1.1 Skills and occupations

The task-based approach has deepened our understanding of labour markets. The impacts of technological change, innovation, outsourcing and business cycles can be quite distinct according to workers’ skills. The task content and skill requirements are also related to the disappearance and emergence of occupations, and employment growth. Due to the heterogeneity across skills, workers with the same formal education can have completely different trajectories during business cycles.

Autor, Levy and Murnane (2003) is the starting point of the contemporary analysis of occupations as a bundle of tasks. In their simple model, “computer capital” would substitute routine tasks (manual or cognitive) and be complementary to non-routine tasks. This hypothesis became known as RBTC. The testable predictions of the model are that an exogenous reduction in the price of computer capital would reduce the demand for labour in routine tasks and increase the demand for labour in non-routine tasks. The analysis of the changes in the tasks of each occupation showed that they did indeed shift in the directions predicted by the model. Autor, Levy and Murnane (2003) also introduce the classification of tasks in a two-by-two matrix: routine vs. non-routine and manual vs. cognitive. In this study, we will follow their typology. Spitz-Oener (2006) backs up the predictions by Autor, Levy and Murnane, (2003) and shows that analyses focused on occupations conceal changes in tasks within each occupation over time. Using keywords to classify tasks, she found similar results to the United States for Germany since 1979, namely there was a general increase in occupational complexity and the demand for analytical and interactive activities has gained importance in the mix of tasks performed by the workers.

Based on these insights and further evidence on the labour market dynamics in the United States, Acemoglu and Autor (2011) developed a model that became the workhorse of task-based empirical analysis by clarifying the concepts of tasks, skills and occupations.2 In a nutshell, performed tasks generate products, and occupations are composed of tasks. Skills are the endowments workers require to perform tasks. Both the task composition of occupations and the skill endowments of workers are endogenous and can be changed over time. This framework allows the allocation of tasks by skill to be determined by technology and relative prices. The distribution of tasks across skills occurs in a Ricardian way. The allocation takes place according to the comparative advantages of the skills in performing the tasks. However, the cut-off points for assigning tasks to skills are determined endogenously in the model. Thus, the model’s main contribution is to facilitate studies of the less obvious consequences of price changes, endowments and production technologies. Acemoglu and Autor (2011) also update Autor, Levy and Murnane’s (2003) skill categories to the more recent O*NET database.

The task-based approach has proven valuable for understanding the current phenomenon of disappearing middle-skill occupations in developed countries. Recent technologies, such as computers and robots, have thus far substituted routine tasks, as the RBTC predicts. In the longer term, machines will not be limited to simple tasks, advancing over more complex, analytical and creative ones. 3 As a result, Acemoglu and Restrepo (2017) point out that advances in artificial intelligence and robotization may further increase inequality and polarization.

As long as routine tasks performed by high-skilled workers are substituted, their productivity increases. As a consequence, the remaining workers will be more valuable to firms and should observe growing returns to non-routine skills augmented by capital. At the same time, the supply of dislocated medium-skilled workers rises, lowering the relative wages of routine task-intensive occupations. Finally, low-skilled workers perform NRM labour market polarization: wages and employment increase at two extremes of the skill distribution. 4

To date, the evidence on job (and wages) polarization in Brazil is mixed. Maloney and Molina (2019) study the impact of technological change in several middle-income countries, including Brazil, and observed some signs of polarization. Yet Ariza and Bara (2020) report that the employment change curve over the wage distribution is flat. According to Firpo and Portella (2019), other factors must have contributed to offset polarization trends. Ferreira, Leite and Litchfield (2008) estimate that the main factor responsible for the fall in wage inequality was the decline in return for experience. Firpo and Portella (2019) suggest that age-biased technical change or skill obsolescence explain this fact. This means that the skills of older people cannot keep up with the new tasks that arise. Almeida, Corseuil and Poole (2017) analyse the impact of computers and the internet on the Brazilian labour market and find evidence in favour of polarization. Herdeiro, Menezes-Filho and Komatsu (2019) identify a more complex temporal dynamic of polarization, as technological advances favoured middle-skilled workers between 1981 and the late 1990s, However, between 2004 and 2015, this trend was reversed and the relative demand for middle-skilled workers began to decline.

In Brazil, some task-based studies exist, however with different focus and skill measurement methods. For example, Ehrl and Monasterio (2019) and (2021) study the (long-term) effects of skill concentration in local labour markets. Neves Jr., Azzoni and Chagas (2017) estimate the skill premium by city size. Albuquerque et al. (2019) and Maciente, Rauen and Kubota (2019) predict the impact of automation on the Brazilian labour market, and Adamczyk, Monasterio, and Fochezatto (2021) on the public sector. Ehrl (2018) analyses the heterogeneous effects of intermediate goods importing on the skill structure within firms. Arellano-Bover (2020) explores how on-the-job learning depends on workers’ skills. These papers either apply Maciente’s (2013) mapping between O*NET and Brazilian Occupational Classification (Classificação Brasileira de Ocupações – CBO) or define discrete skill measures using occupation groups.

1.2 Career changes, business cycles and skills

Automation, RBTC and other forces cause changes in the task set of occupations prompting some individuals to switch occupations. As is well known, these changes do not always occur without frictions. Recent literature uses individual data and the task-based approach to understand transitions between jobs and their impact on wages and the length of unemployment, among others.

Gathmann and Schönberg (2010) build a model of employment switches that takes into account task-specific human capital, also using Autor, Levy and Murnane’s (2003) classification. They tested it on German administrative data at four points in time (1979, 1985, 1991/92, and 1998/99) and found that 40 per cent of wage growth stems from the accumulation of human capital within the same set of tasks. Moreover, they identify that workers have larger wage reductions when they move to occupations with tasks that are more distinct from the ones they previously held. Bachmann, Cim and Green (2019) also examine long-term trends in the German labour market between 1974 and 2018, and identify an accelerating polarization tendency that is more relevant than short-term effects stemming from economic cycles. Again, workers in routine occupations are most likely to lose their jobs and struggle to get out of unemployment. Worse still, these problems of routine workers have become more severe over the decades.

Regarding the United States labour market, Cortes (2016) observes similar results to the papers with German data. The author highlights that workers who migrated from RC to NRC occupations reap gains, while those who moved from RM to NRM occupations are worse off than those who remained in the same occupation class. Likewise, Ross (2017) identifies that a fall in routine tasks in the same occupation leads to falling wages, while an increase in abstract tasks affects wages upwards. Furthermore, the returns from routine tasks fell while abstract tasks rose over time, lending support to the RBTC hypothesis.

