Job Attributes and Occupational Changes: A Shift-Share Decomposition by Gender and Age Group for Canada, 2006–2016

Abstract:Les changements technologiques ont deux effets de premier ordre sur la nature du travail. Premièrement, les nouvelles technologies peuvent entraîner des changements au sein des professions des personnes sur le marché du travail, et, deuxièmement, elles peuvent pousser ces personnes à passer d'une profession à l'autre. Afin de quantifier ces effets, la présente étude procède à un rapprochement des données détaillées sur les professions tirées des recensements canadiens de 2006 et de 2016 avec des données détaillées qui associent chaque profession à des ensembles de tâches, d'activités et de compétences requises pour cette profession. Les résultats révèlent que l'importance des attributs liés aux interactions sociales et aux tâches cognitives non routinières a augmenté de manière considérable. De plus, la majeure partie de cette augmentation s'est produite au sein de professions étroitement définies. Les hommes ont été plus touchés par les changements observés que les femmes.Abstract:Technological change has two first-order effects on the nature of work. First, new technologies can cause changes within workers' occupations, and second, it can push workers to move between occupations. To quantify these effects, we match detailed occupational data from the 2006 and 2016 Canadian censuses to detailed data that associates each occupation with sets of tasks, activities, and skills required for that occupation. Our results show that the importance of attributes related to social interactions and non-routine cognitive tasks has increased significantly. Moreover, most of the increase occurred within narrowly defined occupations. Men were more affected by the observed changes than women.


Introduction
While many studies focus on the impact of technological changes on levels of employment, these changes are much more likely to impact the nature of work (Autor and Handel 2013). 1 Our study focuses on the evolving nature of work.We consider the impact of technological change on the job market to be two-fold: 1) new technologies induce changes in the tasks performed within occupations, and 2) new technologies can induce changes in the occupational composition of employment. 2 In order to quantify these two effects, it is necessary to consider the occupations as being sets of tasks, activities, and required skills.
For example, the most well-known study following this approach is Autor, Levy, and Murnane (2003) who classify tasks associated with a job according to a two-by-two matrix contrasting routine versus non-routine tasks and manual versus cognitive tasks.Non-routine tasks are considered least likely to be automated and require a greater degree of adaptation and critical judgment (see Autor 2015).Within this framework, Autor, Levy, and Murnane (2003) found that during the 1960-1998 period, computerization affected job skill demands and task changes, causing reduced labour input of routine tasks (either manual or cognitive) and increased labour input of non-routine cognitive tasks.
However, in this article, we use a richer descriptive approach to evaluate the impact of technological change on the nature of work that follows that of Freeman, Ganguli,  and Handel (2020).They matched job characteristics from the Occupational Information Network (O * NET) database, developed by the United States Department of Labor's Employment and Training Administration, to occupational employment from the Occupational Employment Statistics for 2005 and 2015.Building indices for 17 O * NET job attributes, they used a shift-share decomposition to reveal that, in the United States, changes within occupations are by far the determining factor in the variation in attributes, almost always being of a magnitude higher than changes between occupations.
Our article makes four distinct contributions.First, it is the first to apply Freeman, Ganguli, and Handel's (2020)  shift-share decomposition to Canadian data.It also extends the choice of job attributes that are analyzed.Third, it examines if variations in attributes vary by gender due to differences in occupational sorting.Finally, it also shows how decompositions vary by age groups.
Following Freeman, Ganguli, and Handel (2020), our approach relies on matching occupational data from the 2006 and 2016 Canadian censuses to O * NET.Census data provide the most comprehensive information on occupations for all Canadian workers.Additionally, our choice of 2006 and 2016 census data allows the most direct comparisons with the original study's results. 3This period is long enough to allow us to detect underlying trends in the evolution of the labour market.Furthermore, this period is rich in technological changes and evolution, particularly the rise of technologies linked to numerical intelligence, significantly impacting the labour market.
We select from 24 O * NET attributes divided into six categories: (1) physical work, (2) social interactions, (3) cognitive skills, (4) verbal cognitive skills, (5) non-cognitive skills, and (6) knowledge.Categories (1) and (5) contain attributes related to jobs that involve more manual tasks, while attributes in categories (3) and ( 4) are related occupations that emphasize cognitive tasks.Category (2) consists of attributes related to social interactions; these tasks are difficult to automate, as they often require direct physical proximity and the ability to interact spontaneously with other individuals (Acemoglu and Autor  2011; Autor and Dorn 2013).Category ( 6) is based on knowledge in different subjects.
We then construct aggregate indices for a particular job attribute and decompose the change in the index between 2006 and 2016 using a shift-share decomposition as described by Freeman, Ganguli, and Handel (2020).This method breaks down the variation in the index of the attribute into two components: (1) variations within the occupation and (2) variations due to changes in occupational structure.
Similar to Freeman, Ganguli, and Handel (2020), we found that indices related to cognitive and social job attributes increased more than the attributes linked to physical work or physical activities.Additionally, we found that the pattern of changes in Canada was due to variations in attributes mostly within narrowly defined occupations rather than between occupations.
Additionally, we go beyond Freeman, Ganguli, and  Handel (2020) by applying the shift-share decomposition by gender, yielding additional interesting results.Although job attributes do not vary by gender, the aggregate indices may vary by gender in a specific year due to gender differences in occupational sorting.First, we found that the labour markets for men and women diverged with respect to physical abilities and physical work, with changes in those attributes often moving in opposite directions, increasing for men and decreasing for women.Job attributes linked to cognitive abilities increased for both genders, and more so for men.As those indices linked to cognitive and social job attributes were higher for women in 2006, some convergence in attributes for men and women over the period was implied.
Finally, we show how those attributes vary by age, both in a cross-sectional context and using synthetic cohort path analysis.

