1] will be produced exclusively by worker 2, and tasks in the interval [i^L, i^H] will be nontraded (produced by both workers for their own use).

As S* → 1, i^L and i^H converge to a single value i*. For any values i^L ≤ 0 and i^H ≥ 1, workers will maximize output by producing all tasks themselves (i.e. autarky).

Figure II, Panels A and B provide a visual illustration of the equilibrium task thresholds under two different values of θ. Figure II, Panel A shows the case where θ is lower and the comparative advantage schedule is flatter, and Panel B shows the impact of increasing θ and making the comparative advantage schedule steeper.

Figure II

Equilibrium Task Thresholds with Different Values of Theta

Panel A illustrates the equilibrium task thresholds i^L and i^H from the model in Section II when |$S^* = \frac{2}{3}$|⁠, θ = 1, and ω* = 1. Panel B illustrates the equilibrium task thresholds i^L and i^H from the model in Section III when |$S^* = \frac{2}{3}$|⁠, θ = 2, and ω* = 1 (see the text for details).

Figure II shows that—all else equal—the size of the nontraded zone [i^L, i^H] is decreasing in θ. This can also be demonstrated by solving equations (12) and (13) for ω, which yields:

\begin{equation} i^{H}-i^{L}=-\frac{\mathrm{ln}S^{*}}{\theta }. \end{equation}

(14)

Equation (14) shows that the size of the range of nontraded tasks (inversely) scales the gains from trade. When trade is costless (i.e., S* = 1), i^L = i^H. On the other hand, equation (14) also shows that there are many values of S* and θ for which autarky is preferable (i.e., whenever i^H − i^L > 1).

As in the case of costless trade, equilibrium can be obtained by solving for the intersection between the two comparative advantage schedules in equations (12) and (13) and the demand for tasks, which is given simply by:

\begin{equation} \omega =\frac{i^{L}}{1-i^{H}}. \end{equation}

(15)

Combining equations (12), (13), and (15) gives two functions with two unknowns (i^H and i^L) and three parameters (⁠|$\bar{A}$|⁠, S*, and θ). Plotting these two implicit functions in the (i^L, i^H) space shows that their intersection defines the unique equilibrium values of i^H and i^L.

Finally, equilibrium wages for workers 1 and 2 are given by:

\begin{equation} w_{1}=P^{*}A_{1}^{i^{H}}(S^{*}A_{2}\omega )^{1-i^{H}}\mathrm{exp}\left[\int _{0}^{i^{H}}\mathrm{ln}\alpha _{1}(i)di+\int _{i^{H}}^{1}\mathrm{ln}\alpha _{2}(i)di\right], \end{equation}

(16)

\begin{equation} w_{2}=P^{*}A_{2}^{1-i^{L}}(S^{*}A_{1}\omega ^{-1})^{i^{L}}\mathrm{exp}\left[\int _{0}^{i^{L}}\mathrm{ln}\alpha _{1}(i)di+\int _{i^{L}}^{1}\mathrm{ln}\alpha _{2}(i)di\right]. \end{equation}

(17)

II.D. Interpreting θ

The variance parameter θ admits at least two interpretations. The first concerns the task content of occupations. What kinds of jobs are characterized by greater productivity dispersion over tasks? One can interpret θ as a measure of predictability. Some jobs require workers to perform the same set of tasks repeatedly, whereas others are unpredictable or require a wide range of tasks depending on the situation.

Although existing data do not allow me to directly measure the variance of tasks for a particular occupation, the closest analog is routineness. Autor, Levy, and Murnane (2003) define a task as “routine” if it can be accomplished by following explicit programmed rules. Relatedly, Bresnahan (1999) argues that computers change the workplace by “organizing, routinizing and regularizing tasks that people- and paper-based systems did more intuitively but more haphazardly.” The idea behind both of these statements is that there is a well-established, correct way to perform some tasks. For example, tasks such as complex mathematical calculations require high levels of cognitive skill but are also routine according to this definition.

Thus one interpretation of θ is that it indexes the share of tasks for which there is no single best approach. As θ increases, a lower share of tasks are routine. Thus the return to social skills should be decreasing in the routineness of an occupation. I examine this prediction in Section IV by estimating variation in the returns to social skill across occupations at a particular point in time.

A second interpretation is that θ is a general production technology parameter that applies to all occupations, but is changing over time. Autor, Levy, and Murnane (2003) show that the United States has experienced relative employment declines in routine-intensive occupations since the 1970s, and Goos, Manning, and Salomons (2014) document this same pattern over a number of Western European countries.

One empirical limitation of this line of work is that it only measures shifts in the distribution of employment across occupations, not within them. Yet it is likely that all occupations are becoming less routine. Indeed the driving causal force in Autor, Levy, and Murnane (2003) is an exogenous decline in the price of computer capital, a phenomenon that presumably affects all occupations to some degree. Case studies that accompany quantitative work on SBTC focus on how occupations such as bank tellers and machinists have changed in response to computerization (Autor, Levy, and Murnane 2002; Bresnahan, Brynjolfsson, and Hitt 2002; Bartel, Ichniowski, and Shaw 2007).

In the model, any general increase in the variance of job tasks θ will lead to an increase in the return to social skills. Thus increases in the variability of workplace tasks should accompany increases in team production. The organizational economics literature strongly supports this conclusion. Studies of the impact of ICT suggest that job design has shifted away from the unbundling of discrete tasks and toward increased job rotation and worker “multitasking” (e.g., Bresnahan 1999; Lindbeck and Snower 2000; Bloom and Van Reenen 2011).

A key theme in studies of ICT and organizational change is the reallocation of skilled workers into flexible, team-based settings that facilitate group problem solving (e.g., Caroli and Van Reenen 2001; Autor, Levy, and Murnane 2002; Bresnahan, Brynjolfsson, and Hitt 2002; Bartel, Ichniowski, and Shaw 2007). Dessein and Santos (2006) develop a model where organizations optimally choose the extent to which employees are allowed to use discretion in response to local information—whether to follow a rigid script or to be “adaptive.” They show that when the business environment is more uncertain—which could be interpreted as a measure of θ—organizations endogenously allow for more ex post coordination among employees.

This literature suggests that the variance of job tasks has increased greatly over time, even within occupations. Thus if we interpret θ as a measure of nonroutineness, the return to social skills should have grown over time for workers in all occupations. In addition, we should be able to observe increases over time in the importance of jobs that require social interaction.

II.E. Empirical Predictions

The model generates several predictions, which I summarize here. The first four predictions concern variation in the returns to social skill across workers and occupations at a particular point in time:

• There is a positive labor market return to both cognitive skill and social skill. This is evident from the expressions for equilibrium wages in equations (16) and (17). I examine this prediction using data from NLSY79, which contains direct measures of worker skills.

• Cognitive skill and social skill are complements. This is true because the second derivatives of w₁ and w₂ with respect to A and S* are positive.¹³ Intuitively, social skills are relatively more valuable when a worker is more productive overall, because she has more of value to “trade” with her fellow worker. I examine this prediction by interacting measures of cognitive skill and social skill from the NLSY79 in a Mincerian earnings regression, with a positive interaction indicating complementarity. Weinberger (2014) finds evidence for growing complementarity between cognitive skills and social skills across two cohorts of young men. The model provides a theoretical foundation for those results. This prediction contrasts with existing models of job assignment where workers have multiple skills. Such models typically feature matching of workers to firms according to Roy-type selection, and skills are assumed to be additively separable for tractability (e.g., Heckman and Scheinkman 1987; Lindenlaub 2014; Lise and Postel-Vinay 2014). Although one could certainly write down a model that simply asserts that cognitive skill and social skill are complements, the model above develops complementarity from first principles.

• Workers with higher social skill sort into nonroutine occupations. This follows because S* and θ are complements—that is, the second derivative of wages with respect to S* and θ is positive.¹⁴ Thus an increase in θ will yield a relatively greater gain for workers with higher social skills, leading to a greater incentive (among both workers and firms) for high S workers to sort into nonroutine occupations. To see this, consider a simple extension of the model where there are two occupations (1 and 2) that differ in the variance of productivity, θ₁ > θ₂. All workers earn higher wages in higher θ occupations, and thus in the absence of labor market frictions all workers will sort into occupation 1.¹⁵ However, if jobs in 1 are limited, workers with higher social skills will obtain them first because they earn relatively higher wages in occupation 1 due to complementarity between S* and θ.¹⁶ The NLSY79 includes multiple observations of the same worker, which allows me to estimate changes in the returns to skill when workers switch occupations. I estimate models with worker fixed effects and interactions between skills and the task content of occupations.