Arellano-Bover (2020) show that young OECD workers who faced recessions, despite an education upgrading effect from postponing labour market entry, still had shortfalls in their skills two decades later. And, among these, the ones who suffered the most were those with parents with little schooling. The enduring persistence of these adverse effects lends even more relevance to the issues examined in this study. Additionally, Lazear and Spletzer (2012) and Carrillo-Tudela et al. (2016) show that shifts between occupations are pro-cyclical. When the economy is growing, turnover increases because workers migrate to higher-wage occupations. During a crisis, the gains from occupational changes are lower, and employees tend to avoid change and wait for the economy to recover.

The task-based approach is also helpful to understand the phenomenon of jobless recoveries. As early as 2003, Bernanke (2003) argued that during economic recoveries, the mismatch between workers’ skills and available jobs would increase. The current consensus is that those engaged in routine tasks once again tend to lag, accelerating the job polarization process in developed countries (Jaimovich and Siu 2020; Hershbein and Kahn 2018; Graetz and Michaels 2017). However, more research is needed as Graetz and Michaels (2017) find that outside the United States, modern technology is not responsible for jobless recoveries.

Wiczer (2015) presents a search and matching model that takes into account job-specific skills. In his model, wage loss for workers who change occupation is proportional to the skill-distance between occupations. This causes workers to seek employment in occupations similar to those they held and employers to seek workers in the same occupation. This mechanism creates a long tail of unemployment duration that increases the long-term unemployment rate. Wiczer (2015) then estimates the skill-distance between occupations in the United States. His calibrated model explains more than two thirds of long-term unemployment between 1995 and 2013.

As far as we know, in-depth studies on occupational transitions in middle-income countries are rare. Parra and Christian (2021) analyse workers’ jobs and occupational switches in Vietnam. Their study is one of the few papers that took transitions between formal and informal labour markets into account.

2.1 Data

The main data source for this research is the Annual Social Information Report (Relação Anual de Informações Sociais - RAIS) produced by the Brazilian Ministry of Economics. RAIS is an administrative register of employer-employee data covering the population of formally employed workers in the public and private sectors. It is widely recognized as the most reliable source of information about the formal labour market in Brazil (Dix-Carneiro 2014). We are able to draw a 10 per cent random sample from the population of workers based on the last digit of their PIS (Programa de Integração Social) identification number. This unique number makes it possible to follow workers over time. Our analysis focuses on male and female employees between 20 and 60 years of age with complete information in the selected variables. 5 Our sample and the present estimates are therefore representative of the entire formal Brazilian labour market.

The informal sector represents almost half of the labour force in Brazil. Of course, it would be ideal if our data also included informal workers, but studies have shown that the informal sector in Brazil acts as a buffer for the unemployed in face of adverse labour demand shocks (Dix-Carneiro and Kovak 2019). Like unemployment rates, informality is counter-cyclical and the transition rates from unemployment to informal jobs are twice as high as those from unemployment to formal jobs (Ulyssea 2020; Dix-Carneiro et al. 2021). In other words, the boundaries between unemployment and the informal sector are quite fluid and more comparable to each other than to formal jobs in the highly regulated formal labour market. Thus, this study deals with the internal dynamics of workers in the formal labour market, and its inflows and outflows, irrespective of whether they are unemployed or working in informal activities.

The data are organized in the form of employment spells. That is, the number of observations per worker-year corresponds to the number of non-overlapping formal employment relations. Labour market entry and exit are permitted at any time during the longest possible observation period from 2003 to 2019. The present analysis takes into account workers’ wages, the number of hours worked per week, their occupation, tenure on the job, education level and their employers’ size, sector, legal nature and geographical location. Since we are mainly interested in how workers’ characteristics relate to employment transitions, the last year is only utilized to calculate change variables (such as occupation change, wage growth and employment distance). The final full sample from 2003 to 2018 contains 50,082,074 observations from 6,358,348 individuals. As usual, the administrative data do not provide information about why an individual leaves the formal labour market. There is, however, a variable that indicates why the labour contract ended, which allows us to calculate the share of retired or deceased workers. Besides, left- and right-censoring of employment trajectories require careful treatment. On the one hand, duration analysis (for the probability of leaving the formal labour market) is adequate for this situation. On the other, to analyse occupation changes, we follow the convention in the literature and restrict the sample to workers with at least two consecutive employment spells. See, for example, Bachmann, Cim and Green (2019) or Cortes et al. (2020). In this consecutive employment sample, we end up with 46,923,670 observations from 5,433,062 individuals.

A key feature of this paper is the derivation of skill measures that are unique and specific to Brazil. To this end, we combine information from the well-known United States O*NET based skill measures with the detailed task descriptions in the Brazilian Occupational Classification (CBO) from the year 2002. The CBO describes the occupations and organizes them in a hierarchy, making it possible to systematize information regarding the workforce according to the occupational characteristics and the nature and content of the work involved. It describes the functions, duties and tasks that make up each occupation, and the content of the work in terms of the set of knowledge, skills and training required for the performance of responsibilities for each occupation (CONCLA, 2019). The details of our skill derivation procedure are explained in Section 2.3.

2.2 Occupational distance measure

In order to find a continuous metric for the distance between a pair of occupations, we proceed in two steps. Firstly, we apply Natural Language Processing (NLP) algorithms to find weights that express the relevance of each word in the Activities Matrix 6 which describes the tasks inherent to each occupation. Secondly, we chose a distance measure and use the weights from the previous step to derive the occupational distance matrix. Previous studies, such as Gathmann and Schönberg (2010), use a distance metric based on the absolute difference between analytical tasks scores. Instead, we proceed by comparing directly the weights of words that describe each occupation’s tasks.

