Introduction

1 The task approach has attracted considerable attention in labor economics since the seminal work of Autor, Levy and Murnane [2003]. Tasks are the building blocks of production. Firms discharge some through machines and contractors. They combine the remainder into jobs, whose content depends on employees’ abilities and market conditions. A study of task assignment can thus provide valuable insight into the evolution of labor markets. Task data are seldom available at the individual level. Therefore, economists have typically examined jobs after they have been grouped into occupational classifications of mostly administrative origin (Stinebrickner, Stinebrickner and Sullivan [2019]). This approach treats each occupation as a bundle of tasks, the demand for which shifts with such shocks as automation and offshoring. Yet occupations evolve (Levy and Murnane [1996]; Autor, Levy and Murnane [2003]; Spitz-Oener [2006]). Autor ([2015], 7) writes: “[As] the routine cash-handling tasks of bank tellers receded.., banks recognized the value of tellers.. as salespersons, forging relationships with customers and introducing them to additional bank services like credit cards, loans, and investment products.” This observation suggests that job content is flexible: firms adapt assignments to changes in the relative costs of production factors, as bank clerks exemplified by assuming more cognitive tasks. [1] In consequence, there is no exact mapping from a job title to a set of tasks (Autor and Handel [2013]; Arntz, Gregory and Zierahn [2017]): today’s tellers share few duties with their counterparts from the 1970s, just as they cater to different clients at multinational banks and regional institutions, yet their jobs receive the same occupational code. Unlike tasks, occupations are not a precise economic concept: they are statistical tools, the result of complex algorithms and specific classifications. This paper explores the relation between job content and market conditions. In particular, we assess the impact of changes in the supply of skilled labor in France from 1991 to 2013. Thanks to public investment in higher education, university graduates increased their share of the workforce from 18 to 36% over this period. We exploit individual data from five surveys of work conditions, which allows us to compare jobs within occupations. Table 1 in the second section presents our task measures. Following the literature, we group them into three indexes for analysis: routine, cognitive and social. Our argument is threefold. First, job content is heterogeneous within occupations (Autor and Handel [2013]; Arntz, Gregory and Zierahn [2017]). For example, consider again bank clerks. As Figure 1 shows, there is significant variation in their tasks. If we divide the sub-sample by the number of tasks in each group, no cell contains more than 13% of observations and 90% report tasks in multiple categories. Second, university graduates hold a comparative advantage in cognitive work (Spitz-Oener [2006]; Acemoglu and Autor [2011]). Third, higher average educational attainment increased the supply of cognitive tasks and reduced their relative price, so workers spent more time on routine tasks instead. Figure 1 illustrates this shift: bank clerks were given more routine and fewer cognitive tasks as the share of university graduates rose from 14% of tellers in 1991 to 58% in 2013.

Figure 1

Distribution of the sum of task indicators for bank clerks by year

Distribution of the sum of task indicators for bank clerks by year

Note: Routine tasks (resp. Cognitive, resp. Social) comprise 4 (resp. 5, resp. 3) indicators. The figure presents the distribution of the sum of these indicators for bank clerks, in 1991 and 2013. See the second section for a full description of the task indicators.

2 The fourth section formalizes these ideas into a model. Workers supply one unit of labor, which they share between a routine and a cognitive task. Skilled workers hold a comparative advantage in the cognitive task. Unlike Autor, Levy and Murnane [2003] or Acemoglu and Autor [2011], we assume that workers do not specialize. We also assume that skills and tasks can be “unbundled,” using Choné and Kramarz [2022]’s terminology. Firms combine tasks into output. The model predicts two effects from an increase in the skill supply. Because skilled workers perform more cognitive work than the unskilled, a composition effect raises the cognitive content of aggregate output. On the other hand, a substitution effect obtains at the individual level: as the relative price of cognitive tasks decreases, each worker supplies more routine and fewer cognitive tasks.

3 To test these predictions, we regress our task indexes on the share of university graduates within each labor market by year. We assume that a separate labor market exists for each occupation within each region of France. Since schooling, migration and labor supply are endogenous, we forecast the graduate share on the basis of previous surveys to construct instruments. The first is the graduate share among workers who will still be under the minimum retirement age by the next survey. The second supposes that the contingents of each skill group in each labor market will evolve at the national rate between surveys (Bartik [1991]). We define both instruments in terms of birth regions rather than region of residence on account of migration. The exclusion restriction assumes that temporary local shocks are orthogonal to the initial distribution of graduate shares across labor markets (Goldsmith-Pinkham, Sorkin and Swift [2020]).

4 Our first specification includes fixed effects for occupation, region and year. Therefore, we obtain identification from variation in task assignment across workers within each labor market and the coefficients inform us about the presence of a substitution effect at the individual level. The second does not include occupation effects. The resulting coefficients combine variation in job content at the individual level with variation between occupations (hence, labor markets). As a consequence, they inform us about the existence of a composition effect.