Occupational Data
We use data on employment in occupations from the 2006 and 2016 censuses.The Census contains detailed information on the number of workers aged 15 and over doi:10.3138/cpp.2022-009 in each occupational category according to the National Occupational Classification (4-digit NOC).
The NOC is a system developed through a collaborative partnership between Employment and Social Development Canada and Statistics Canada.The NOC presents a systematic classification structure that categorizes all occupational activities in Canada.The 2016 NOC (the current version) includes approximately 30,000 job titles, further organized in a four-level hierarchical structure.
The first level includes ten major occupational categories, the second consists of 40 major groups, the third includes 140 intermediate groups, and the fourth is divided into 500 groups (Statistics Canada 2018).We use this last level for our analyses.
The content of the NOC is updated every five years, and a structural review is conducted every ten years.Therefore, we used the gateways

Job Attributes Data
The O * NET is an occupational information system created in 1998 by the US Department of Labor.O * NET has its own occupational classification, called O * NET-SOC because it is based on the Standard Occupation Classification (SOC) of the US Bureau of Labor Statistics. 4 The O * NET database "describes the occupational characteristics of occupations from a set of exhaustively defined descriptors [...] to describe the skills required to exercise a given trade[...]" (Hart 2010).Descriptors (or attributes) are classified into a multitude of categories that describe characteristics related to workers or jobs.
Figure 1 presents the six categories of the O * NET® Content Model.Among the attributes describing the characteristics of workers are abilities, skills, knowledge, and the level of education and training required (i.e., education and training).The attributes describing the characteristics related to jobs include work activities, work contexts, and tasks, among others.
For all O * NET-SOC occupations, these attributes are assigned a value according to a specific scale and criteria.These values are partially updated regularly; each annual version of the O * NET database contains new information for around 100 occupations, while the other 700-800 remain largely unchanged.However, sometimes two or three updates occur in the same year.
The values assigned to the different attributes are from questionnaires specific to each category measured on a scale of 1 to 5. Questionnaire responses on education and training, knowledge, work activities and work context are collected from approximately 20-40 job holders for each occupation.Approximately 25 experts address the occupations for which it is more difficult to create such a panel of job incumbents.Questionnaire responses are collected from a group of eight work analysts on abilities; ratings are then based on the title, definition, tasks of the occupation, and the results from the work activities and work contexts questionnaires (Donsbach et al. 2003).Since 2008, work analysts have been completing the skills questionnairepreviously completed by job incumbents-thus creating a break in this series (Handel 2016). 5he job attribute measure for each occupation in the censuses is obtained by matching the NOC codes with the O * NET database using a gateway created by the Brookfield Institute for Innovation + Entrepreneurship.This step matches the abilities from the O * NET 20., 7 These attributes are chosen as they are related to job characteristics that have received considerable attention in the literature regarding the impact of technological changes on the future of work: automation, repetitiveness of work, decision latitude, interpersonal relationships, level of education, and knowledge required within an occupation.