The Term Frequency–Inverse Document Frequency (TF-IDF) method calculates the relative weights of words describing the tasks. The idea of the TF-IDF is that the derived weights increase proportionally to the frequency of occurrence of a word in a given text extract in relation to the number of texts that contain the word. Following NLP jargon, corpus is a collection of documents, document is a text object, and text object is a sequence of words. In this subsection, corpus is the set of CBO activities descriptions and the documents are these descriptions. Given the set of occupations’ activities $D$, a word $w$ in the activity description, $d?D$, the calculation of the corresponding weight ($weigh{t}_{w,d}$) according to the TF-IDF method is given by

where $f_{w,d}$ is the relative frequency of $w$ in $d$, defined as the number of occurrences of each word relative to the total number of words in the activity description $d$. 7 $\left|D\right|$ is the total number of activities descriptions and $f_{w,D}$ is the number of activities in $D$ in which $w$ appears. High values of $weigh{t}_{w,d}$ imply that $w$ is an important word in $d$ (first term), but not very common in $D$ (second term) in equation (1). In that sense, high TF-IDF weights signal that the word $w$ serves well to discriminate between the task sets of occupations.

A limitation of the TF-IDF method is the incapability to identify the grammatical variations of the same word, especially regarding verbs (Qaiser and Ali 2018). Spellings such as “analysed”, “analyse”, or ”analysing” are treated as being different words. In order to avoid this issue, a stemming procedure is carried out before applying the TF-IDF method. Therefore, inflections of the same term are converted to a single common root. In addition, stop-word removal discards phrase-connecting words, such as conjunctions, numbers and special characters that have no semantic value for the analyses.

The resulting vector, ${\mathcal{W}}_d=\left(weigh{t}_{1,d}\dots weigh{t}_{W,d}\right)$, collects the TF-IDF weights of all words from 1 to $W$, as $W$ being the number of different words that appear in the Activities Matrix. The matrix $\mathcal{W}=\left(\mathcal{W}{\mathrm{\text{'}}}_1\dots \mathcal{W}{\mathrm{\text{'}}}_D\right)$ ends up with 2,322 words describing 2,641 occupations. This matrix is 99 per cent sparse since most of the terms are not shared between occupations, which is common in TF-IDF analysis.

We apply the cosine similarity measure to derive distances between occupations as vectors in high dimensional sparse matrices. The cosine similarity computes the angle between two vectors, that is the weights of words that compose the activities of two given occupations. Because the cosine measure is insensitive to the absolute length of each document, it works well for sparse matrices. 8 Based on Aggarwal (2015), the cosine similarity measures for any pair of occupations $i$ and $j$ is defined as

Therefore, we have a symmetric matrix reflecting the bilateral similarity between occupations. $Co{s}_{ij}$ is a continuous measure in the interval $\left[0,1\right]$, where occupations with value 1 are identical to each other in terms of tasks, and occupations with 0 do not share words describing their tasks. Finally, to have a direct measure for the distance between occupations, we use $OccDis{t}_{ij}?\mathrm{l}\mathrm{o}\mathrm{g}\left(1/Co{s}_{ij}\right)$ in the following analysis.

2.3 Skill measures

Our aim in this section is to derive a novel methodology to generate skill measures that (a) are highly comparable across countries, (b) have modest data requirements, (c) do not rely on cross-country occupation matching, and (d) replicate an international standard.

Our method of skill measurement has three advantages over the previous ones. The first advantage is that we do not need ad hoc lists of words related to each type of task. Second, the method does not assume that the occupations perform the same activities across countries. Third, our method can measure the skills even of occupations that are not listed in a database such as O*NET. A large number of Brazilians work in occupations that cannot be found in O*NET, such as elevator operators in commercial buildings (CBO 5141-05 - Ascensorista) or bus ticket/money collectors (CBO 5112-15 - Cobrador).

We have chosen to follow the skills categories introduced by Autor, Levy and Murnane (2003) which were then adapted to the O*NET data by Acemoglu and Autor (2011). This will be the only relevant ad hoc decision in the following methodology. In order to replicate the United States data-based skills, we use the O*NET as an initial stepping stone. The remainder of the procedure relies exclusively on the Brazilian CBO data. The derivation of skill measures is carried out in five steps:

• • 1. Build a dictionary of verbs from the O*NET that are associated with occupations with the highest skills in each of the groups (NRC, RC, NRM, and NRM);

    2. Translate these verbs to Portuguese by means of a deep learning translation service;

    3. Derive weights of these Portuguese verbs by means of an NLP technique;

    4. Apply these weights by skill to the text tasks descriptions of the CBO occupations;

    5. Normalize skill measures.

Our choice skill measures are the four categories from Acemoglu and Autor (2011), namely: NRC, RC, NRM and RM. 9 Having calculated the original task measures of Acemoglu and Autor (2011), we rank occupations by their skill score and select the top 10 per cent of occupations. Repeating this procedure for all four skill categories defines the selected occupations’ elements whose words are then translated into Portuguese. This resulted in around 6,100 tasks.

The conversion of the task file from English to Brazilian Portuguese is done with DeepL Pro, a neural machine learning translation service. 10 The task descriptions contain information about contexts that would be undesirable for our purposes. For example: “clean a car” and “design a car” contain verbs associated with different skills, but the noun “car” would have the same weight as the verbs and would harm the classification. Therefore, we prefer to extract only the verbs from the task descriptions.

In this subsection, we use two corpora: verbs that are associated with skills in O*NET and the verbs in the activities descriptions by occupation in CBO. From the 96,900 words in the translated task descriptions of O*NET, 996 verbs have correspondence in the CBO task descriptions. The matrix $V$ contains the weights (as defined below) for the 996 verbs (columns) in the four skill categories $i?$ {NRC, RC, NRM, RM} (rows). Our final skill scores for each occupation are the result of the following matrix product:

where the matrix $Occup$ is defined by the raw count of the 996 verbs (rows) in the 2,478 CBO occupations (columns). 11

The usual way to identify the weights of terms in a document for $V$ is through the TF-IDF method (see section 2.2). Here, however, TF-IDF has not turned out to be the best option. The reason is simple: if a verb is present in all four skill categories, the value of the IDF is 0, and obviously TF-IDF is equal to 0. 12 For example, if a given verb$w_v$ appears 100 times in total with the following distribution: 90 times in the RM skill category, eight times in NRM, and once both in RC and NRC, the TF-IDF value of the verb $w_v$ will be zero. The same value would be obtained if $w_v$ appeared with equal frequency in the four skills. Thus, relevant information about the distribution of the verb $w_v$ would be lost.

2.4 Model

2.4.1 Transition regressions and testable hypotheses

The research question in this paper is: how are skills related to labour market transitions in Brazil? In particular, the following hypotheses will be tested in our regressions and duration analysis:

• [H1] Workers’ previous skill endowments affect the probability of transitioning to a different occupation.