5 Our results are twofold. For a given occupation, a higher graduate share is associated with more routine, fewer cognitive and fewer social tasks. The opposite pattern holds across occupations: the average job involves fewer routine, more cognitive and more social tasks. Hence, we find evidence for both theoretical predictions of an individual substitution effect and an aggregate composition effect. The estimates are significant but modest: the task indexes shift by 2 to 16% of a standard deviation for a rise in the graduate share of 10 percentage points around the nationwide share in 1990.

6 We examine task compensation as well. We show that an increase in the routine index lowers hourly wages by 0.6 to 1%, an increase in the cognitive index of one standard deviation raises them by 0.9 to 2.1% and an increase in the social index lowers them by 0.4 to 1.33%. The wage effects of routine and cognitive tasks decrease in magnitude between surveys, which is also consistent with the model. Our estimates are similar to those in Autor and Handel [2013], though theirs are based on different measures of job content and American data.

7 The task literature has greatly improved our understanding of labor markets. For example, Autor, Levy and Murnane [2003] argue that computers replaced labor in routine activities, raising the cognitive content of occupations and reshaping the occupational structure within industries. Similar analyses have shed light on automation (Agrawal, Gans and Goldfarb [2019]; Atack, Margo and Rhode [2019]; Gregory, Salomons and Zierahn [2022]; Spitz-Oener [2006]), employment polarization (Acemoglu and Autor [2011]; Firpo, Fortin and Lemieux [2011]; Autor and Dorn [2013]), gender gaps (Black and Spitz-Oener [2010]), immigration (Peri and Sparber [2009]), mobility (Gathmann and Schönberg [2010]), offshoring (Blinder [2009]; Jensen and Kletzer [2010]), part-time work (Elsayed, De Grip and Fourge [2017]), social skills (Deming [2017]), trade (Autor, Dorn and Hanson [2015]), the role of technology and of “Techies” (Harrigan, Reshef and Toubal [2023]), and more. This paper shows that market conditions influence task assignment within occupations. This finding highlights the need for nuance in discussing the future of work (Acemoglu and Restrepo [2019a], [2019b]; Arntz, Gregory and Zierahn [2017]). It does not suffice to examine the typical tasks in an occupation at present to forecast its susceptibility to automation or outsourcing. As we noted earlier, occupations evolve: workers may perform un-automated tasks more intensively, firms may develop new tasks for idle employees, etc. Rising educational attainment may facilitate this adjustment by preparing workers for lifelong learning and flexible roles.

8 The paper continues as follows. The second section presents the data. The third section discusses stylized facts. The fourth section introduces the model. The fifth section describes our empirical approach. The sixth section contains the results. The seventh section concludes.

Data

9 This section describes our data. Our sources are the French Labor Force Survey (Enquête Emploi, LFS), the Work Conditions Survey (Enquête Conditions de Travail, WCS) and the Work Organization Survey (Enquête Techniques et Organisation du Travail, WOS).

The Work Conditions Survey (WCS) and the Work Organization Survey (WOS)

14 INSEE conducted its first WCS in 1984. A supplementary survey of the outgoing group of the LFS, it enquired into sources of stress at work, whether physical (e.g., loud noises) or psychological (e.g., interacting with the public). INSEE repeated the exercise in 1991, 1998 and 2005. The WOS was a similar supplement to the LFS, focused on job content and the organization of work. It was undertaken in 1987 and 1993. The WCS and WOS averaged 20 000 respondents per wave. The Directorate for Research, Studies and Statistics at the Labor Ministry (Direction de l’animation de la recherche, des études et des statistiques, DARES) took responsibility over the WCS in 2013. It became an independent survey and involved 33 673 respondents in its first wave.

15 Researchers have often drawn task data from two sources from the US: the Dictionary of Occupational Titles and the O*NET. Both files provide scores for a large number of occupations in terms of activities, aptitudes and requirements (Autor, Levy and Murnane [2003]; Jensen and Kletzer [2010]). The WCS and WOS offer a significant advantage over these data: access to individual responses. As a consequence, we can explore heterogeneity within occupations and the joint task distribution across workers. [6]

16 We measure job content along three dimensions: routine, cognitive and social. We borrow this approach from the extensive literature on automation and offshoring (Autor, Levy and Murnane [2003]; Jensen and Kletzer [2010]; Handel [2012]). Following Spitz-Oener [2006], we construct task indexes by selecting relevant variables from the WCS and the WOS, transforming them into indicators and averaging the indicators. We selected variables on three criteria: they unambiguously pertain to one of our three categories, they are available across all years and the underlying questions are identically phrased across surveys. Note that they surveys report respondents’ original answers and interviewers did not help them interpret the questions.