• Workers earn more when they switch into nonroutine occupations, and their relative wage gain is increasing in social skill. This follows from the logic of prediction iii above. While the prediction for occupational sorting on social skills is clear, the impact of sorting on wages is less clear, for two reasons. First, in the absence of frictions, occupational sorting implies that wages will adjust until the marginal worker is indifferent between occupations. Second, the wage equations in (16) and (17) show a clear spillover of one worker’s skill to the other worker’s wages. Thus wage returns cannot be identified without information about labor market frictions and about the skills of the other workers. My solution is to study whether within-worker sorting into nonroutine occupations increases wages. While the magnitude of the coefficient will not have an economic interpretation because of the issues raised above, a positive sign is consistent with the predictions of the model.¹⁷ Moreover, because S and θ are complements, any wage gain from switching into a less routine occupation should be increasing in the worker’s social skills.

In addition, the model yields two predictions about changes in the return to social skills over time:

• Growth in the relative importance of jobs requiring social skills. The decline of routine employment is widely known (e.g., Autor, Levy, and Murnane 2003). However, I show that jobs requiring social skills have also experienced relative employment and wage growth in the United States over the past several decades. Indeed, these are largely the same types of jobs. I show that there is a strong negative correlation between measures of an occupation’s routineness and its social skill intensity. Thus the decline of routine employment can also be understood as growth in social skill-intensive employment. Importantly, this is not due to growth of higher-skill jobs more generally—in fact, employment and wage growth for high math, low social jobs (including many STEM occupations) has been relatively slow.

• Increasing returns to social skills over time. I explore this prediction by comparing the returns to social skills across the 1979 and 1997 waves of the NLSY. This compares youth entering the labor market in the 1980s and early 1990s to their counterparts in the early 2000s. I construct comparable age cohorts and include an identical set of covariates, which allows me to estimate changes in the returns to skills over time holding other factors constant. I also study whether the wage returns from sorting into social skill–intensive occupations have increased with time.

III. Data

III.A. O*NET and Census/ACS Data

I study changes in the the task content of work using data from O*NET. O*NET is a survey administered by the U.S. Department of Labor to a random sample of U.S. workers in each occupation. The O*NET survey began in 1998 and is updated periodically. I use the 1998 O*NET to most accurately reflect the task content of occupations in earlier years, although results with later versions of O*NET are generally similar.

The O*NET survey asks many different questions about the abilities, skills, knowledge, and work activities required in an occupation. The questions are rated on an ordinal scale, with specific examples that illustrate the value of each number to help workers answer the question accurately. Because the scale values have no natural cardinal meaning, I follow Autor, Levy, and Murnane (2003) and convert average scores by occupation on O*NET questions to a 0–10 scale that reflects their weighted percentile rank in the 1980 distribution of task inputs.

Autor and Dorn (2013) create a balanced and consistent panel of occupation codes that cover the 1980–2000 censuses and the 2005 American Community Survey (ACS). I extend their approach with the ACS data through 2012, updating the occupation crosswalk to reflect changes made in 2010 and making a few minor edits for consistency (see the Online Data Appendix for details).

I focus on changes in three key indicators of task content. First, I measure an occupation’s routine task intensity as the average of the following two questions: (i) “how automated is the job?” and (ii) “how important is repeating the same physical activities (e.g. key entry) or mental activities (e.g. checking entries in a ledger) over and over, without stopping, to performing this job?”¹⁸ Second, I closely follow Autor, Levy, and Murnane (2003) and define nonroutine analytical (math) task intensity as the average of three O*NET variables that capture an occupation’s mathematical reasoning requirements.¹⁹ Third, I define an occupation’s social skill intensity as the average of the four items in the O*NET module on “social skills”: (i) coordination, (ii) negotiation, (iii) persuasion, and (iv) social perceptiveness.²⁰

The measures of routineness and social skill intensity are strongly negatively correlated. Online Appendix Table A1 shows that the occupation-level correlation between routine task intensity and social skill task intensity is −0.68. This strong negative correlation drops only slightly (−0.56) after adding controls for 10 other widely used O*NET task measures. This suggests that a strong predictor of whether an occupation is routine is whether it requires social skills.

O*NET is the successor of the Dictionary of Occupational Titles (DOT), which was used by Autor, Levy, and Murnane (2003) and many others to study the changing task content of work. Online Appendix Figure A2 shows that the two data sources yield similar results for analogous task measures. I use the O*NET in this article because it is a more recent data source that is updated regularly and because it contains many more measures of the task content of work than the DOT.

III.B. NLSY79

My main data source for worker skills and wages is the NLSY79. The NLSY79 is a nationally representative sample of youth aged 14 to 22 in 1979. The survey was conducted yearly from 1979 to 1993 and then biannually from 1994 through 2012, and includes detailed measures of premarket skills, schooling experience, employment, and wages. My main outcome is the real log hourly wage (indexed to 2013 dollars), excluding respondents under the age of 23 or who are enrolled in school. Following Altonji, Bharadwaj, and Lange (2012), I trim values of the real hourly wage that are below 3 and above 200. The results are robust to alternative outcomes and sample restrictions such as using log annual earnings or conditioning on 20 or more weeks of full-time work.

I use respondents’ standardized scores on the Armed Forces Qualifying Test (AFQT) to proxy for cognitive skill, following many other studies (e.g., Neal and Johnson 1996; Altonji, Bharadwaj, and Lange 2012). Altonji, Bharadwaj, and Lange (2012) construct a mapping of the AFQT score across NLSY waves that is designed to account for differences in age-at-test, test format, and other idiosyncrasies. I take the raw scores from Altonji, Bharadwaj, and Lange (2012) and normalize them to have mean 0 and standard deviation 1.

Several psychometrically valid and field-tested measures of social skills exist, but none are used by the NLSY or other panel surveys of adult workers. As an alternative, I construct a premarket measure of social skills using the following four variables:

• Self-reported sociability in 1981 (extremely shy, somewhat shy, somewhat outgoing, extremely outgoing)

• Self-reported sociability in 1981 at age 6 (retrospective)

• The number of clubs in which the respondent participated in high school

• Participation in high school sports (yes/no)

I normalize each variable to have a mean of 0 and a standard deviation of 1. I then take the average across all four variables and restandardize so that cognitive skills and social skills have the same distribution. The results are not sensitive to other reasonable choices, such as dropping any one of the four measures or constructing a composite using principal component analysis.

The first three questions measure behavioral extroversion and prosocial orientation—both of which have been shown in meta-analyses to be positively correlated with measures of social and emotional intelligence (Lawrence et al. 2004; Declerck and Bogaert 2008; Mayer, Roberts, and Barsade 2008). Participation in team sports in high school has been associated with leadership, prosocial orientation, and teamwork ability and has been shown to positively predict labor market outcomes in adulthood (Barron, Ewing, and Waddell 2000; Kuhn and Weinberger 2005; Weinberger 2014). The measures of participation in sports and clubs used here are very similar to those used in Weinberger (2014).

A key concern is that this measure of social skills may simply be a proxy for unmeasured cognitive or “noncognitive” skills. The correlation between AFQT and social skills is about 0.26 in the analysis sample, which is consistent with the modest positive correlations (between 0.25 and 0.35) found between IQ and social and emotional intelligence across a variety of meta-analyses and independent studies (Mayer, Roberts, and Barsade 2008; Baker et al. 2014).

To account for possible bias from unmeasured ability differences, I control for completed years of education in addition to AFQT in some specifications. I also construct a measure of “noncognitive” skills using the normalized average of the Rotter Locus of Control and the Rosenberg Self-Esteem Scale—which are also used by Heckman, Stixrud, and Urzua (2006). This “noncognitive” skill measure is modestly positively correlated with both AFQT (0.30) and the social skills composite (0.20). To the extent that my measure of social skills is an imperfect or even poor proxy for the underlying construct, the results may understate its relative importance.