Empirical Model
In this section, we discuss how we match the O * NET data to the occupational structure of the Canadian censuses and continue with the definitions of the occupational attribute indices. 8

Data Preparation
An underlying assumption while matching the O * NET attributes to the Canadian censuses is that the value of these attributes also applies to the Canadian context.This means that each NOC occupation should correspond to at least one O * NET occupation (Vu 2019).
However, the first step in constructing the sample involves finding the correspondence between the occupational structures of the 2006 and 2016 censuses.This is because the structure of the NOC was revised entirely in 2011.Thus, the data on numbers of workers by occu- In several cases, the correspondence is simple.The occupation from the NOC corresponds to the same occupation under the two O * NET classifications.However, in three cases, we made arbitrary decisions.
First, as the 2006 and 2010 O * NET-SOCs are more granular than the 2016 NOC, the same NOC occupation may have several O * NET equivalents.We take the average values associated with the equivalent O * NET occupations in those cases.Third, some NOC occupations correspond to O * NET occupations for which there is no data in either the 20.3 or 10.0 O * NET file.These occupations are excluded from the analysis that follows. 11inally, we are left with 446 of the 500 occupations included in the 2016 NOC.Upon identifying the workers for whom matching with the O * NET could not be carried out, we obtained a coverage rate of 85.7 percent of Canadian workers.In our analyses differentiated by gender, the coverage rate is slightly lower for women than for men (83.0%versus 88.2%, respectively). 12Indices First, let us define A pt as being the value of the O * NET attribute A of the occupation p of the NOC 2016 in year t, and W pt as the proportion of the active population that is part of from occupation p in year t.The aggregate index of the attribute, i.e., the value of the attribute for the entire working population, will therefore be: The first term of the equation represents the contribution of the variation in attributes within occupations, i.e., the variation within occupations (ΔA p ), weighted by the proportion of the labour force that is part of the occupation in the first year reference (W p2006 ).The second term of the decomposition represents the contribution of the variation resulting from the movement of workers between occupations, i.e., the variation in the proportion of the active population occupying the occupation p (ΔW p ), weighted by the value of A during the first reference year (A p2006 ).Finally, the third term is an interaction term that captures the residual variation.

Results
We begin by presenting the results for the Canadian population as a whole and continue with shift-share decompositions differentiated by gender and age.

Aggregate Results
Table 2 shows the value of the job attribute indices for 2006 and 2016, the variation between the two years, and the results of the decomposition of the variation into within, between, and interaction components.
We note that the aggregate index of the vast majority of attributes (21 out of 24) increased between 2006 and 2016, while in Freeman, Ganguli, and Handel (2020)  more mean changes were negative.This is the case for some attributes in the physical work category and for mathematics in the knowledge category.Specifically, in the physical work category, 3 out of 4 changes were negative in Freeman, Ganguli, and Handel (2020), while we observe a decline only for degree of automation.In the knowledge category, the more striking difference is the divergence between mathematics, which increased in our data and declined in Freeman, Ganguli, and Handel's.However, similar to Freeman, Ganguli, and Handel (2020), our results also show the biggest increase in the knowledge category to be computer and electronics.Finally, another difference is the strong increase we see in the category of knowledge, where administration and management didn't increase as much as it did in the US.Since Canada was starting from a lower index value in 2006, we observe some convergence for this knowledge between the two countries.
Additionally, in Canada, we find that the attributes that declined or had very little growth are related to manual jobs involving the physical work and non-cognitive skills categories.The automatization, overall body coordination, and manual dexterity job attributes declined between 2006 and 2016, while bending or body twisting, multilimb coordination, and stamina increased marginally.This is also the case for mechanical knowledge-often associated with blue-collar jobs-experiencing slower growth than the other three knowledge indices.
Consistent with Deming (2017), we find that social skills are increasingly important in the labour market.The aggregate index of the two O * NET attributes related to social interactions-coordinating or directing other people and being in contact with other people-experienced the strongest growth between the two reference years.Additionally, we find that knowledge-computers and electronics and business administration-and cognitive skills-ease of conceiving ideas and creativity-also increased.
The shift-share decomposition of the variation in attributes is presented in the last three columns of Table 2. Like Freeman, Ganguli, and Handel (2020), we find that the effect of the variation in the importance of the attributes within occupations is generally greater than that caused by the movement of workers between occupations; this is the case for 17 of the 24 attributes (with just over 70% of the variation due to within changes).The absolute variation within occupations (0.098) is, on average, approximately four times larger than the absolute variation between occupations (0.027).This case also applies to Canada, as recent changes in the nature of work depend more doi:10.3138/cpp.2022-009on changes within occupations themselves than changes in the distribution of employment between occupations.
Furthermore, we observe that the effect of the variation between occupations is negative for all the attributes generally related to manual jobs (i.e., physical work and non-cognitive skills categories).Notably, this effect is positive for all the attributes of social interactions and verbal cognitive skills.This implies a net shift of workers from blue-collar jobs to jobs where social interactions are important.This reorganization of the Canadian workforce is consistent with Charles, Hurst, and Notowidigdo (2018), who describe the marked decline of the manufacturing sector during the first decade of the 2000 s.Additionally, Autor and Dorn (2013) hypothesize that low-skilled workers migrated to the services sector because these jobs were more difficult to automate than manufacturing jobs because they were linked to social interactions.