17 Table 1 shows the means of each indicator by category and year. Routine tasks denote a lack of autonomy or initiative. Cognitive tasks involve decision-making. Indeed, our cognitive tasks capture autonomy (e.g., handling incidents) rather than intellectual difficulty (e.g., complex calculations). Social tasks require interaction with clients or the public. [7]

18 Hence, the way a firm is organized, in particular its hierarchical structure, clearly affects the amount of autonomy granted to workers (Caliendo, Monte and Rossi-Hansberg [2015]). This cognitive measure will therefore reflect less each worker’s university education than usual measures. In particular at the occupation level (see Table 1 below), we see that these measures are likely to capture complex decisions made at work, and not a theoretical view imposed outside of its context. Hence, the role given to autonomy (for cognitive tasks), repetition (for routine tasks), the public (for social tasks) makes the firm-specific structure central in our measure of tasks, something that cannot be captured by definitions of tasks uniquely based on ONET-style measures.

Table 1

Task measures by category

	2-digit occupation with highest incidence	Percentage of positive responses
		1991	1993	1998	2005	2013	All
Routine tasks
Production norms to be fulfilled within the day	Drivers	38	42,6	43	42,1	45,9	42,3
Repetitive movements	Unskilled manufacturing workers	29,6	24,5	28,7	27,9	41,2	30,4
Work rhythm determined by machinery	Unskilled manufacturing workers	12,8	11,5	13,6	13,9	18	14
Cognitive tasks
Choosing strategy to achieve goals	Senior technicians in the private sector	83,5	83,9	86,9	81,4	80,2	83,2
Departing from deadlines	Senior technicians in the private sector	35,7	37,1	36,2	36,6	34,3	36
Departing from instructions	Professionals in arts and culture	24,6	22,4	28	30,3	28,2	26,7
Handling incidents	Senior managers in the private sector	50,2	53,4	56,6	52,1	50,7	52,6
Social tasks
Contact with the public	Sales workers	60,7	61,3	62,4	68,5	70,9	64,8
Work rhythm determined by external demands	Sales workers	45,9	45	54,3	53,5	58	51,4

Task measures by category

19 Note that we limit the sample to the period from 1991 to 2013. We discard the 1984 WCS and the 1987 WOS because the LFS did not contain all of the variables of interest at the time. See the third section for further discussion.

20 We also construct an indicator of computer usage from the WCS and the WOS. We use it as a control variable to account for the influence of technological shocks on the skill supply and the demand for tasks.

Stylized facts

21 This section presents stylized facts about the French labor market.

22 Figure 2 illustrates the increase in educational attainment in France since 1990. University graduates constituted 36% of the employed workforce in 2012, up from 18%. The proportion of workers with secondary degrees rose from 12 to 19% in this period. By contrast, the fraction of workers without degrees fell from 41 to 20%. This upskilling process is largely due to sustained public investment in higher education. As education minister under President François Mitterrand in the mid-1980s, Jean-Pierre Chevènement initiated an effort to raise the high-school graduation rate to 80%. Modernization plans for tertiary education followed in 1990 (Université 2000) and 1999 (Université du troisième millénaire), which included the creation of eight universities and dozens of technical colleges. A very thorough analysis of this episode, performed mostly at the city-level using a matched employer-employee data source (the DADS) together with firm-level measures (balance-sheet from FICUS, in particular) can be found in Nimier-David [2023].

Figure 2

Share of university graduates and skill premium by year

Share of university graduates and skill premium by year

Note: The sample consist of employed wage workers. The skill premium is the ratio of median hourly wages of university graduates and less educated workers.

23 Table 2 summarizes the evolution of job content in this period. Workers performed 1.05 out of three possible routine tasks on average in 2013, up from 0.8 in 1991; in consequence, our index increased by 30% between 1991 and 2013. We observe a smaller rise in the incidence of social tasks, 21%. Cognitive tasks remained stable. These trends may seem surprising, but they are broadly consistent with cross-country evidence from the European Working Conditions Survey in Handel [2012]. Moreover, there is little evidence of polarization in the French labor market (Van Reenen [2011]; Verdugo, Fraisse and Horny [2012]; Verdugo [2014]) [8] and the impact of computerization was limited (Card, Kramarz and Lemieux [1999]; Goux and Maurin [2000]). Hence, it is plausible to find an increase in routine tasks and no change in cognitive tasks.

Table 2

Variation in task assignment

	Routine tasks			Cognitive tasks			Social tasks
	Mean	St. dev.	R²	Mean	St. dev.	R²	Mean	St. dev.	R²
1991	0.268	0.305	0.228	0.485	0.272	0.200	0.533	0.415	0.246
1993	0.262	0.300	0.231	0.492	0.257	0.178	0.532	0.405	0.180
1998	0.284	0.309	0.243	0.519	0.259	0.164	0.584	0.405	0.196
2005	0.280	0.309	0.203	0.501	0.267	0.119	0.610	0.393	0.154
2013	0.350	0.327	0.217	0.484	0.265	0.151	0.645	0.377	0.175
All	0.289	0.312	0.223	0.496	0.264	0.156	0.581	0.402	0.193

Variation in task assignment

Note: The table shows summary statistics for the average of task indicators in each group. The R² refers to a linear regression on individual characteristics and fixed effects (education, occupation, region and year).