The NLSY79 includes information on each respondent’s occupation, which I match to the O*NET and DOT codes using the census occupation crosswalks developed by Autor and Dorn (2013). The NLSY also includes census industry codes, and I control for industry fixed effects in some specifications.

Mean self-reported sociability is 2.32 at age 6 and 2.88 as an adult, so on average respondents viewed themselves as less sociable in childhood than as adults. About 39% of respondents participated in athletics in high school, and the mean number of clubs was just above 1. Kuhn and Weinberger (2005) and Weinberger (2014) study the returns to leadership skills among a sample of white males who begin as high school seniors, leading to college-going rates that are about three times higher than in the NLSY79. Compared to those samples, the NLSY79 respondents are more disadvantaged and more representative of the U.S. population.

III.C. NLSY97

I investigate the growing importance of social skills by comparing the return to skills in the NLSY79 to the NLSY97. The NLSY97 is a nationally representative panel survey of youth age 12–16 in 1997 that follows a nearly identical structure to the NLSY79. My measure of social skills in the NLSY97 is two questions that capture the extroversion factor from the commonly used Big 5 personality inventory (e.g., Goldberg 1993). Following the procedures above, I normalize these two questions, take the average, and then renormalize them. The NLSY97 does not ask questions about clubs or participation in high school sports. Like the NLSY79, the NLSY97 also includes information on noncognitive skills (the Big 5 factor conscientiousness), as well as education, occupation, and industry.

When estimating changes in the return to skills over time in Section V, I modify the construction of the social skills measure from the NLSY79 so that it only uses the first two items on sociability. This maximizes the comparability of the two measures of social skills across NLSY waves. Finally, when comparing NLSY waves I restrict the sample to ages 25–33 to exploit the overlap in ages across surveys. This means I am comparing the returns to social skills for youth of similar ages during the late 1980s and early 1990s, compared to the more recent 2004–2012 period.

IV. NLSY79 Results

IV.A. Labor Market Returns to Skills and Complementarity

The first two predictions of the model are that there will be a positive return to skills in the labor market, and that cognitive skill and social skill are complements. I regress log hourly wages on both measures of skill and their interaction, controlling for a variety of other covariates:²¹

\begin{eqnarray} \ln \left(wage_{ijt}\right)&=&\alpha +\beta _{1}COG_{i}+\beta _{2}SS{}_{i}+\beta _{3}COG_{i}*SS_{i}+\gamma X_{ijt}\nonumber\\ &&+\,\delta _{j}+\zeta _{t}+\epsilon _{ijt}. \end{eqnarray}

(18)

The results are in Table I. The baseline model includes controls for race-by-gender indicators, indicators for region and urbanicity, and age (indexed by j) and year (indexed by t) fixed effects. Each observation is a person-year, and I cluster standard errors at the individual level.

Table I

Labor Market Returns to Cognitive Skills and Social Skills in the NLSY79

Outcome is log hourly wage (in 2012 dollars)	(1)	(2)	(3)	(4)	(5)	(6)	(7)
Cognitive skills (AQT, standardized)		0.206***	0.206***	0.189***	0.126***	0.190***	0.126***
		[0.007]	[0.007]	[0.007]	[0.008]	[0.007]	[0.008]
Social skills (standardized)	0.107***	0.055***	0.049***	0.043***	0.029***	0.044***	0.029***
	[0.006]	[0.006]	[0.006]	[0.006]	[0.006]	[0.006]	[0.006]
Cognitive * Social			0.019***	0.019***	0.011*	0.017***	0.010*
			[0.006]	[0.006]	[0.006]	[0.006]	[0.006]
Noncognitive skills (standardized)				0.048***	0.040***	0.046***	0.040***
				[0.006]	[0.006]	[0.006]	[0.006]
Cognitive * Noncognitive						0.008	0.001
						[0.006]	[0.006]
Demographics and age/year fixed effects		X	X	X	X	X	X
Years of completed education					X		X
R-squared	0.300	0.343	0.344	0.347	0.359	0.347	0.359
Observations	126,251	126,251	126,251	126,191	126,191	126,191	126,191

Notes. Each column reports results from an estimate of equation (18), with real log hourly wages as the outcome and person-year as the unit of observation. The data source is the National Longitudinal Survey of Youth 1979 cohort (NLSY79). Cognitive skills are measured by each NLSY79 respondent’s score on the Armed Forces Qualifying Test (AFQT), and are normalized to have a mean of 0 and a standard deviation of 1. I use the AFQT score crosswalk developed by Altonji, Bharadwaj and Lange (2012). Social skills is a standardized composite of four variables (i) sociability in childhood, (ii) sociability in adulthood, (iii) participation in high school clubs, and (iv) participation in team sports; see the text for details on construction of the social skills measure. My measure of noncognitive skills is the normalized average of the Rotter and Rosenberg scores in the NLSY. The regression also controls for race-by-gender indicator variables, age, year, census region, and urbanicity fixed effects, plus additional controls as indicated. Standard errors are in brackets and are clustered at the individual level. ***p < .01, **p < .05, *p < .10.

Column (1) shows that the return to social skills is positive and statistically significant. A one standard deviation increase in social skills increases real hourly wages by 10.7%. Column (2) adds the AFQT, my measure of cognitive skill. A one standard deviation increase in cognitive skill increase hourly wages by 20.6%. The addition of cognitive skill lowers the coefficient on social skills to 5.5% but it remains highly statistically significant.

Column (3) tests for complementarity by adding the interaction of cognitive skills and social skills, following the results in Weinberger (2014). The interaction is positive, large (0.019), and statistically significant at the less than 1% level. Column (4) adds controls for noncognitive skills. Noncognitive skills are highly predictive of wages (0.048, p < .001), but their inclusion barely changes the coefficients on cognitive skill and social skill, suggesting that each measure contains independent information about productivity. Column (5) adds controls for years of completed education. Controlling for education reduces the coefficient on all the skill measures, yet they remain statistically significant predictors of wages.

One concern is that cognitive skill and social skill are noisy measures of the same underlying ability. In that case, the estimated complementarity between cognitive skills and social skills reflects measurement error. I test this in column (6) by adding an interaction between cognitive skill and noncognitive skill. If wages are determined by a single ability that is measured by all three skills with error, all of the interaction terms will be positive. Yet column (6) shows that the interaction between cognitive skills and noncognitive skills is not statistically significant. Moreover, it drops to 0 after adding controls for education, even as the coefficient on the cognitive skill and social skill interaction remains statistically significant (column (7)). Complementarity holds only for cognitive skills and social skills.

Online Appendix Tables A3 and A4 show that the labor market return to social skills is positive and statistically significant for all race, gender, and education subgroups, in both logs and levels, respectively. I find some evidence of greater returns to skills and greater skill complementarity among respondents who have at least some college education, which is consistent again with Weinberger (2014).

IV.B. Occupational Sorting on Skills

I next examine the third prediction of the model—workers with higher levels of social skill will sort into nonroutine and social skill–intensive occupations. I estimate regressions like equation (18) above but with the task content of occupations (measured using O*NET) as the dependent variable. The baseline model is identical to equation (18), and I control for the covariates in Table I plus years of completed education and industry fixed effects.

The results are in Table II. Column (1) shows that a one standard deviation increase in social skills decreases the routine task intensity of a worker’s occupation by 1.88 percentiles, and the coefficient is highly statistically significant. I also find a negative coefficient on cognitive skills and the interaction between cognitive skills and social skills. Column (2) adds controls for math task intensity and three other related O*NET cognitive task measures. This causes the sign on cognitive skills to flip but has little impact on the other coefficients. Conditional on overall cognitive task intensity, workers in routine occupations have higher cognitive skills (0.161, p < .001) and significantly lower social skills (−0.149, p < .001). Combined with the negative coefficient on the interaction, these results imply that workers with high cognitive skills and low social skills sort into routine occupations.