• [H2] Workers’ previous skill endowments affect the probability of transitioning to a different sector.

• [H3] Previous skill endowments affect the distance between occupations in case of an occupation change.

• [H4] Previous skill endowments affect the wage differential when workers move to a different occupation.

• [H5] Previous skill endowments affect the probability of a worker experiencing a period out of the formal labour market.

• [H6] The size of these marginal effects depends on workers’ age, tenure, employer size and the business cycle.

The dependent variables that allow us to assess hypotheses H1 to H6 are defined according to the difference in the employment outcome of worker $i$ in the current period $t$ and the sequential employment spell (which must not necessarily be in year $t+1$). We use the following variables:

• $?Occ4$ is an indicator variable for whether we observe that a worker has a different 4-digit occupation in the following employment spell.

• $?Occ2$ is defined similarly to the previous definition but based on 2-digit occupations.

• $?Occ4?$ is an indicator variable for whether we observe that a worker moves to a different 4-digit occupation with a higher average wage. Average wages are defined as the mean log wage for a given year and 4-digit occupation based on the observations in our sample.

• $?Occ4?$ is an indicator variable for whether we observe that a worker moves to a different 4-digit occupation with a lower average wage.

• $?Sect$ indicates whether a worker will move to a different sector.

• $\mathrm{l}\mathrm{o}\mathrm{g}\left(OccDist\right)$ is the log of the inverse of the occupational distance defined in equation (2).

• $?Wage$ is the difference between the log wage in the next and current employment spell.

• $OFLM$ is an indicator variable for whether a worker experiences a period of at least one month out of the formal labour market between the current and the next employment spell.

Our baseline employment transition regressions take the following form

where the vector of control variables $X_{it}$ includes worker $i$’s number of working hours, level of education, tenure, gender, age and firm size. $D_s$, $D_r$, and $D_t$ are dummies for sector, federal state and year, respectively. The coefficients of interest ${?}_k$ then indicate how the current skill levels are linked to employment transitions of observably equivalent workers in the near future.

In line with the definitions above, we use NRC, NRM, and RM skills in the regressions, while RC skill is the omitted reference category. In case the dependent variable is an indicator, we apply logit regressions; otherwise, equation (4) is estimated with pooled OLS. Additionally, we also include interaction terms between the worker’s skill level and moderator variables such as worker $i$’s age, tenure, employer size, GDP growth and residence in the federal state. Standard errors are clustered at the occupation level because the explanatory variables of interest - workers’ skills - vary only across occupations. This procedure yields more conservative results than clustering at the worker level.

2.4.2 Duration analysis

The most appropriate way to analyse the probability of leaving the formal labour market is via duration analysis. This type of approach estimates the probability of leaving, given that this event has not occurred before, conditional on the set of observable variables (Lancaster 1979). The exit probability at time $t$ upon survival to time $t$ is known as the hazard rate. If the duration time $t$ follows a Weibull distribution, which is the most frequently used in the literature, the hazard rate $?$ takes the following form

where $Z_{it}$ is composed of the same control variables defined in equation (4) and $p$ is a parameter that indicates whether the exit probability has an increasing, decreasing or constant form. The parameters of the duration model are obtained by maximum-likelihood estimation, where standard errors are also clustered at the occupation level.

According to the nature of our worker panel, the survival data includes multiple records per subject with continuous-time measurement and the possibility of multiple exits. The exit event is defined as leaving the formal labour market for at least one month. Because we cannot distinguish whether the individual becomes unemployed, self-employed, or gets a job out of the formal labour market, we consider only exits for individuals that eventually return to the formal labour market. If an individual does not return to the formal labour market, we have no way of knowing the direction of his or her transition into a lower or higher wage, or distant occupation.

3.1 Descriptive statistics

3.1.1 Skill scores and workers’ characteristics

The descriptive statistics section presents aggregate trends in the Brazilian labour market that contribute to the understanding of the subsequent analysis of employment transitions at the worker level. The following results are derived from the full sample of the RAIS data from 2003 to 2018.

First, it is crucial to check whether our novel skill measures produce adequate results. The continuous skill measures’ mean values in the most aggregated occupation groups are certainly plausible. There are nine groups, excluding police and military professionals, and taking into account both groups of industrial workers (discrete and continuous processes) in CBO.

Table 1 shows that the groups with the highest wages - professionals, managers and technicians - present the highest scores in NRC skills. These occupations are responsible for abstract and highly specialized tasks, intensive in hard and soft skills since most of the work is done by interacting with other people. Managers and technicians have the highest scores in RC skills as well. Besides, administrative and service occupations mostly require RC and, to a lesser extent, NRC skills. As expected, the highest scores on NRM skills belong to workers in maintenance, service and manufacturing 2 (continuous processes). Manufacturing 1 (discrete processes) and agricultural workers score highest in RM skills as those workers mainly perform repetitive tasks in controlled environments, with little or no necessity for improvisation.

Table 1. Skill scores by occupations.

Note: the cells show the mean and standard deviation (in parenthesis) for the skill measures (non-routine cognitive (NRC), routine cognitive (RC), non-routine manual (NRM) and routine manual (RM)) by 1-digit occupations over the entire observation period 2003–2018. Calculations are based on the full sample and apply weights for employment duration and hours worked. The 1-digit occupations are ranked according to their mean wage in descending order, namely: directors and managers; science and arts professionals; medium-level technicians; administrative service workers; workers in the production of industrial goods and services (1), workers in repair and maintenance; workers in the production of industrial goods and services (2); service workers and sellers; and agricultural workers.

Source: elaborated by the authors.

Figure 1 provides another way to compare the plausibility of our skill measures to previous findings from other countries. Following Autor, Katz and Kearney (2008), the figure shows how the skill scores vary across the wage distribution in 2018. Additional information for previous years is provided in the Appendix (see Figure B1). RM and NRM intensive occupations dominate at the lower end of the wage distribution, exhibiting a negative and decreasing slope over the wage distribution. NRC skills show a mirror-inverted pattern, beginning at a low level and rising at an accelerated slope over the upper half of the wage distribution. RC skills are most frequently used in occupations between the 20${?}^{th}$ and the 80${?}^{th}$ percentile. While the RC skill intensity is also increasing in wages, it is almost flat for the most part in the middle of the distribution. Regarding the evolution over time, Figure B1in the Appendix indicates that skill intensities of high- and middle-wage workers remained quite constant. The most pronounced change is that low-wage employment now uses fewer NRC skills, whereas NRM skills are becoming more important. There is also a clear tendency of middle-wage workers to increasingly use NRC skills, in line with the RBTC hypothesis.