24 Table 2 also reveals that covariates do not provide much insight into the distribution of job content. It shows the coefficient of determination from yearly linear regressions of task indexes on individual characteristics and fixed effects for education level, occupation and region. This model explains a quarter of the variation in task assignment at most—further evidence that job content is heterogeneous within occupations (Autor and Handel [2013]; Arntz, Gregory and Zierahn [2017]).

25 As Figure 2 shows, the skill premium shrank throughout the period: university graduates’ median hourly wage was 38% larger than other workers’ in 2012, down from 66% in 1990. Wage inequality fell as a result (Charnoz, Coudin and Gaini [2011]; Verdugo, Fraisse and Horny [2012]; Verdugo [2014]). Table 3 displays the growth in each decile of hourly wages between 1991 and 2005. [9] The ratio of the ninth to the first decile decreased by 8.3%. As the table shows, differences in job content between wage deciles diminished too. Routine tasks are more frequent at the bottom of the wage distribution; however, the ratio of the average routine score in the ninth wage decile to the average in the first rose by 18% between 1991 and 2005. Cognitive and social tasks are more common in the top of the distribution, but the ratio of cognitive scores decreased by 15% and that of social scores, by 10%.

26 Figure 3 presents average task indexes by education level in 1991 and 2013. As Spitz-Oener [2006] notes, university graduates perform fewer routine, more cognitive and more social tasks than other workers. However, these patterns weakened over time. Routine and social activities increased in incidence in both groups, but the change was larger among graduates in proportional terms. On the other hand, graduates discharged fewer cognitive tasks in 2013 than 1991. [10] As we discuss in the following sections, up-skilling may partly explain these changes: higher educational attainment may have reduced the relative price of cognitive tasks, leading workers to spend more time on routine tasks instead.

Figure 3

Change in task assignment by education level

Change in task assignment by education level

Note: The figure shows the average of task indicators in each group by education level and year.

Table 3

Task assignment and hourly wages by decile of hourly wages

	Hourly wages		Routine tasks		Cognitive tasks		Social tasks
	Level, 1991	Change, 91–05	Level, 1991	Change, 91–05	Level, 1991	Change, 91–05	Level, 1991	Change, 91–05
1	6.625	0.215	0.336	–0.011	0.396	0.123	0.482	0.220
2	7.494	0.193	0.327	–0.002	0.409	0.109	0.495	0.202
3	8.351	0.181	0.317	0.008	0.424	0.092	0.507	0.187
4	9.207	0.168	0.302	0.028	0.445	0.065	0.520	0.169
5	10.192	0.151	0.287	0.042	0.464	0.049	0.531	0.156
6	11.246	0.154	0.267	0.060	0.489	0.033	0.544	0.142
7	12.712	0.154	0.240	0.084	0.521	0.016	0.559	0.125
8	14.988	0.142	0.202	0.117	0.562	–0.008	0.573	0.109
9	19.485	0.114	0.151	0.171	0.627	–0.047	0.583	0.094

Task assignment and hourly wages by decile of hourly wages

Note: Wages are shown in constant euros (base 2015). The table shows average task indexes within each wage decile and the proportional change in averages from 1991 to 2005.

27 We conclude this section with an overview of economic conditions in the 1990s and 2000s. Growth was slow and unsteady. Real GDP per capita expanded at an average yearly rate of 1.8% between 1990 and 2013. There were four recessions in this period (in 1992, 2001, 2008 and 2012). Unemployment was persistently high, averaging 8.8%, and the shares of both fixed-term contracts and part-time jobs increased (from 6 to 11% and from 11 to 17%, respectively). These decades are also noteworthy for the reduction of the legal workweek from 39 to 35 hours between 2000 and 2002, which inflated hourly wages and compressed their distribution (Aeberhardt, Givord and Marbot [2016]).

Theoretical framework

28 This section develops a simple theoretical framework for our empirical analysis of the interaction between the supply of skilled workers and task assignment. We adapt the model by Peri and Sparber [2009].

Task Demand

29 Consider an economy in autarky. A representative firm combines tasks into a consumption good (y). Tasks may be routine (r) or cognitive (c). For simplicity, we assume that production does not require capital. The production technology is:

30 where σ controls the elasticity of substitution between inputs (n.b. σ > 0). [11] Production does not require both tasks (unless σ → 1), but this functional form implies that the firm will always mix them in equilibrium, as we observe in the data.

31 The firm purchases task services on frictionless labor markets. [12] The consumption good is the numeraire. Therefore, profits are: y – w_r r – w_cc, where w_r is the price of a unit of routine tasks (analogously for w_c). By combining the necessary conditions for profit maximization, we find the relative demand for tasks:

32 where ω is the price ratio: ω ≡ w_r / w_c. The firm decreases the routine content of production in response to an increase in the cost of routine tasks. This inverse relationship is important for our results, though its precise functional form is not.