Table II

Occupational Sorting on Skills in the NLSY79

	Routine		Social skills
Outcomes are O*NET task measures	(1)	(2)	(3)	(4)
Cognitive skills (AQT, standardized)	−0.055*	0.161***	0.345***	−0.044**
	[0.030]	[0.032]	[0.028]	[0.019]
Social skills (standardized)	−0.188***	−0.149***	0.208***	0.119***
	[0.022]	[0.024]	[0.020]	[0.014]
Cognitive * Social	−0.058***	−0.054**	0.014	0.013
	[0.021]	[0.023]	[0.019]	[0.014]
Demogs, age/year, education fixed effects	X	X	X	X
Controls for O*NET cognitive tasks		X		X
Observations	133,599	133,599	133,599	133,599
R-squared	0.204	0.237	0.305	0.668

Notes. Each column reports results from an estimate of equation (18), with the indicated 1998 O*NET task intensity of an occupation as the outcome and person-year as the unit of observation. The task measures are percentiles that range from 0 to 10 and are weighted by labor supply to conform to the 1980 occupation distribution. The additional O*NET cognitive task measures are nonroutine analytical, number facility, inductive/deductive reasoning, and analyze/use information. See the text and Online Appendix for details on the construction of each O*NET task measure. The data source is the National Longitudinal Survey of Youth 1979 cohort (NLSY79). Cognitive skills are measured by each NLSY79 respondent’s score on the Armed Forces Qualifying Test (AFQT), and are normalized to have a mean of 0 and a standard deviation of 1. I use the AFQT score crosswalk developed by Altonji, Bharadwaj and Lange (2012). Social skills is a standardized composite of four variables: (i) sociability in childhood, (ii) sociability in adulthood, (iii) participation in high school clubs, and (iv) participation in team sports (see the text for details on construction of the social skills measure). My measure of noncognitive skills is the normalized average of the Rotter and Rosenberg scores in the NLSY. The regression also controls for race-by-gender indicator variables, age, year, census region, and urbanicity fixed effects, plus additional controls as indicated. Standard errors are in brackets and are clustered at the individual level. ***p < .01, **p < .05, *p < .10.

Columns (3) and (4) estimate parallel specifications but with the social skill intensity of a worker’s occupation as the outcome. The results are generally similar but opposite in sign. Overall, the results in Table II confirm the prediction that workers with higher social skills sort into nonroutine and social skill–intensive occupations. This suggests that estimates of the return to skills within occupations should be interpreted with caution.

IV.C. Returns to Skills by Occupation Task Intensity

Table II shows clearly that workers sort into occupations where their skills are more rewarded. This makes it difficult to estimate the returns to worker skills controlling for occupation. However, if we are willing to assume that labor market frictions prevent perfect sorting of workers to occupations, we can estimate how the return to skills changes when the same worker switches occupations. Labor market frictions may be particularly important early in one’s career, when skills are imperfectly observed by employers (e.g., Altonji and Pierret 2001).

The model predicts that workers will earn more when they switch into nonroutine and social skill–intensive occupations and that the wage gain from switching will be increasing in social skill. I explore these predictions by estimating:

\begin{eqnarray} \ln \left(wage_{ijt}\right) & = & \beta _{1}COG_{i}*T_{ijt}+\beta _{2}SS_{i}*T_{ijt}+\beta _{3}COG_{i}*SS_{i}*T_{ijt}\nonumber \\ & & + \,\gamma X_{ijt}+\eta _{i}+\delta _{j}+\zeta _{t}+\epsilon _{ijt}. \end{eqnarray}

(19)

where T_ijt indexes the task content of a worker’s occupation (with the main effect included in the X_ijt vector), η_i is a worker fixed effect and the rest of the terms are defined as above. Note that with worker fixed effects only the interactions between skills and T_ijt are identified, not the returns to skills themselves.

The results are in Table III. The baseline specification in column (1) shows that workers earn significantly higher wages when they sort into routine occupations. However, I do find that the wage return from sorting into nonroutine occupations is increasing in social skills, which is consistent with the predictions of the model. Column (2) replaces routine with social skill task intensity. Workers who switch into a job that is 10 percentiles higher in the O*NET measure of social skill intensity earn about 3.9% higher wages. Moreover, the worker’s wage gain is significantly increasing in her social skills. For example, the estimates imply a wage gain of 3.9% for a worker of average social skills but 8.9% when the worker has social skills that are one standard deviation above the mean.

Table III

Returns to Skills by Occupation Task Intensity in the NLSY79

Outcome is log hourly wage (in 2012 dollars)	(1)	(2)	(3)
Routine task intensity	0.0136***		0.0212***
	[0.0012]		[0.0014]
Cognitive * Routine task intensity	−0.0034***		0.0005
	[0.0013]		[0.0015]
Social skills * Routine task intensity	−0.0025**		−0.0008
	[0.0013]		[0.0015]
Cognitive * Social * Routine task intensity	−0.0008		−0.0011
	[0.0012]		[0.0014]
Social skill task intensity		0.0039***	0.0176***
		[0.0013]	[0.0016]
Cognitive * Social skill task intensity		0.0113***	0.0112***
		[0.0015]	[0.0018]
Social skills * Social skill task intensity		0.0050***	0.0041**
		[0.0015]	[0.0018]
Cognitive * Social * Social skill task intensity		0.0021	0.0011
		[0.0015]	[0.0023]
Worker fixed effects	X	X	X
Observations	126,251	126,251	126,251
Number of individuals	11,050	11,050	11,050

Notes. Each column reports results from an estimate of equation (19), with real log hourly wages as the outcome and person-year as the unit of observation. The data source is the National Longitudinal Survey of Youth 1979 cohort (NLSY79). Cognitive skills are measured by each NLSY79 respondent’s score on the Armed Forces Qualifying Test (AFQT), and are normalized to have a mean of 0 and a standard deviation of 1. I use the AFQT score crosswalk developed by Altonji, Bharadwaj and Lange (2012). Social skills is a standardized composite of four variables: (i) sociability in childhood, (ii) sociability in adulthood, (iii) participation in high school clubs, and (iv) participation in team sports (see the text for details on construction of the social skills measure). My measure of noncognitive skills is the normalized average of the Rotter and Rosenberg scores in the NLSY. All models control for worker fixed effects, age, year, census region, and urbanicity fixed effects, plus additional controls as indicated. The interactions between cognitive/social skills and 1998 O*NET task intensities measure whether the returns to skills vary with the task content of the worker’s occupation. The task measures are percentiles that range from 0 to 10 and are weighted by labor supply to conform to the 1980 occupation distribution. See the text and Online Appendix for details on the construction of each O*NET task measure. Standard errors are in brackets and are clustered at the individual level. ***p < .01, **p < .05, *p < .10.

Column (3) includes both the routine and social skill measures together. This causes the interactions between skills and routine task intensity to fade to near 0, while the coefficients on the social skill interactions remain statistically significant and even increase slightly. Thus the social skill O*NET task measure is a better predictor of the returns to social skills when both measures are included together. The results in Table III are robust to including industry fixed effects as well as other specific job attributes such as union status or whether a position involves supervising workers. In addition, in results not reported I find that interactions between skills and math task intensity are not statistically significant. This shows that relatively higher returns to skill in social skill–intensive occupations are not simply a proxy for job complexity or overall skill requirements.

While Krueger and Schkade (2008) do not estimate within-worker wage changes, their compensating differentials explanation implies that workers are willing to accept a wage penalty for a job with more social interaction. However, the wage gains from switching into a social skill–intensive occupation shown in Table III are not consistent with a compensating differentials story. Instead, the results support the predictions of the model, which suggest that higher social skills are more beneficial in occupations where there is more potential gain from task trade.

V. The Growing Importance of Social Skills

V.A. Employment and Wage Growth in Social Skill–Intensive Occupations

I begin by presenting trends in employment and wage growth in the United States between 1980 and 2012. Figure III replicates Figure I of Autor, Levy, and Murnane (2003) for the 1980–2012 period using the three O*NET task measures described above. By construction, each task variable has a mean of 50 centiles in 1980. Thus subsequent movement should be interpreted as changes in the employment-weighted mean of each task relative to its importance in 1980. The data are aggregated to the industry-education-sex level, which implicitly controls for changes in task inputs that are due to changes in the industry and skill mix of the U.S. economy over time. There is no adding-up constraint for tasks in a given year, and so changes over time can also reflect changes in total labor supply.