Figure 1. Skill scores over the wage distribution in 2018.

Source: elaborated by the authors.

It is informative that the skill intensities over the wage distribution are similar to the United States experience. Autor, Katz and Kearney (2008), for example, report that NRC skills are monotonically rising, NRM skills are markedly decreasing, while RC skills are most frequent in middle-wage occupations. These patterns are as expected and are not related to taking the United States task descriptions in the O*NET as a starting point for our method. The observed patterns in Figure 1 reflect that NRC skills require a large investment in human capital (such as a university degree). Consequently, the return to these skills is highest and occupations with high NRC requirements are located at the top of the wage distribution.

The pattern in Figure 1 and the order of relative wages are in line with other studies of the Brazilian labour market. Controlling for observed and unobserved characteristics in a fixed-effects model, Detoni, Freguglia and Corseuil (2020) find higher returns for NRC intensive occupations, followed by RC, RM, and NRM skills. Sulzbach (2020) shows the increasing value of tasks related to cognitive skills since 2003, while routine skills saw a decreasing wage compensation. This evidence is in line with RBTC, that is advancing automation substitutes previous routine tasks performed by humans, and therefore degrades their wages (Acemoglu and Autor 2011).

In order to evaluate how the socioeconomic characteristics of workers relate to their skills, we define discrete skill groups following Goos and Manning (2007) and Cortes (2016), among many others. These categories are listed in Data Appendix 1. Table 2 shows the descriptive statistics by comparing occupational and socioeconomic characteristics of each skill group between 2003 and 2018. 14 NRC workers receive higher wages, followed by RC, RM, and NRM workers. This ordering remains stable through 2018, but all occupational groups experienced real gains over our observation period. In particular, NRM and RM occupations gained 33.8 per cent and 29.9 per cent higher real wages than in 2003. This catching up may not be a sign of high demand but rather be influenced by the important adjustments of the minimum wage, which grew by 76.5 per cent in real terms since 2003.

Regarding gender, men dominate the RM occupations in both years, scoring 83 per cent on average. On the other hand, women increased their participation in the three remaining categories and are predominant in NRC, RC, and NRM occupations. The most relevant effect of the increased female labour market participation can be seen in NRM jobs, which had 58 per cent of its positions occupied by men in 2003, decreasing to 47 per cent in 2018. These NRM occupations encompass service and cultural services workers, technicians in transport services and agricultural producers.

Table 2. Descriptive statistics by skill groups: 2003–2018.

Note: the cells show the mean and standard deviation (in parenthesis) for selected employment transition indicators in 2003 (columns (1) to (4)) and in 2018 (columns (5) to (8)) for each of the discrete skill groups: non-routine cognitive (NRC), routine cognitive (RC), non-routine manual (NRM) and routine manual (RM). Inflation is adjusted by the official consumer price index (IPCA) to 2018 prices. Calculations apply weights for employment skill duration and hours worked and are based on the consecutive employment sample.

Source: elaborated by the authors.

Likewise, the evolution of workers’ age reflects, at least partly, the demographic transition in Brazil (IBGE 2018). All skill groups exhibit a higher average age in 2018 in comparison to 2003. Additionally, the most pronounced increase in average age occurs between 2013 and 2018, suggesting that firms primarily dismissed younger, less experienced workers during the economic recession. Workers in non-routine, cognitive-intensive occupations have longer employment tenure, increasing from 87.3 months in 2003 to 92.3 in 2018. RM workers are most prone to employment changes, that is to say that they present the lowest average tenure. Analogous to age, tenure shows a crescent trend that has accelerated since 2013 for all skill groups.

The variable “occupation change” in Table 2 shows the fraction of workers that moved into a different 4-digit occupation. 15 We can see that occupational changes became less frequent over the years. In 2003, 17 per cent of NRC workers transitioned to another occupation, whereas only 10 per cent did in 2018. The same pattern is seen for RM, RC, and NRM workers. The wide definition that considers occupation group changes based on a higher level of aggregation (the 2-digit CBO classification) shows the same declining pattern. In both cases, the change occurs gradually over time and does not seem to be influenced by the business cycle. The sector change variable presents a stable pattern over time, ranging from 5 per cent for NRC workers to 10 per cent for the RM skill group. The log occupational distance suggests that, without accounting for other observable characteristics, RM workers move the farthest away from the previous occupation. The mean occupational distance in 2003 for RM workers was 2.49, increasing slightly to 2.51 in 2018. NRM and RC workers decrease their mean occupational distance. NRC workers present a stable, relatively low occupational distance.

3.1.2 Employment changes and transitions

Figure 2 shows the employment change by occupational groups between 2003 and 2018, in total, and for three sub-periods. Over the last two decades, the largest relative change has been observed for the professional group, with a growth of 51 per cent relative to 2003, followed by 44 per cent growth for managers and 35 per cent in service occupations. In general, all aggregated groups increased the number of formal employees, in line with the 34 per cent increase of our 10 per cent random sample, which grew from 1.8 million in 2003 to 2.4 million full-time equivalent workers in 2018.

The three graphs for the subperiods in Figure 2 show that the Brazilian economy experienced a period of rapid GDP expansion between 2003 and 2009, with a mean growth rate of 4.2 per cent. The 2008 financial crisis did little harm to the economy, as GDP growth decelerated only slightly to a mean of 3.3 per cent between 2008 and 2013, before the economy plunged into a severe recession from 2014 onward (IBGE 2019b). The business cycle helps to explain the patterns of employment change shown in Figure 2. Between 2013 and 2018, occupations in manufacturing of continuous processes (manufacturing 2) suffered the strongest retraction of 30 per cent in employment, followed by maintenance workers and manufacturing of discrete processes (manufacturing 1). The recession was severe in sectors such as construction and transformation, falling to 85 per cent and 70 per cent of their production level in 2014, respectively (IFI 2018). Between 2003 and 2013, in contrast, all 1-digit occupation groups grew.