Task Supply

33 The economy comprises a measure p of skilled workers (s = 1) and a measure 1 – p of unskilled workers (s = 0). Each worker is endowed with one unit of labor. They do not derive utility from leisure. Hence, they apportion x_s of their time to the supply of r_s in routine tasks and 1 – x_s to the supply of c_s in cognitive tasks. The resulting task supply is:

34 where β ∈ (0, 1), α_rs > 0 and α_cs > 0. The curvature parameter β implies that workers become less productive as they repeat tasks, which may reflect technical limitations (e.g., fatigue) or a preference for variety at work. The scale parameters α_rs and α_cs determine total productivity. We assume that skilled workers enjoy a relative advantage at cognitive tasks: α_c1/α_r1 > α_c0/α_r0.

35 Because savings bear no interest and the model is static, workers do not save. Therefore, they maximize utility by maximizing their income,

36 through the choice of x_s. The optimal allocation (x^*_s(ω)) satisfies:

37 where x^*_s(·) is an increasing function.

38 Equation 3 has three implications. First, the supply of routine tasks increases with their relative price. Second, unskilled workers perform more routine tasks than the skilled: r₀ > r₁; conversely, c₁ > c₀. This property is a consequence of their relative advantages and finds support in the data (see Figure 3). Third, workers do not specialize: r_s > 0 and c_s > 0 for all s. This feature is due to the non-linearity in the task supply (see Equation 2). Empirical evidence support it too: fewer than 15% of workers in our sample report tasks in a single category. Our model thus differs from Acemoglu and Autor [2011] or Autor, Levy and Murnane [2003], where the task supply is linear and workers perform one task each. [13]

Equilibrium and Comparative Statics

39 Equilibrium obtains when prices, w_r and w_c, ensure that each task market and the goods market clear:

40 We can find the equilibrium in two steps. Equations 1 and 3 fix relative prices. We can then determine absolute prices by clearing the goods markets.

41 This paper investigates the impact of changes in the supply of skilled labor on task assignment. The model has implications for our empirical analysis. To see this, first combine Equation 1 with Equation 3:

42 Implicit differentiation then reveals that the price ratio, ω, is an increasing function of the share of skilled workers, p. Simple manipulations of previous optimality conditions yield . For a fixed ω, a higher p induces an expansion in the aggregate supply of cognitive tasks, since skilled workers spend more time on cognitive tasks than the unskilled. Equilibrium requires that cognitive tasks become relatively cheaper—i.e., ω must go up. [14] As the price ratio rises, each worker supplies more routine and fewer cognitive tasks, while the firm demands more cognitive and fewer routine tasks. Therefore, the new equilibrium differs in two aspects. There is a substitution effect toward the routine at the worker level by Equation 3. Nonetheless, the cognitive content of aggregate output increases through a composition effect. These effects are generally nonlinear in both the initial share of skilled workers and the magnitude of the shift in p between equilibria.

Empirical approach

47 The model implies that an increase in the skill supply should correlate with more routine and fewer cognitive tasks at the individual level (the substitution effect) but with fewer routine and more cognitive tasks at the aggregate level (because of a composition effect). To test the first prediction, we use variation in job content across individual workers as they are exposed to different supply shocks in each labor market. To test the second, we use variation between labor markets. Even though social tasks are not part of the model, we include them in the empirical analysis in light of the rising importance of social skills in modern labor markets (Deming [2017]).

48 We assume that workers segregate into distinct labor markets by administrative region, two-digit occupation and year. Our analysis encompasses 2,415 markets by this definition (21 × 23 × 5). [16] We use the share of university graduates as a proxy for the skill supply.

49 Consider observation i in occupation o_i region r_i and year t_i. Write y_i for i’s individual task score, p_o,r,t for the graduate share in their labor market, x_i for a vector of individual characteristics (see second section) and u_i for the residual. We consider two specifications. The first includes fixed effects for occupation, region and year:

50 We are interested in γ₁ and γ₂. The quadratic term helps us take nonlinearities into account (see fourth section, subsection “Equilibrium and Comparative Statics”). Because it includes a full set of fixed effects, this regression only exploits variation in task assignment within labor markets for identification; hence, it assesses the extent of substitution between tasks at the individual level. The second specification leaves occupation effects out:

51 This regression mixes variation at the individual level and variation between occupations (hence, between labor markets), so it provides insight into the task content of aggregate output. So far as γ^ʹ₁ and γ^ʹ₂ differ from γ₁ and γ₂, it informs us about the existence of a composition effect at an aggregate level.

52 Ordinary least squares need not yield consistent estimates of γ₁, γ₂, γ^ʹ₁ and γ^ʹ₂. One concern is measurement error, given that we estimate the graduate share from the LFS. Although we pool observations across three years for additional precision (see second section), we may still lack power for some occupations in less populated regions. Another worry is endogeneity in schooling decisions, migration and workforce participation. For example, a local technological shock could affect both the relative demand for cognitive tasks (because routine tasks are automated, say) and the skill supply (because skilled workers immigrate from other regions, say).