Gender Differences
There is every reason to believe that technological changes will have different impacts on men and women, as long as we observe segregation according to gender in occupations requiring different skills.For example, Cortes, Jaimovich, and Siu (2023) observe that since 1980, the probability of working in a well-paid occupation for men with a university education has fallen, while it is the opposite for women with an equivalent level of education.Black and Spitz-Oener (2010) observe that women in West Germany have experienced increases in non-routine analytic and non-routine tasks and a decrease in routine tasks compared with men.
To compute our indices by gender, we use census data on the total number of men and women, as well as the disaggregated number of men and women from each of three age groups (35-44, 45-54, and 55-64 y) in each occupation.
Although job attributes do not vary by gender, the aggregate indices may vary by gender in a specific year due to gender differences in occupational sorting.Moreover, over time, indices vary differently by gender through two mechanisms.First, differences in occupational sorting will lead to differences in how the indices vary by gender.Second, further differences could occur if more women move into different types of occupations than do men between the two periods. 13ables 3 and 4 show the results of the shift-share decomposition for men and women, respectively.Comparing the two tables yields four important results about how the indices vary by gender.
First, the aggregate index for most attributes increased over the period, and more for men than women.As per Table 3, the only decreasing indices for men are memorization in the category cognitive abilities and gross body coordination and manual dexterity in the category physical abilities; most of the reductions are very small.
For women, we observe declines that are concentrated in the categories physical work (2 out of 4 attributes) and  physical abilities (3 out of 4 attributes).Notably, the mechanical knowledge attribute decreases slightly for women.Overall, we observe rising aggregated indices for cognitive and social abilities for both genders.However, it is mixed for physical abilities or work, which is declining for women but, on average, stable for men.
Second, considering the total variation Δ in the aggregate index for the different O * NET attributes, we note that the average variation in attributes (in absolute value) is approximately 64 percent higher for men than women (0.136 versus 0.083), suggesting that the nature of work has changed more for men than for women.
Third, we observe that attributes related to blue-collar jobs (i.e., physical work and physical abilities) are higher for men, while attributes in cognitive, social, and knowledge categories are higher for women, for both 2006 and 2016.These three findings indicate nuances of how the labour market varies for men and women over the period.This is the case in particular for aggregate indices regarding physical work and physical abilities attributes, for which the declines are more substantial for women (Table 4) than men (Table 3), increasing the gender differences in the aggregate indices between 2006 and 2016.However, the growth of attributes related to social interactions, cognitive abilities, social abilities, and knowledge was more substantial for men, indicating that men are improving their skills in the labour market.
These findings are consistent with the hypothesis of Cortes, Jaimovich, and Siu (2023) that social skills have become more important in well-paid occupations relative to others, explaining the better performance of women in the labour market.doi:10.3138/cpp.2022-009 Finally, considering the shift-share decomposition results, we again observe significant disparities between the two genders.First, the absolute variation within occupations is almost 80 percent higher for men (0.134 versus 0.075), while the absolute variations caused by workers' movement between occupations are comparable (0.022 versus 0.022 for men and women, respectively).For both genders, most of the changes in the labour market occurred within occupations, with men facing a greater impact.