Figure III

Worker Tasks in the U.S. Economy, 1980–2012 (Update of Autor, Levy, and Murnane (2003) Figure I)

Figure III is constructed to parallel Figure I of Autor, Levy and Murnane (2003). O*NET 1998 task measures by occupation are paired with data from the IPUMS 1980–2000 censuses and the 2005–2013 American Community Survey samples. Consistent occupation codes for 1980–2012 are from Autor and Dorn (2013) and Autor and Price (2013). Data are aggregated to industry-education-sex cells by year, and each cell is assigned a value corresponding to its rank in the 1980 distribution of task input. Plotted values depict the employment-weighted mean of each assigned percentile in the indicated year. See the text and Online Appendix for details on the construction of O*NET task measures.

Like Autor and Price (2013), I find that the labor input of routine tasks has continued to decline, and that nonroutine analytical (math) task inputs stopped growing and even declined modestly after 2000. However, social skill task inputs grew by 24% from 1980 to 2012, compared to only about 11% for nonroutine analytical tasks. Moreover, while nonroutine analytical task inputs have declined since 2000, social skills task inputs held steady (growing by about 2%) through the 2000s. Not surprisingly, the decline in routine tasks mirrors the growing importance of social skills between 1980 and 2012.

Since the math and social skill task measures are highly correlated, growth in the importance of social skills could simply reflect general skill upgrading. I address this by dividing occupations into four mutually exclusive categories based on whether they are above or below the median percentile in both math and social skill task intensity. I then compute the share of all labor supply-weighted employment in each category and year.

Figure IV plots the growth of employment shares—relative to a 1980 baseline—in each category. Jobs with high math and high social skill intensity grew by about 7.2 percentage points as a share of the U.S. labor force between 1980 and 2012. Low math, high social skill jobs grew by about 4.6 percentage points, for a total increase of 11.8 percentage points in the employment share of social skill–intensive occupations since 1980. In contrast, the employment share of jobs with high math but low social skill intensity shrank by about 3.3 percentage points over the same period. This includes many of the STEM jobs shown in Figure I. The basic pattern in Figure IV is robust to choosing cutoffs other than the 50th percentile for each type of task.

Figure IV

Cumulative Changes in Employment Share by Occupation Task Intensity, 1980–2012

Each line plots 100 times the change in employment share (relative to a 1980 baseline) between 1990 and 2012 for occupations that are above and/or below the 50th percentile in nonroutine analytical and social skill task intensity as measured by the 1998 O*NET. Consistent occupation codes for 1980–2012 are updated from Autor and Dorn (2013) and Autor and Price (2013). See the text and Online Appendix for details on the construction of O*NET task measures and for examples of occupations in each of the four categories. Source: 1980–2000 census, 2005–2013 ACS.

One possible explanation for the slow growth of high math, low social skill jobs is that employers cannot find workers to fill technical and math-intensive positions. In that case, we would expect relatively greater wage growth for these occupations. Figure V plots the change since 1980 in real hourly wages for occupations in each of the four categories. I find that wages for high math, low social skill jobs grew by only about 5.9% between 1980 and 2012, compared to about 26% for high math, high social skill occupations.

Figure V

Cumulative Changes in Real Hourly Wages by Occupation Task Intensity, 1980–2012

Each line plots the percent change in median hourly wages (relative to a 1980 baseline and in constant 2012 dollars) between 1990 and 2012 for occupations that are above and/or below the 50th percentile in nonroutine analytical and social skill task intensity as measured by the 1998 O*NET. Consistent occupation codes for 1980 to 2012 are updated from Autor and Dorn (2013) and Autor and Price (2013). See the text and Online Appendix for details on the construction of O*NET task measures and for examples of occupations in each of the categories. Source: 1980–2000 census, 2005–2013 ACS.

Online Appendix Figures A3 and A4 show that employment and wage growth for social skill–intensive occupations has occurred throughout the skill distribution and is not concentrated in particularly low- or high-paying jobs.

Online Appendix Tables A5 and A6 estimate employment and wage growth for jobs requiring different bundles of tasks in a multivariate framework. The results generally support the growing importance of social skills after controlling for changes in sex, education, and industry mix. I find particularly strong employment growth for jobs that are high in both math and social skills. This pattern has accelerated since 2000. Finally, I note that the strong growth of social skill–intensive jobs is robust to excluding all managerial, health care, and education occupations from the sample, although these jobs are important drivers of the overall trend.

Overall, the evidence from aggregate labor market data suggests that jobs requiring social skills have experienced strong relative employment and wage growth since 1980.

V.B. Increasing Returns to Social Skill across NLSY Waves

Here I present direct evidence on the growing importance of social skills by studying changes in the returns to skills across the 1979 and 1997 waves of the NLSY. The cognitive skill and social skill measures are designed to be closely comparable across waves. As a reminder, I restrict the age range to 25–33 and use an alternative definition of social skills for this analysis to maximize comparability across waves (see Section III for details). I estimate:

\begin{eqnarray} y_{ijt}&=&\alpha +\sum _{s=1}^{^{S}}\left[\beta _{s}SKILL_{i}+\gamma _{s}\left(SKILL_{i}*NLSY97_{i}\right)\right]\nonumber\\ &&+\,\zeta X_{ijt}+\delta _{j}+\zeta _{t}+\epsilon _{ijt}. \end{eqnarray}

(20)

The skill vector includes cognitive skills, social skills and their interaction, and noncognitive skills in some specifications. The interaction between skills and an indicator for being in the NLSY97 sample allows me to directly test the hypothesis that the returns to skills have changed over time. The X_ijt vector includes a standard set of demographic controls, age and year fixed effects, and an indicator variable for whether the respondent is in the NLSY97 sample. To study changing selection into the labor force, I allow y_ijt to be either an indicator for full-time employment or the log real hourly wage (conditional on employment).

The results are in Table IV. Columns (1)–(3) show results for full-time employment. Column (1) shows that a one standard deviation increase in cognitive skills increases the probability of full-time employment by 6.8 percentage points, relative to a baseline mean of about 85%. However, the interaction with the NLSY97 sample indicator is not statistically significant, suggesting that the returns to cognitive skill in terms of full-time work have not changed very much across survey waves.

Table IV.

Labor Market Returns to Skills in the NLSY79 versus NLSY97

	Full-time employment			Log real hourly wage
	(1)	(2)	(3)	(4)	(5)	(6)
Cognitive skills (AQT, standardized)	0.068***	0.042***	0.040***	0.203***	0.129***	0.116***
	[0.003]	[0.003]	[0.003]	[0.005]	[0.006]	[0.006]
Cognitive skills * NLSY97	0.008*	0.005	0.009*	−0.052***	−0.061***	−0.045***
	[0.004]	[0.004]	[0.005]	[0.008]	[0.008]	[0.008]
Social skills (standardized)	0.007***	0.005**	0.004*	0.020***	0.015***	0.011**
	[0.002]	[0.002]	[0.002]	[0.004]	[0.004]	[0.004]
Social skills * NLSY97	0.023***	0.021***	0.019***	0.017**	0.019**	0.016**
	[0.004]	[0.004]	[0.004]	[0.008]	[0.008]	[0.008]
Cognitive * Social	−0.007***	−0.006**	−0.007**	0.006	0.003	0.002
	[0.003]	[0.003]	[0.003]	[0.004]	[0.004]	[0.004]
Cognitive * Social * NLSY97	−0.006	−0.006	−0.006	−0.004	−0.004	−0.002
	[0.004]	[0.004]	[0.004]	[0.008]	[0.008]	[0.008]
Noncognitive skills (standardized)			0.008**			0.041***
			[0.003]			[0.005]
Noncognitive skills * NLSY97			0.013***			0.016*
			[0.005]			[0.009]
Demographics and age/year FE	X	X	X	X	X	X
Years of completed education		X	X		X	X
R-squared	0.081	0.096	0.097	0.309	0.333	0.337
Observations	104,613	104,252	104,206	77,845	77,631	77,599

Notes. Each column reports results from an estimate of equation (20), with an indicator for being employed full-time as the outcome in columns (1)–(3), real log hourly wages as the outcome in columns (4)–(6), and person-year as the unit of observation. The data are a pooled sample of two cohorts of youth: the National Longitudinal Survey of Youth 1979 (NLSY79) and 1997 (NLSY97) waves. I restrict the age range to 25–33, which allows for a comparison of NLSY respondents at similar ages across survey waves. Cognitive skills are measured by each NLSY respondent’s score on the Armed Forces Qualifying Test (AFQT), and are normalized to have a mean of 0 and a standard deviation of 1. I use the AFQT score crosswalk developed by Altonji, Bharadwaj, and Lange (2012), which adjusts for differences across survey waves in age-at-test and test format. Social skills is a standardized composite of two variables that measure extroversion in both the NLSY79 (sociability in childhood and sociability in adulthood) and in the NLSY97 (two items from the Big 5 personality inventory that measure extroversion). The noncognitive skill measures are a normalized average of the Rotter and Rosenberg scores in the NLSY79, and two items from the NLSY97 that measure the Big 5 personality factor conscientiousness. The regression also controls for an indicator for whether the respondent was in the NLSY97 wave, race-by-gender indicator variables, age, year, census region, and urbanicity fixed effects, plus additional controls as indicated. Standard errors are in brackets and are clustered at the individual level. ***p < .01, **p < .05, *p < .10.