Concerning the whole period, agriculture seems to be an outlier, and in fact this sector is usually excluded. Nevertheless, agriculture is an important component of the Brazilian economy and we aim to be as inclusive as possible in the present analysis. The relative stability of agricultural employment can be explained by the fact that Brazil is among the most productive and competitive agro-exporters worldwide thanks to state-driven innovation, large properties and continuous expansion of its agricultural frontier over the last decades that has made land available for commercial agriculture (Hopewell 2016). Interestingly, back in 1970, Brazil was a net importer of agricultural goods. Agriculture is directly responsible for (only) 5.3 per cent of GDP but employs 15 per cent of the country’s labour force (IBGE, 2019a, 2019b).

Figure 2. Employment changes by occupation group, 2003-2018.

Note: the 1-digit occupations are ranked according to their mean hourly wage in 2003. See Appendix Table 1 for the precise definition of the occupation groups. Calculations are based on the consecutive employment sample and apply weights for employment duration and hours worked.

Source: elaborated by the authors.

Jobs in agriculture and services (such as food service workers, security guards, cleaners, hotel and housekeepers and sales attendants) are often referred to as “low-skill” occupations; manufacturing, maintenance, and administrative workers as “medium-skill” occupations; and managers, professionals and technicians as “high-skill” workers 16 (Autor 2019). Using these categories, Figure 2 shows that low-skill occupations grew on average between 2003 and 2013 and fell slightly over the period 2013-2018. The high-skill occupations followed the same pattern but grew at higher rates. Middle-skill occupations, in contrast, present a pattern of steadily declining growth, culminating in a heavy contraction of up to -20 per cent in 2013-2018. In sum, Figure 2 shows the first sign of employment polarization with low-skill and high-skill occupations growing relatively more than medium-skill occupations.

Figure 3 describes the changes in employment shares along with the wage distribution over the last two decades, following the seminal work of Autor, Katz and Kearney (2008). The total change and the breakdown for three sub-periods allow us to analyse whether Brazil experienced employment polarization similar to developed countries. Only the most recent period, 2013 to 2018, shows employment growth in occupations at the top and bottom of the wage distribution and a retraction in medium-income occupations. This is additional evidence of wage polarization since the left and right tails of the wage distribution are increasing relative to its middle.

Over the last observation period (2013-2018), the Brazilian formal labour market resembles the change in the United States labour market between 1990 and 2000. In both countries, employment losses are most pronounced between the 40${?}^{th}$ and 60${?}^{th}$ wage percentile. The magnitude of the most extreme aggregate changes in a single percentile in Brazil, range from +8 per cent to -3.5 per cent, while those reported by Autor, Katz and Kearney (2008) over a period twice as long are between -13 per cent and +16 per cent.

Figure 3. Employment changes over the wage distribution 2003-2018

Note: The figure shows the employment changes in 4-digit occupations ranked according to their 2003 mean wage using a locally weighted smoothing for the entire observation period, as well as three different subperiods, as indicated in the above four graphs. Wages are deflated to 2018 prices using the IPCA. Calculations are based on the entire RAIS sample and apply weights for employment skill duration and hours worked.

Source: elaborated by the authors.

The polarization pattern is not seen when we consider other time intervals, like 2003 to 2008, 2008 to 2013, or the full period. Although high-skilled employment is continually expanding, the low-skilled occupations experienced negative growth rates and changes in the middle of the skill distribution were close to zero. Jaimovich and Siu (2020) and Hershbein and Kahn (2018) show that in the United States, routine-biased technical change and jobless recoveries go hand in hand. Increasing automation of routine and manual tasks can accelerate these trends. That is, Brazilian firms may have used the recent recession as an “opportunity” to displace and substitute middle-skill workers. It remains to be seen whether the employment polarization in Brazil is a transitional phenomenon related to the economic downturn or the beginning of a secular trend that arrived later than in developed countries.

Lastly, we can look at transitions between occupations in different skill groups, as well as transitions into or out of the labour market and into retirement. Table 3 shows a transition matrix based on the discrete skill classification. The share of workers who remained in the same classification in 2003 and 2018 range from 27 per cent for NRM to 34.2 per cent in RM occupations. In general, routine skilled workers are less likely to move into another occupational group. Transitions from manual to cognitive occupations and vice-versa are least common.

Table 3. Transition matrix between skill groups 2003-2018

Note: the cells show the percentage of workers from a given skill category (non-routine cognitive (NRC), routine cognitive (RC), non-routine manual (NRM), and routine manual (RM) in 2003 (column 1) that are either out of the formal labour market (OFLM), employed in the same or in another skill group (row 1) in 2018, or retired in or before 2018. Per definition, the row sum is equal to 1. Calculations are based on the entire sample and apply weights for employment skill duration and hours worked.

Source: elaborated by the authors.

Although the observation period spans 16 years, the share of transitions out of the formal labour market is strikingly high. About 46 per cent of the NRC workers in 2003 were outside the formal labour market in 2018. Another 8.9 per cent of all NRC workers observed in 2003 moved into retirement by or before 2018. So overall, the effects of aging, self-employment (in the growing gig-economy) and increasing unemployment, among others, contribute to the fact that 55 per cent of all NRC workers are not in the formal labour market 16 years later. The high turnover is only slightly affected by transition periods within a calendar year, because the RAIS data allow us to observe the individual over the entire year and not only at a specific date.

The differences in the job-to-retirement rates are also remarkable. Among the RM skilled, only 3.3 per cent moved from formal employment directly into retirement. For RM and NRM workers, the rate is higher (5.1 per cent) but still much below that of the most skilled NRC workers. Two more observations from Table 3 confirm that the RM workers are the most vulnerable group. Their transition rate out of the formal labour market is highest at 49.4 per cent and, despite being the largest group in terms of the aggregate employment share (30-37 per cent), only 27 per cent of those who were not observed (OFLM) in 2003 but entered the labour market in one of the following years were in a RM occupation in 2018. This picture is in line with RM workers being hardest hit by the economic recession (see Figure 2). In general, the transition matrix reveals a high job market turnover. This is a known problem in the Brazilian economy that deteriorates productivity and wages. Rocha, Pero and Corseuil (2019) indicate that the annual turnover rate at the firm level in Brazil fluctuates around 45 per cent, ranging from 38 per cent to 64 per cent. The high turnover negatively impacts productivity by reducing specific learning in organizations.

3.2 Propensity to change occupation and sector

This section analyses how skill endowments influence the probability of workers moving into different occupations or sectors. That is, we aim at providing answers to the testable hypotheses H1, H2 and H6 listed in section III.D. In the next section, the consequences of these changes on wages and occupational distance are explored further.