Skills Over the Lifecycle
Figures 2 to 7 show how the 24 job attributes, grouped according to the six categories outlined in Table 1, vary by age for men and women.Each figure shows the value of the index for each of three age groups (35-44, 45-54,  and 55-54 y), for men in the right panel and women in the left panel. 14Dotted lines show the cross-section values for 2006 and plain lines for 2016.Additionally, dotted arrows illustrate cohort effects, considering that a 35-44-year-old in 2006 will be aged 45-54 years in 2016.Thus, each of the arrows can be interpreted as encompassing both age and time effects on the attribute level. 15igure 2 shows the evolution of the attributes that describe the type of work.We observe no significant changes for men except for a small increase in attribute (C), pace determined by equipment.Attributes in that category changed more for women with a decrease in attribute (A), automation.We also note an increase in attribute (B), time making repetitive motions, for both genders.Comparing the three age groups, attribute (A), degree of automation, shows a significant increase for women in the older age group.
Figure 3 shows strong increases in the attributes for both men and women.Note that the equilibrium outcome for these attributes is stronger for women but increases for men, showing some convergence between genders.That is, gender lines for 2016 are closer together than gender lines for 2006.Figure 3 highlights the general importance of social interactions in a transforming labour market.
As expected, Figure 4 shows overall strong increases in attributes related to cognitive abilities except for memorization, for both genders, and deductive reasoning, for women only.
Two additional findings are worth highlighting.First, as we noted, the labour market seems to be changing more for men than for women; men are experiencing strong catch-up effects toward women.Second, the effects are similar among the six chosen attributes, as they increase in the first ten years in the job market but decrease slowly thereafter.This may be due to possible links between aging and cognitive abilities.Notably, in all cases, time effects dominate age group effects, meaning that all age categories experience increases in cognitive abilities attributes as shown by the cohort effects.
Figure 5 mirrors the previous findings shown in Figure 3.We observe changes in attributes related to social abilities.Increases are not as stark as those shown previously, but they are stronger for men than for women.Despite women requiring these attributes more, a convergence between genders is seen.Another finding is that written abilities, both expression and comprehension, increase more than oral abilities.
Figure 6 also mirrors the previous findings shown in Figure 2, showing minimal-overall decreasingchanges in the attributes.Only attribute (C), gross body coordination, shows more changes, with women experiencing a decline in this attribute.
Finally, Figure 7 shows similar graphs for four knowledge types.Similar to Figure 4 on cognitive abilities, most knowledge attributes show dramatic increases from 1996 to 2006.This is the case for administration and management, mathematics, and computer and electronics.However, in the case of mathematics, the increase is mostly seen in men.Interestingly, for men, administration and management increases continuously over the lifecycle (peaking in the 45-55 y age group), whereas other knowledge types follow the same lifecycle pattern as other cognitive abilities.Mechanical knowledge does not increase and remains much higher for men than for women.