In contrast, the association between social skills and the probability of full-time work has increased more than fourfold. A one standard deviation increase in social skills is associated with an increase in the probability of full-time employment of only about 0.7 percentage points (p = .006) in the NLSY79 sample, compared to 3.0 percentage points in the NLSY97 sample (p < .001).

The NLSY97 sample was in the 25–33 age range between 2004 and 2012, which matches up closely to the labor market trends shown in Section V.A. In results not shown, I find that the difference in returns to skills across NLSY waves is slightly larger for men, which suggests that differences in women’s labor force participation across the past few decades are not directly driving the results.

Column (2) adds controls for years of completed education, which reduces the impact of skills overall but has almost no impact on the change in returns to skills over time. Column (3) adds controls for noncognitive skills. The impact of a one standard deviation gain in noncognitive skills on the probability of full-time work has increased from 0.8 to 2.1 percentage points. However, the coefficients on social skills are qualitatively unchanged.

Columns (4)–(6) study changes in the impact of skills on wages, among workers who are employed full-time. The large change in the impact of skills on full-time work in columns (1)–(3) suggests that these results should be interpreted with caution, although under reasonable assumptions about labor market sorting they provide a lower bound estimate of the changing return to skills.

Interestingly, the wage return to cognitive skills has declined modestly over time. The estimates in column (4) imply that a one standard deviation increase in cognitive skills increased wages by 20.3% in the NLSY79 but only 15.1% in the NLSY97.²² This is consistent with Castex and Dechter (2014), who also study the changing returns to cognitive skill using the NLSY79 and NLSY97.

In contrast, the return to social skill among full-time workers has grown significantly across NLSY waves. The estimates in column (4) imply that a one standard deviation increase in social skills yields a wage gain of 2.0% in the NLSY79, compared to 3.7% in the NLSY97. Adding controls for years of completed education and noncognitive skills has little impact on the estimates. Overall, the results in Table IV show that social skills are a significantly more important predictor of labor market success for youth in the 2004 to 2012 period, compared to the late 1980s and 1990s.

V.C. Changes in the Relative Returns to Skill Across Occupations

Finally, I study (i) whether the wage gain from sorting into social skill–intensive occupations has changed across survey waves, and (ii) whether this wage gain (if any) is increasing in a worker’s social skills. I estimate:

\begin{eqnarray} \ln \left(wage_{ijt}\right) & = & \sum _{s=1}^{^{S}}\left[\beta _{s}\left(SKILL_{i}*T_{ijt}\right)+\vartheta _{s}\left(T_{ijt}*NLSY97_{i}\right)\right.\nonumber \\ & &\left.+\,\gamma _{s}\left(SKILL_{i}*T_{ijt}*NLSY97_{i}\right)\right]\nonumber\\ &&+\,\zeta X_{ijt}+\eta _{i}+\delta _{j}+\phi _{t}+\epsilon _{ijt}. \end{eqnarray}

(21)

Equation (21) takes the same general form as equation (19), with worker fixed effects and interactions between skills and occupation task intensities from O*NET. The key difference is that I also include three-way interactions between skills, task measures, and an indicator for being in the NLSY97 panel.

The results are in Table V. Columns (1) and (2) include only the two-way interactions between the task measures T_ijt and the NLSY97 indicator. In column (1), I find that the wage gain for a worker who switches into a more social skill–intensive occupation is significantly greater in more recent years. The within-worker wage return to a 10 percentile increase in skill intensity is equal to 0 in the NLSY79 wave, compared to about 2.1% in the NLSY97 wave.²³ Column (2) adds the math task measure plus an interaction with the NLSY97 indicator. In contrast to the results for social skills, the wage return to math-intensive occupations declined from about 1.7% to 0.8% between the 1979 and 1997 NLSY cohorts. Thus the evidence in Table V suggests that the wage gain from sorting into social skill–intensive jobs has increased over time.

Table V

Returns to Skills by Occupation Task Intensity in the NLSY79 versus NLSY97

Outcome is log hourly wage (in 2012 dollars)	(1)	(2)	(3)	(4)
Social skill task intensity	0.0004	−0.0096***	−0.0095***	−0.0096***
	[0.0013]	[0.0015]	[0.0015]	[0.0015]
Social skill task intensity * NLSY97	0.0210***	0.0253***	0.0217***	0.0225***
	[0.0036]	[0.0041]	[0.0040]	[0.0040]
Math task intensity		0.0175***	0.0177***	0.0177***
		[0.0015]	[0.0015]	[0.0015]
Math task intensity * NLSY97		−0.0082**	−0.0085**	−0.0099***
		[0.0035]	[0.0034]	[0.0034]
Cognitive skill * Social skill task intensity			0.0069***	0.0074***
			[0.0013]	[0.0016]
Cognitive skill * Social skill task intensity * NLSY97			0.0114***	0.0047
			[0.0036]	[0.0044]
Social skill * Social skill task intensity			0.00008	0.0011
			[0.0013]	[0.0016]
Social skill * Social skill task intensity * NLSY97			0.0040	0.0069*
			[0.0032]	[0.0038]
P (Social skill * Social skill intensity in NLSY97 > 0)			0.108	0.023
P (All skills * Social skill intensity in NLSY97 > 0)			0.000	0.001
P (All skills in NLSY97 > All skills in NLSY79)				0.000
Observations	77,845	77,845	77,845	77,845
Number of individuals	14,998	14,998	14,998	14,998

Notes. Each column reports results from an estimate of equation (21), with real log hourly wages as the outcome and person-year as the unit of observation. The data are a pooled sample of two cohorts of youth: the National Longitudinal Survey of Youth 1979 (NLSY79) and 1997 (NLSY97) waves. I restrict the age range to 25–33, which allows for a comparison of NLSY respondents at similar ages across survey waves. Cognitive skills are measured by each NLSY respondent’s score on the Armed Forces Qualifying Test (AFQT), and are normalized to have a mean of 0 and a standard deviation of 1. I use the AFQT score crosswalk developed by Altonji, Bharadwaj and Lange (2012), which adjusts for differences across survey waves in age-at-test and test format. Social skills is a standardized composite of two variables that measure extroversion in both the NLSY79 (sociability in childhood and sociability in adulthood) and in the NLSY97 (two items from the Big 5 personality inventory that measure extroversion). The regression also controls for age, year, census region, and urbanicity fixed effects, plus additional controls as indicated. The interactions between cognitive/social skills and 1998 O*NET task intensities measure whether the returns to skills vary with the task content of the worker’s occupation. The task measures are percentiles that range from 0 to 10 and are weighted by labor supply to conform to the 1980 occupation distribution. See the text and Online Appendix for details on the construction of each O*NET task measure. Standard errors are in brackets and are clustered at the individual level. ***p < .01, **p < .05, *p < .10.

Columns (3) and 4 add the three-way interactions with skills shown in equation (22). I add summary tests of statistical significance across multiple coefficients on skills at the bottom of Table V. The complementarity between social skills and jobs requiring social interaction has increased in the NLSY97 sample. The coefficient on the triple interaction in column (3) is positive but not statistically significant, and the sum of the coefficients barely fails to reject at the 10% level (p = .108). Column (4) adds interactions between skills, the NLSY97 indicator, and math task intensity. To conserve space, I do not show these coefficients. However, adding math task intensity makes the triple interaction between social skills, social skill task intensity, and the NLSY97 indicator larger and more precise, and it is now statistically significant at the 10% level (0.0069, p = .071). In contrast, the triple interaction with math task intensity (not shown) is negative and not statistically different from 0.