Table 4 presents the results from the logit regressions where the binary dependent variables are the two definitions of occupation change and if a worker migrates to a different sector. The upper part of the table reports the marginal effects of the continuous skill variables. The lower part shows the predicted change probabilities for workers with average characteristics in each discrete skill group. 17

Table 4. Occupation change, sector change and skill endowments

Note: the upper part of the table presents the marginal effects of skill scores on the probability of either moving to a different occupation or sector applying logit regressions according to equation (4). The lower part of the table presents the predicted probabilities for workers from each of the four discrete skill groups (non-routine cognitive (NRC), routine cognitive (RC), non-routine manual (NRM), and routine manual (RM)) with average characteristics. The dependent variables in columns (1) to (4) assume the value 1 if an occupation change between the current and the next employment spell is observed. Columns (1) and (2) use the 4-digit occupation classification, whereas columns (3) and (4) are based on the wider (2-digit) occupational classification. The estimation in columns (2) and (4) are similar to columns (1) and (3), respectively, but the sample is further restricted to workers who change sectors. In column (5), the dependent variable indicates whether the worker switches to a different sector and the consecutive employment (CE) spell sample is used as in columns (1) and (3). The estimations control for workers’ tenure, gender, age and age squared, education, number of hours worked, type of employment contract, firm size and legal nature, as well as year, sector, and federal state fixed effects. Standard errors in parenthesis are clustered at the occupation level. ${?}^{\mathrm{*}}$ denotes significance at ten, ${?}^{\mathrm{*}\mathrm{*}}$ at five and ${?}^{\mathrm{*}\mathrm{*}\mathrm{*}}$ at 1 per cent level.

Source: elaborated by the authors.

The first column of Table 4 shows that higher NRC skills are positively related to the probability of moving to a different job. This relation is stronger than for any other type of skill, yet none of the marginal effects is statistically significant. Once the observable individual characteristics are taken into account, an average worker in a typical NRC job has an occupation change probability of 16 per cent. RC and NRM workers change occupations at slightly higher rates (19 per cent), whereas RM workers are most likely to switch occupations (22 per cent). Note that these predicted probabilities are higher and of a different order than the unconditional observed probabilities in Table 2. That is, characteristics such as workers’ age, tenure, sectoral affiliation, employment contract type and firm size result in the lowest occupational and sectoral change probabilities for NRC workers. In contrast, NRM skills seem to be best employed in the same occupation, especially when one moves to a different sector. 18

More precisely, column (2) shows that workers who use 10 per cent more NRM skills have a 9 per cent lower occupation change probability, ceteris paribus. By conditioning the sample based on sectoral change, we observe a jump in predicted probabilities to above 62 per cent for all skill groups. The differences in predicted probabilities across skill groups are less pronounced in this case. As in the previous estimations, NRC workers are most likely to maintain their previous occupation in the new sector, and RM workers are more likely to switch jobs as they change sector. However, column (5) shows that NRM skills seem to booster sectoral switching the most. So, in this case, the work activity of NRM (predominantly service jobs) seems to be much more relevant than the sector in which it is applied. The high occupational and sectoral change probabilities of RM workers seem to stem from their unfavourable labour market position, including low wages, higher job instability and unemployment rates.

Table 5. Occupation change, sector change, skill endowments and moderators

Note: the table shows the marginal effects of the skill measures (non-routine cognitive (NRC), routine cognitive (RC), non-routine manual (NRM), and routine manual (RM)) evaluated at the lowest and highest tertile of the moderator variable’s distribution. These marginal effects are derived from logit regressions according to equation (4) with interactions between the skill measures and a moderator variable, as indicated in the second row. The dependent variables assume the value 1 if an occupational or sectoral change between the current and the next employment spell is observed. The estimations control for workers’ tenure, gender, age and age squared, education, number of hours worked, type of employment contract, firm size, legal nature, as well as year, sector, and federal state fixed effects. The consecutive employment (CE) spell sample is used and the number of observations is equal to 39,086,764 in all estimations. Standard errors in parenthesis are clustered at the occupation level. ${?}^{\mathrm{*}}$ denotes significance at ten, ${?}^{\mathrm{*}\mathrm{*}}$ at five, and ${?}^{\mathrm{*}\mathrm{*}\mathrm{*}}$ at one percent level.

Source: elaborated by the authors.

Table 5 reveals some heterogeneity across skills for the moderator variables of age, tenure, GDP growth, and firm size. Higher tenure seems to increase the probability of NRC workers assuming different occupations and sectors. For example, column (2) indicates that NRC workers in the lower tertile of the age distribution have a marginal effect on the occupation change probability equal to 0.37, which increases to 0.77 in the upper age tertile. In other words, when an average worker aged 28 increases their NRC skill set by 10 per cent, the occupation change probability increases from 16 per cent to 19.7 per cent, while the probability for a worker aged 43 would increase to 24.7 per cent. This observation seems to be in line with the age-biased technical change in Brazil, highlighted in Firpo and Portella (2019) whereby some of the older NRC workers, whose skills are particularly complemented by new technologies, would rather switch to a different occupation and constantly face the need for adjustments in their job. Contrarily, higher tenure and higher manual skills seem to decrease the propensity for sectoral switching. GDP growth does not seem to change the marginal effects of skills in a relevant way. Lastly, firm size appears to decrease the marginal effects of either skill type, both regarding occupational and sectoral changes.

3.3 Consequences of occupation changes

Tables 6 and 7 show how skills relate to the distance between occupations and to wages when workers switch occupations according to hypotheses H3, H4 and H6. Note that these findings complement our previous observations from the Transitions Matrix between skill groups in Table 3. The difference is that the regressions in Table 6: (a) make workers more comparable with each other, (b) apply the continuous skill measures, and (c) consider only job-to-job transitions. Similar to the observations in Table 5, the skill measures’ coefficients are insignificant in the baseline specification because the apparently substantial heterogeneity between workers leads to high standard errors. Once we use interaction regressions or distinguish whether workers move to higher or lower-paying occupations in Tables 6 and 7, we uncover some significant heterogeneity among workers.