Conclusion
Recent research on the impact of technological change on employment levels has led to the publication of numerous reports raising the possibility that millions of jobs are at risk of being replaced by robots or computers in the near future.While research has found that automation and computerization have negatively impacted the number of intensive routine jobs (Autor and Dorn 2013;  Autor, Levy, and Murnane 2003), some believe that jobs involving non-routine tasks are at higher risk of being displaced than before due to continuing technological change, including recent advances in the field of artificial intelligence.This hypothesis would be coherent with the deceleration of employment growth in abstract taskintensive occupations after 2000 (Beaudry, Green, and  Sand 2016).
In this paper, we study the variations in 24 job attributes in Canada between 2006 and 2016 by linking the occupational structure of the Canadian census with job attributes from O * NET.We decompose this variation into changes within occupations and changes from workers' movement between occupations using a shift-share decomposition as in Freeman, Ganguli, and  Handel (2020).
We find most of the variation in job attributes to be within the occupation.In the ten years under study, we interpret this finding as technological change affecting tasks within a job, rather than destroying jobs and causing movements in the occupational structure.This doi:10.3138/cpp.2022-009Third, investigating changes by gender, we notice that, while job attributes related to physical work or physical abilities decrease more on average for women than men, there is some convergence in how cognitive or social job attributes are evolving between genders.
Finally, considering the varying evolution of the job attributes over the lifecycles of men and women, we find evidence of converging job attributes between genders and some differences in their age profiles.Physical abilities and physical job attributes tend to be flat or declining, whereas cognitive abilities follow an inverse U-shaped curve.result suggests a certain skepticism toward predictions of massive upheavals in employment and underlines the importance of continuing education and training so that workers can develop the new skills required in their evolving occupations.
Second, considering the variations in individual job attributes over those years, we observe that the attributes related to manual tasks have experienced weak and sometimes negative growth in their importance in the labour market, while the importance of attributes related to social interactions has increased significantly.These findings are consistent with the decline of the manufacturing sector and the increased importance of social skills described by Deming (2017). doi:10.3138/cpp.2022-009 provided by Statistics Canada to link the 2006 (based on the 2006 NOC-S [NOC for Statistics; Statistics Canada 2007]) and 2016 occupational categories (Statistics Canada 2020).

Figure 1 :
Figure 1: The O * NET Content Model (2023) Source: The United States Department of Labor, Employment & Training Administration (USDOL/ETA).Reproduced under the CC BY 4.0 license.
3 (United States Department of Labor, Employment & Training Administration 2016) with the occupations from the 2016 census and those from the O * NET 9.3 (United States Department of Labor, Employment & Training Administration 2006) with the 2006 census.Freeman, Ganguli, and Handel (2020) also used O * NET to analyze the variations of 17 O * NET attributes from three questionnaires completed by job holders for the categories work contexts, education and training, and knowledge between 2005 and 2015.The attributes analyzed in this work build on those analyzed by pation from the 2006 census (classified according to the 2006 NOC-S) are not directly comparable to those from the 2016 census (classified according to the 2016 NOC).To do this, we use the concordance between the 2006 NOC-S and the 2011 NOC (Statistics Canada 2020), as the 2016 NOC structure is the same as the 2011 NOC. 9 With the occupational structure data from the 2006 and 2016 censuses classified according to the 2016 NOC, we match it to the O * NET data.The O * NET's occupational classification, O * NET-SOC, is revised on an irregular basis. 10O * NET version 10.0 uses the 2006 O * NET-SOC, while O * NET version 20.3 uses the 2010 O * NET-SOC.Therefore, correspondence is necessary between the 2016 NOC and these two versions of the O * NET-SOC to associate O * NET 10.0 values with the 2006 census data and O * NET 20.3 values with the 2016 census data.Accordingly, we first find the correspondence between the 2016 NOC and 2010 O * NET-SOC using the existing bridge from the Brookfield Institute for Innovation + Entrepreneurship (2018).This correspondence shows an association of at least one occupation from the 2010 O * NET-SOC with each occupation from the 2016 NOC.We then use the correspondence available from the O * NET to go from O * NET-SOC 2010 to O * NET-SOC 2009 then from O * NET-SOC 2009 to O * NET-SOC 2006.This exercise results in a correspondence matrix with 500 occupations from the 2016 NOC and the corresponding occupations from the 2010 and 2006 O * NET-SOCs.
possible to carry out a shift-share decomposition by breaking down the variation of this aggregate index (A t ) between 2006 and 2016 into three distinct terms according to:

Figure 2 :
Figure 2: Physical Work; (A) Degree of Automation, (B) Time Making Repetitive Motions, (C) Pace Determined by Equipment, (D) Time Bending or Twisting

Figure 3 :
Figure 3: Social Interactions; (A) Coordinate or Lead Other, (B) Contact with Other

Table 1 :
Selected O * NET Job Attributes Second, sometimes, a NOC occupation corresponds to a 2010 O * NET-SOC occupation that does not exist in the 2006 O * NET-SOC.These occupations are excluded from our analysis.