In sum, comparing the returns to skills and the impact of job changes across survey waves suggests that social skills have become more important over time and that growth in the return to social skills has been greater for workers who sort into social skill–intensive occupations.

VI. Conclusion

This article presents evidence of growing demand for social skills over the past several decades. What explains the growing importance of social skills in the labor market? One reason is that computers are still very poor at simulating human interaction. Reading the signals of others and reacting is an unconscious process, and skill in social settings has evolved in humans over thousands of years. Human interaction in the workplace involves team production, with workers playing off of each other’s strengths and adapting flexibly to changing circumstances. Such nonroutine interaction is at the heart of the human advantage over machines.

I formalize the importance of social skills with a model of team production in the workplace. Because workers naturally vary in their ability to perform the great variety of workplace tasks, teamwork increases productivity through comparative advantage. I model social skills as reducing the worker-specific cost of coordination or trading tasks with others. Workers with high social skills can trade tasks at a lower cost, enabling them to work with others more efficiently.

The model generates intuitive predictions about sorting and the relative returns to skills across occupations, which I investigate using two panel surveys, the NLSY79 and NLSY97, that contain comparable measures of worker skills and repeated observations of occupational choice and wages. I find that the wage return to social skills is positive even after conditioning on cognitive skill, noncognitive skill, and a wide variety of other covariates, and that cognitive skill and social skill are complements. I also find that workers with higher social skills are more likely to work in social skill–intensive occupations, and that they earn a relatively higher wage return when they sort into these occupations.

I show evidence of strong relative employment and wage growth for social skill–intensive occupations between 1980 and 2012. Jobs that require high levels of cognitive skill and social skill have fared particularly well, while high math, low social skill jobs (including many STEM occupations) have fared especially poorly. I also study changes in the returns to social skill between the NLSY79 and NLSY97, using nearly identical measures of skills and other covariates across survey waves. I find that social skills were a much stronger predictor of employment and wages for young adults age 25 to 33 in the mid-2000s, compared to the 1980s and 1990s. In contrast, the importance of cognitive skills has declined modestly.

This article argues for the importance of social skills, yet it is silent about where social skills come from and whether they can be affected by education or public policy. A robust finding in the literature on early childhood interventions is that long-run impacts on adult outcomes can persist can even when short-run impacts on test scores “fade out” (e.g., Deming 2009; Chetty et al. 2011).

It is possible that increases in social skills are a key mechanism for long-run impacts of early childhood interventions. Heckman, Pinto, and Savelyev (2013) find that the long-run impacts of the Perry Preschool project on employment, earnings and criminal activity were mediated primarily by program-induced increases in social skills. The Perry Preschool curriculum placed special emphasis on developing children’s skills in cooperation, resolution of interpersonal conflicts, and self-control. Recent longitudinal studies have found strong correlations between a measure of socioemotional skills in kindergarten and important young adult outcomes such as employment, earnings, health, and criminal activity (Dodge et al. 2014; Jones, Greenberg, and Crowley 2015).

If social skills are learned early in life, not expressed in academic outcomes such as reading and math achievement, but important for adult outcomes such as employment and earnings, this would generate the fade-out pattern that is commonly observed for early life interventions. Indeed, preschool classrooms focus much more on the development of social and emotional skills than elementary school classrooms, which emphasize “hard” academic skills such as literacy and mathematics. Still, these conclusions are clearly speculative, and the impact of social skill development on adult labor market outcomes is an important question for future work.

Supplementary Material

An Online Appendix for this article can be found at The Quarterly Journal of Economics online. Data and code replicating the tables and figures in this paper can be found in Deming (2017), in the Harvard Dataverse, doi:10.7910/DVN/CYPKZH.

*

Thanks to Pol Antràs, David Autor, Avi Feller, Lawrence Katz, Sandy Jencks, Richard Murnane, and Lowell Taylor for reading early drafts of this article and providing insightful feedback. Thanks to Felipe Barrera-Osorio, Amitabh Chandra, Asim Khwaja, Alan Manning, Guy Michaels, Luke Miratrix, Karthik Muralidharan, Devah Pager, Todd Rogers, Doug Staiger, Catherine Weinberger, Marty West, and seminar participants at PSE, LSE, CESifo, Yale, Columbia, Harvard, MIT, Michigan State, Northwestern, UBC, Simon Fraser, Cornell, University of Chicago, and the NBER Education and Personnel meetings for helpful comments. Special thanks to David Autor and Brendan Price for sharing their data and programs and to Madeleine Gelblum for excellent research assistance throughout the writing of this article. Olivia Chi, Lauren Reisig, and Stephen Yen also provided superb research assistance. Extra special thanks to Lisa Kahn and Chris Walters for “trading tasks” with me. All errors are my own.

1.

Following Beaudry, Green, and Sand (2016), Figure I displays employment growth for what the U.S. Census refers to as managerial, professional, and technical occupation categories. Autor and Dorn (2013) create a consistent set of occupation codes for the 1980–2000 censuses and the 2005–2008 ACS—I follow their scheme and update it through the 2010 census and the 2011–2013 ACS (see the Online Data Appendix for details). Following Beaudry, Green, and Sand (2016), “cognitive” occupations include all occupation codes in the Data Appendix between 1 and 235. I group occupation codes into larger categories in some cases for ease of presentation (e.g., engineers, managers).

2.

Acemoglu and Autor (2011) develop a Ricardian model of the labor market with three skill groups, a single skill index, and comparative advantage for higher-skilled workers in relatively more complex tasks. I follow a long literature that treats teamwork as a trade-off between the benefits of increased productivity through specialization and the costs of coordination (Becker and Murphy 1992; Bolton and Dewatripont 1994; Lazear 1999; Garicano 2000; Garicano and Rossi-Hansberg 2006).

3.

A related literature studies job assignment when workers have multiple skills (Heckman and Scheinkman 1987; Yamaguchi 2012; Lindenlaub 2014; Lise and Postel-Vinay 2014). Models of this type would treat social skill as another addition to the skill vector, with Roy-type selection and linear (or log-linear) wage returns rather than the specific pattern of complementarity between cognitive skill and social skill.

4.

Krueger and Schkade (2008) show that gregarious workers sort into jobs that involve more social interaction. They interpret this as a compensating differential, suggesting that workers have preferences for interactive work. However, if skill in social interaction had no value in the labor market but interactive jobs were preferred by workers, compensating differentials imply that interactive jobs should pay less all else equal.

5.

Autor and Dorn (2013) explain the rise of low-wage service occupations as computers replacing routine production tasks rather than service tasks (which are more difficult to automate). However, this does not explain the growth of social skill–intensive jobs at the top of the wage distribution. Autor, Dorn, and Hanson (2015) compare the impact of import competition from China to technological change and find that the impact of trade is concentrated in manufacturing and is larger among less-skilled workers. Oldenski (2012) shows that production requiring complex within-firm communication is more likely to occur in a multinational’s home country. Karabarbounis and Neiman (2014) show that the share of corporate value-added paid to labor has declined, even in labor-intensive countries such as China and India, suggesting that offshoring alone is unlikely to explain the decline of routine employment and the growth in social skill–intensive jobs.

6.

In McCann et al. (2015), workers who specialize in communication become managers of a team, and the communication skills of the other workers on the team are irrelevant. Models with communication or “people” tasks face the challenge of specifying exactly what is being produced. Are workers who spend an entire day in meetings communication task specialists? The model here treats communication as a friction. Workers who spend more time in meetings—conditional on total output—have lower social skill.

7.

Woolley et al. (2010) randomly assign individuals to groups and then ask the groups to perform a variety of tasks. Group performance is positively correlated with the “average social sensitivity” of group members as measured by a test called “Reading the Mind in the Eyes.” This test was originally developed to assist in the diagnosis of autism and Asperger syndrome, but has since been demonstrated as psychometrically valid and able to detect subtle differences in individual social sensitivity (e.g., Baron-Cohen et al. 2001).