53 Therefore, we use instrumental variables for identification. We construct two instruments by projecting the graduate share on the basis of earlier waves of the LFS. [17] For concreteness, consider observation i from 1991. We use three surveys to compute the instruments for i: 1983, 1984 and 1985. [18] The first instrument exploits retirements. It is the graduate share among such workers as were born in the same region as i, have the same occupation as i and will not reach the minimum retirement age by 1991 (viz. 60 years). This definition helps us cancel the effect of endogenous education (by dropping incoming cohorts) and migration (by using birth regions). It does not involve actual retirements, lest we introduce bias from participation decisions. This instrument is relevant because retirements boosted the share of graduates in this period by removing less educated cohorts from the labor force. The second is a Bartik instrument (Bartik [1991]). Consider such workers that were born in the same region and have the same occupation as i. To construct the instrument, we multiply the contingent of each skill group within this population by the corresponding growth rate across the entire workforce and compute the implied graduate share. In computing the national growth rates, we exclude i’s birth region and occupation. The exclusion restriction requires the same assumption for both instruments: the initial distribution of graduate shares across labor markets must be orthogonal to local shocks (Goldsmith-Pinkham, Sorkin and Swift [2020]). Following these authors, to assess the validity of the design, they note that “in some settings there is a pre-period, as in a standard difference-in-differences design.” Hence, by testing for (the absence of) pre-trends, it becomes possible to assess the assumption that the common shock caused the changes. To do so, we free-ride on Nimier-David [2023] who examines the same episode as we do, directly in a difference-in-differences framework at the local level. By using recent techniques developed by Chaisemartin and d’Haultfoeuille [2020], he shows repeatedly the absence of any pre-trend in all the variables he examines; most particularly the share of graduates. Results shown below comfort us in this assessment of the validity of our instruments.

54 Table 4 shows coefficients from linear regressions of the graduate share on the instruments. Both are highly correlated with the graduate share. The Bartik instrument is slightly stronger, perhaps because it takes the average education of incoming cohorts into account. Covariates reduce the coefficients, but they remain significant at any conventional level. [19] The consistency of the results across the two sets of instruments lends further weight to our choice of identification strategy.

Table 4

Linear regression of the graduate share on instruments

	(1)	(2)	(3)	(4)
Retirement instrument	1.000***	0.418***
	(0.006)	(0.022)
Bartik instrument			0.960***	0.476***
			(0.005)	(0.022)
Partial F-statistic	3.2104	3.5102	3.5104	4.9102
Controls
Individual characteristics		16		16
Occupation fixed effects		22		22
Region fixed effects		20		20
Year fixed effects		4		4
Fit
Observations	94,990	94,253	94,990	94,253
Adjusted R²	0.963	0.980	0.972	0.982

Linear regression of the graduate share on instruments

Note: The table shows coefficients from ordinary linear regressions. Standard errors (in parentheses) are clustered by occupation, region and year. The outcome is the share of university graduates by occupation, region and year. Both instruments are a projection of the share of graduates by birth region and occupation. The retirement instrument supposes that the graduate share will only evolve between surveys because of retirements. The Bartik instrument supposes that the contingents of graduates and nongraduates will evolve at the national rate in each local market between surveys. See fifth section for further detail and discussion. The partial F-statistic tests the joint significance of the instruments. * significant at the 10% level, ** 5%, *** 1%.

Results

The Impact of the Skill Supply on Task Assignment

55 Table 5 displays our main results: estimates of the impact of changes in the graduate share on job content. Each column presents one combination of task index and covariates. The first two columns show coefficients from regressions of the routine score; columns 3 and 4, of the cognitive score; columns 5 and 6, of the social score; the last two columns, of the ratio between the routine score and the sum of the three scores. [20] Odd columns display coefficients from regressions with fixed effects for occupation; even columns, without them. Each column contains estimates by both two-stage and ordinary least squares.

Table 5

Impact of changes in the graduate share on task assignment

	Routine tasks		Cognitive tasks		Social tasks		Routine tasks (share)
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
Retirement instrument
Share of university graduates by occupation, region and year	0.207*** (0.075)	–0.676*** (0.032)	–0.043 (0.086)	0.416*** (0.028)	0.040 (0.122)	0.742*** (0.054)	0.052 (0.048)	–0.647*** (0.029)
Squared share of university graduates by occupation, region and year	–0.433*** (0.090)	0.448*** (0.036)	–0.439*** (0.123)	–0.254*** (0.032)	–0.404*** (0.161)	–0.720*** (0.062)	0.028 (0.056)	0.487*** (0.032)
Bartik instrument
Share of university graduates by occupation, region and year	0.136* (0.079)	–0.735*** (0.033)	–0.100 (0.078)	0.447*** (0.027)	0.018 (0.119)	0.783*** (0.054)	0.043 (0.051)	–0.693*** (0.031)
Squared share of university graduates by occupation, region and year	–0.334*** (0.084)	0.521*** (0.037)	–0.338*** (0.104)	–0.292*** (0.032)	–0.459*** (0.151)	–0.768*** (0.062)	–0.044 (0.054)	0.543*** (0.034)
No instrument
Share of university graduates by occupation, region and year	0.026 (0.054)	–0.729*** (0.031)	–0.061 (0.045)	0.404*** (0.025)	–0.004 (0.082)	0.810*** (0.049)	0.004 (0.033)	–0.686*** (0.028)
Squared share of university graduates by occupation, region and year	–0.132*** (0.054)	0.512*** (0.034)	–0.046*** (0.044)	–0.239*** (0.029)	–0.292*** (0.084)	–0.797*** (0.057)	–0.009 (0.031)	0.533*** (0.031)
Controls
Individual characteristics	16	16	16	16	16	16	16	16
Occupation fixed effects	22		22		22		22
Region fixed effects	20	20	20	20	20	20	20	20
Year fixed effects	4	4	4	4	4	4	4	4
Fit
Observations	94,253	94,253	94,253	94,253	94,253	94,253	93,592	93,592
Adjusted R² (OLS)	0.219	0.144	0.156	0.114	0.193	0.085	0.301	0.172