8.

In Becker and Murphy (1992), the benefits of specialization are balanced against the costs of coordinating increasingly specialized workers. In their analysis, coordination costs are features of the economy or of particular sectors. Here I treat coordination costs as attributes of individual workers. The definition of social skills in this article is closely related to the formulation of “iceberg” trade costs between countries as in Dornbusch, Fischer, and Samuelson (1977) and Eaton and Kortum (2002). The main difference is that iceberg trade costs are defined at the country-pair level (i.e., S_ni) and do not necessarily have a common worker (country) component. This is a particular definition of social skill, and it does not rule out other ways that sociability might affect productivity and wages (i.e., taste discrimination by firms, differential rates of on-the-job learning or information acquisition). One convenient interpretation of S is that it represents the probability that a worker will correctly communicate her productivity schedule to another worker.

9.

An earlier draft of this article developed a Ricardian model with multiple workers which closely followed Eaton and Kortum (2002). Adding multiple workers yields identical predictions and has a very similar structure, but requires a strong distributional assumption and comes with much added complexity.

10.

Specifically, imagine that worker j’s productivity in task t is a random variable with a log-normal distribution: |$a_{j}(t)\sim \mathrm{ln}\mathcal {N}(\mu _{j},\sigma ^{2})$|⁠. Then the ratio of worker 1 to worker 2’s productivity in task t, |$G(t)\equiv \frac{a_{1}(t)}{a_{2}(t)}$|⁠, also takes on a log-normal distribution: |$G(t)\sim \mathrm{ln}\mathcal {N}(\mu _{G},\sigma _{G}^{2})$|⁠, with μ_G = μ₁ − μ₂ and |$\sigma _{G}^{2}=2\sigma ^{2}$|⁠. It can be shown that the quantile function for G(t) evaluated at (1 − i) corresponds closely to the chosen functional form for γ(i), with |$\bar{A}\approx \mathrm{exp}(\mu _{1}-\mu _{2})$| and θ ≈ 2σ. See the Online Model Appendix for details.

11.

The marginal task is equal to |$i^{*}=\frac{A_{1}}{A_{1}+A_{2}\mathrm{exp}(\theta (2i^{*}-1))}$|⁠.

12.

The gains from trade can be expressed as |$\Delta Y=\frac{Y^{T}}{Y^{A}}$|⁠, the ratio of worker output under trade to worker output under autarky. This is equal to |$\mathrm{exp}\bigg (\int _{i^{*}}^{1}\mathrm{ln}\bigg [\frac{\gamma (i^{*})}{\gamma (i)}\bigg ]di\bigg )=\mathrm{exp}(\theta [i^{*}-1]^{2})$| for worker 1 and |$\mathrm{exp}\bigg (\int _{0}^{i^{*}}\mathrm{ln}\bigg [\frac{\gamma (i)}{\gamma (i^{*})}\bigg ]di\bigg )=\mathrm{exp}(\theta i^{*2})$| for worker 2.

13.

For simplicity, assume workers have equal cognitive skill, that is A₁ = A₂ = A and thus |$\bar{A}=\omega =1$|⁠. Then worker 1’s production is |$Y_{1}^{S}=A(S^{*})^{1-i^{H}}\mathrm{exp}(\int _{0}^{i^{H}}\mathrm{ln}[\alpha _{1}(i)]di+\int _{i^{H}}^{1}\mathrm{ln}[\alpha _{2}(i)]di)$|⁠. The second derivative with respect to A and S* is |$\frac{d^{2}Y_{1}^{S}}{dAdS^{*}}=(1-i^{H})(S^{*})^{-i^{H}}\mathrm{exp}(\int _{0}^{i^{H}}\mathrm{ln}[\alpha _{1}(i)]di+\int _{i^{H}}^{1}\mathrm{ln}[\alpha _{2}(i)]di)$|⁠, which is always positive. Note that the special case of equal ability matches the empirical work in Section IV, which explicitly conditions on cognitive skill. See the Online Model Appendix for a proof.

14.

The second derivative of wages with respect to θ and S* is positive because S* and θ are complements in the gains from trade. See the Online Model Appendix for a formal proof of this proposition. Since production under task trade is equal to production under autarky times the gains from trade, and wages are equal to output times the exogenous output price P*, S* and θ are complements in output (and thus in wages) when S* and θ are complements in the gains from trade.

15.

Section 3.7 of the Online Model Appendix (equations (78) and (84)) shows that |$\frac{dY}{d\theta }>0$| for all workers when |$\bar{A}=1$|⁠.

16.

Section 3.7 of the Online Model Appendix provides a formal proof of this proposition by assuming that there are two occupations characterized by different values of θ (θ₁ > θ₂). The setup of the model is the same as above, except that each firm hires two workers into a single occupation—that is, firms are either type 1 or type 2. Workers maximize wages and can switch occupations. I show that in the simplified case where all workers have equal cognitive ability, the set of two workers with the highest combined S* will always sort into θ₁. Since wages are increasing in S* and increasing in θ, and since S* and θ are complements, workers with higher social skills will earn relatively higher wages in high θ occupations.

17.

An alternative hypothesis, advanced by Krueger and Schkade (2008), is that gregarious workers have a preference for social interaction, and thus will accept a lower wage to work in a nonroutine occupation.

18.

This definition of routineness differs from the task measures used by Autor, Levy, and Murnane (2003), who use the 1977 Dictionary of Occupational Titles (DOT) measures “set limits, tolerances or standards” (STS) and “finger dexterity” (FINGER). They call these task measures “routine cognitive” and “routine manual” respectively. Autor and Dorn (2013) and other subsequent work combine these two measures into an index of routine task intensity (RTI). Occupations that are at least 50 percentiles higher on the RTI measure compared to my O*NET based measure include telecom and line installers, masons, tilers and carpet installers, pharmacists, and dental assistants. Occupations that rank as much more routine according to the O*NET measure include taxi drivers and chauffeurs, bus drivers, garbage collectors, and computer scientists.

19.

The three O*NET variables are (i) the extent to which an occupation requires mathematical reasoning; (ii) whether the occupation requires using mathematics to solve problems; and (iii) whether the occupation requires knowledge of mathematics. See the Online Data Appendix for details.

20.

O*NET gives the following definitions for the four items designed to measure social skills: (i) coordination—“adjusting actions in relation to others’ actions”; (ii) negotiation—“bringing others together and trying to reconcile differences”; (iii) persuasion—“persuading others to change their minds or behavior”; (iv) social perceptiveness—“being aware of others’ reactions and understanding why they react as they do.” Online Appendix Figure A1 demonstrates that my preferred measure of social skills is strongly correlated with other similar O*NET variables that capture coordination, interaction and team production. See the Online Data Appendix for details.

21.

The formal model is written in levels. However, taking logs in equations (16) and (17) would lead to a regression with the natural log of wages as the outcome and additive separability of cognitive skills and social skills. This implies that cognitive skills and social skills are complements in levels, but not in logs. However, I present main results using log wages to follow standard practice in the literature. Table I shows results for log wages, and Online Appendix Table A2 presents analogous results with hourly wages in levels. I find complementarity in both specifications, although it is stronger in levels than in logs.

22.

Unlike Table I, the results in columns (4)–(6) show little evidence of complementarity between cognitive skills and social skills. The results are different for two reasons. First, the sample in Table IV is restricted to ages 25–33, whereas Table I estimates returns to skills for prime-age workers. Skill complementarity is about 30% smaller when I restrict to ages 25–33 in Table I. Second, the definition of social skills in Table IV only includes self-reported sociability, whereas the measurement of social skills in Table I also includes participation in clubs and sports. Complementarity is about 50% smaller (but still statistically significant at the 5% level) when I use only survey responses to measure social skills in Table I.

23.

Note that this estimate differs from the worker fixed effects models in Table III, because those are estimated using a much larger age range. This suggests that the wage gain from switching to a social skill–intensive occupation was greater for older workers in the NLSY79 survey. Unfortunately, the panel design of the NLSY does not allow me to distinguish between age effects and cohort effects (i.e., whether the larger return for older workers is because the return to social skills increased over time or whether the return is constant but larger for later-career workers.)