Impact of changes in the graduate share on task assignment

Note: Standard errors (in parentheses) are clustered by occupation, region and year. The share of routine tasks is the ratio of routine tasks to the sum of task indexes. See second section for a description of the task indexes. See fifth section for a description of the instruments. * significant at the 10% level, ** 5%, *** 1%.

56 The table shows separate results for each instrument. [21] Our estimates may be difficult to interpret because of the quadratic term in Equation 4 and in Equation 5. For convenience, we define the standardized effect as the change in a given task index for an increase of 10 percentage points in the graduate share around the nationwide share in 1990 (i.e., 5 points below and 5 points above 18%, hence from 13 to 23%). The discussion focuses on estimates by two-stage least squares with the retirement instrument for parsimony’s sake. Our conclusions are robust to the choice of estimator.

57 Consider routine tasks first. As column 1 shows, the effect of an increase in the graduate share on routine tasks is concave within occupations. According to the causal estimates, it peaks when the graduate share nears 24% and turns negative when it reaches 48%. Since the nationwide graduate share was 18% in 1991, we conclude that rising educational attainment caused an increase in the routine job content in France, corroborating the prediction of a substitution effect at the worker level from our model. As a reminder, workers perform more routine tasks in the model because their relative price goes up as the skill supply expands. The magnitude is modest: the standardized effect is 0.005 or 2% of a standard deviation. [22] The causal coefficients are larger than the estimates by linear regression, but they agree in direction. Column 2 repeats this exercise without occupation indicators. We find the opposite pattern: the impact of an expansion in the graduate share is convex and uniformly negative, bottoming out when the graduate share is just past 75%. Unlike the previous regression, this specification uses variation in job content between occupations for identification. Therefore, it captures a mixture of the change in the incidence of routine tasks within occupations and growing employment in cognitive occupations (in which skilled workers specialize). It constitutes evidence of the composition effect in the model. The standardized effect is –0.051 or 16% of a standard deviation.

58 As theory suggests, cognitive tasks mirror the routine. Column 3 implies that the impact of an increase in the graduate share on the cognitive score is negative and concave. The standardized effect is –0.02 or 8% of a standard deviation. By contrast, we find a positive and concave relationship upon dropping the occupation indicators, as column 4 shows. It peaks as the graduate share nears 82%. The standardized effect is 0.032 or 12% of a standard deviation. These estimates are again consistent with the two main predictions of our model: a substitution effect away from cognitive tasks at the worker level and an aggregate composition effect toward cognitive tasks.

59 Columns 5 and 6 examine social tasks. Unlike the routine or the cognitive, social tasks are not part of our model. We analyze them nonetheless for completeness. Perhaps because of complementarities between cognitive and social skills (Deming [2017]), our estimates are broadly similar to the regressions of the cognitive score. If we include occupation indicators, we find a negative impact of an increase in the graduate share on social tasks. The quadratic term is especially salient. The standardized effect is –0.011 or 3% of a standard deviation. Should we exclude occupation indicators, the response function becomes positive and concave. The maximum occurs when graduates represent 52% of employed workers. The standardized effect is 0.048 or 12% of a standard deviation.

60 The last two columns show regressions for the ratio of the routine score to the sum of task scores. Hence, the resulting coefficients combine the individual effects in columns 1 through 6. They are consistent with the regressions of the routine score in the first two columns, but the coefficients are not significant when occupation indicators are included.

The Impact of Task Assignment on Wages

61 Our model predicts that an increase in the skill supply should raise the price of routine tasks and reduce that of cognitive tasks. As a rudimentary test of this prediction, we undertake yearly regressions of wages on task assignment. Because educational attainment increases throughout the period, there should be changes in task prices if the theory is correct. Our coefficients are not causal (DiNardo and Pischke [1997]; Autor and Handel [2013]), since task assignment is a function of workers’ comparative advantages and we do not have instruments for tasks. We present them nonetheless as preliminary empirical evidence and for comparison with Autor and Handel [2013]. Table 6 shows our results. Note that we do not use the 2013 WCS because it did not collect comparable wage data. [23]

Table 6

Impact of job content on wages

	Hourly wages (log)			Monthly wages (log)
	1991	1998	2005	1991	1998	2005
Routine tasks	–0.033*** (0.007)	–0.024*** (0.008)	–0.022*** (0.010)	–0.029*** (0.008)	–0.030*** (0.008)	–0.031*** (0.009)
Cognitive tasks	0.078*** (0.010)	0.050*** (0.010)	0.035*** (0.011)	0.099*** (0.010)	0.081*** (0.010)	0.054*** (0.012)
Social tasks	–0.010* (0.006)	–0.033*** (0.006)	–0.017** (0.007)	–0.015*** (0.006)	–0.018*** (0.007)	–0.014* (0.008)
Controls
Individual characteristics	16	16	16	15	15	15
Occupation fixed effects	22	22	22	22	22	22
Region fixed effects	20	20	20	20	20	20
Fit
Sample	All	All	All	Full time	Full time	Full time
Observations	16,819	17,522	15,514	14,936	14,694	12,790
Mean outcome	2.396	2.420	2.555	7.448	7.483	7.528
Adjusted R²	0.606	0.610	0.548	0.634	0.634	0.599

Impact of job content on wages

Note: The table shows coefficients from ordinary linear regressions. Standard errors (in parentheses) are clustered by occupation and region. See second section for a description of the task indexes. * significant at the 10% level, ** 5%, *** 1%.

62 The first three columns consider hourly wages. Routine and cognitive tasks have opposite effects on pay: routine tasks lower wages, whereas cognitive tasks raise them. The penalty for routine tasks falls from 3.3% of hourly wages in 1991 to 2.4 in 1998 and 2.2 in 2005. These estimates represent the wage loss for performing all three routine tasks in the survey as opposed to none. Conversely, the premium for cognitive tasks declines from 7.8% of hourly wages in 1991 to 5 in 1998 and 3.5 in 2005. The pattern of changes between surveys match our theoretical predictions. The picture is less clear for social tasks: they reduce pay as well, but the coefficient increases in magnitude from 1991 to 1998 and decreases from 1998 to 2005. Deming [2017] also finds a negative correlation between social tasks and compensation.

63 The last three columns examine monthly wages. We restrict the sample to full-time workers. Routine tasks have similar effects on monthly and hourly wages, but the monthly penalty is stable across years. The impact of cognitive tasks on monthly wages is slightly larger than its hourly counterpart. It too shrinks between surveys, going from 10% of monthly wages in 1991 to 5.4% in 2005. By contrast, the role of social tasks is equivocal. The coefficient is positive in 1991 and 2005 but negative in 1998. Although we find a similar dip in 1998 for hourly wages, all coefficients are negative in that case.

64 Our wage regressions are consistent with those in Autor and Handel [2013]. The authors classify tasks into three groups: abstract, routine and manual. (Their abstract tasks correspond to our cognitive. They do not discuss social tasks.) They study the impact of job content on wages within occupations with survey data from the United States. Although they examine a different labor market and measure job content through different variables, their coefficients are surprisingly similar to ours in direction and magnitude. Albeit exploratory, our analyses should provide useful guidance for future research on task assignment.

Conclusion

65 This paper contributes to a growing literature about task assignment. Since the seminal work of Autor, Levy, and Murnane [2003], this research has provided insight into employment polarization, wage inequality and much else. It has mostly studied the influence of job content on an outcome of interest. We take the opposite approach and investigate the determination of job content in equilibrium. In particular, we show that the skill supply affects job content by analyzing the impact of an increase in the share of university graduates in the French workforce from 18% in 1991 to 36% in 2013. We find that higher average educational attainment is associated with more routine, fewer cognitive and fewer social tasks within occupations and with fewer routine, more cognitive and more social tasks across occupations. Recently, Deming [2023] remarked that returns to cognitive skills in the US have stagnated, and even decreased, when the demand for educated workers continued to grow. Both his and our findings show how skills are equilibrium outcomes and not necessarily the result of some inexorable force: technical change.

66 Our results have three methodological implications for future research. First, researchers should explore variation in job content within occupations in greater depth. Second, occupations evolve, so care is needed in analyzing long-term trends in labor markets on the basis of rigid occupational classifications. Third, identification deserves attention as task assignment is endogenous to both worker aptitudes and aggregate conditions.

67 Our approach has two significant limitations. Although we can measure changes in task assignment at the individual level, we cannot distinguish the intensive margin (i.e., changes in tasks for a given worker in a given job) and the extensive margin (i.e., changes through job creation, job destruction and employee turnover). Moreover, we cannot observe innovation in tasks in our data. Task creation has historically offset the pressure of automation on wages and employment (Acemoglu and Restrepo [2019a], [2019b]), such as we experience today. Panel data would help us address these shortcomings (Stinebrickner, Stinebrickner and Sullivan [2019]). It would also allow us to study the influence of job content on workers’ careers, wage inequality and more—a promising avenue for future research.