Table 6. Distance after occupation changes

Note: the table presents the estimated coefficient of skill measures (non-routine cognitive (NRC), routine cognitive (RC), non-routine manual (NRM), and routine manual (RM)) and their interaction effects with a moderator variable as indicated in the third row applying OLS regressions according to equation (4). The second line indicates whether the sample consists of workers changing occupations (columns (4) to (7)) or occupations and sectors (column (3)) or changing to an occupation with higher or lower average wage in columns (1) and (2), respectively. The dependent variable is the log of the occupational distance. The estimations control for workers’ tenure, gender, age and age squared, education, number of hours worked, type of employment contract, firm size and legal nature, as well as year, sector and federal state fixed effects. Standard errors in parenthesis are clustered at the occupation level. ${?}^{\mathrm{*}}$ denotes significance at ten, ${?}^{\mathrm{*}\mathrm{*}}$ at five and ${?}^{\mathrm{*}\mathrm{*}\mathrm{*}}$ at one per cent level.

Source: elaborated by the authors.

The results in columns (1) to (2) in Table 6 indicate that having more non-routine skills decreases the occupational distance. Once we account for interactions among skills and age, tenure and firm size, the majority of marginal interaction effects also become significant. For example, columns (4) and (7) in Table 6 show that the pure effects of age and firm size on occupational distance are positive, while their interaction with skills is negative. That is, having more non-routine skills allows older, more tenured workers to find closer occupations. The same is true for workers that have worked in large companies. It is indeed intuitive that workers who have accumulated more experience on the job or in the labour market would want to utilize those abilities in a similar occupation whenever feasible. The business cycle does not seem to affect occupational distance, and neither shows any heterogeneous effect across skill types. Finally, column (3) shows that workers with more RM skills move to more distant occupations once they change occupation and sector.

Regarding wage changes following a job-to-job transition, Table 7 indicates that only the NRC skills show a positive relation, independent of whether the worker moves to an occupation with higher or lower average wage. When workers change sectors and occupations, the differences between skill groups are insignificant. Extending those regressions with the four interaction terms used previously, Table 7 shows that neither firm size, tenure nor age show meaningful or significant estimates. Only regional GDP growth seems to magnify the wage changes of non-routine skills. In line with the observed employment changes and demand for skill, we observe that during expansionary periods wage increases are higher for NRC workers, whereas the marginal return to RM skills is almost zero.

Table 7. Wages after occupation changes

Note: the table presents the estimated coefficient of skill measures (non-routine cognitive (NRC), routine cognitive (RC), non-routine manual (NRM), and routine manual (RM)) and their interaction effects with a moderator variable as indicated in the third row applying OLS regressions according to equation (4). The second line indicates whether the sample consists of workers changing occupations (columns (4) to (7)) or occupations and sectors (column (3)) or changing to an occupation with higher or lower average wage in columns (1) and (2), respectively. The dependent variables are the log wage change. The estimations control for workers’ tenure, gender, age and age squared, education, number of hours worked, type of employment contract, firm size and legal nature, as well as year, sector and federal state fixed effects. Standard errors in parenthesis are clustered at the occupation level. ${?}^{\mathrm{*}}$ denotes significance at ten, ${?}^{\mathrm{*}\mathrm{*}}$ at five and ${?}^{\mathrm{*}\mathrm{*}\mathrm{*}}$ at one per cent level.

Source: elaborated by the authors.

It is important to note that the present approach does not identify marginal wage returns as Gonzaga and Guanziroli (2019) derive from Mincerian wage regressions. Evidence from Brazil and the United States indicates that NRC wages receive the highest marginal returns, particularly in agglomerations (Ehrl and Monasterio 2019). The present approach only allows us to check whether otherwise observably equivalent workers instantly receive a systematically higher or lower wage in a novel or different occupation, depending on their current skill level. Another caveat is that the present estimates are conditional on changing jobs, inside or in other company. RAIS captures changes in job titles and wages annually, even if the worker remains with the same employer. However, the sample may not be representative of the entire population due to selection bias, as the probability of job changes is related to skills, as shown in Table 5.

3.4 Duration results

This section presents how skills are related to the formal labour market exit probability given that the worker re-enters at a later time. That is, the following duration analyses test and largely confirm hypotheses H5 and H6. The reader should bear in mind that the period out of the labour market may either be unemployment, voluntary time-out or in the informal sector. Temporary illness or injury leave is not accounted as a labour market exit.

Table 8. Risk of transition out of the labour market with moderator variables

Note: the table presents the estimated coefficients of the skill measures (non-routine cognitive (NRC), routine cognitive (RC), non-routine manual (NRM), and routine manual (RM)) from equation (6) and its extension with interaction terms using the moderator variables specified in the second line according to the duration model in equation (6). The estimations are based on the complete sample and include the workers’ tenure, gender, age and age squared, education, number of hours worked, type of employment contract, firm size and legal nature, as well as year, sector and federal state fixed effects. Standard errors in parenthesis are clustered at the occupation level. ${?}^{\mathrm{*}}$ denotes significance at ten, ${?}^{\mathrm{*}\mathrm{*}}$ at five and ${?}^{\mathrm{*}\mathrm{*}\mathrm{*}}$ at one per cent level.

Source: elaborated by the authors.

Table 8 displays the estimated coefficients $?$ according to the model specified in equation (6). That is, variables with positive coefficients increase a worker’s hazard rate, ceteris paribus. Thus, the first column indicates that workers who use manual skills intensively have a greater exit risk than the reference category of RC workers. For a different interpretation of the results, recall that the hazard rate representation can be derived as $\mathrm{e}\mathrm{x}\mathrm{p}\left(?\right)$. So, for example, the elasticity of the hazard rate with respect to RM skill is equal to $6.9=\mathrm{e}\mathrm{x}\mathrm{p}\left(1.93\right)$. Occupations that require 10 p.p. more RM skills have a 69 per cent higher hazard rate.

Column (3) reveals substantial interaction effects between skill endowments and the age profiles of workers. On the one hand, higher age, per se, and having higher NRC skills seems to decrease the exit probability. On the other hand, the combination of being more specialized in either NRC or manual skills and being older increases the labour market exit risk. This result could be due to older, more experienced workers having stronger preferences for self-employment or retirement. The business cycle and tenure, however, are apparently unrelated to the labour market exit risk of employees with different skill types. Lastly, column (5) shows that manual skilled workers are more protected from losing their job when they work in large firms. There seems to be no interaction between NRC skills and firm size.

3.5 Robustness

Our findings are robust with regard to the definition of our skill measures. First, there are no qualitative modifications regarding Acemoglu and Autor’s (2011) choice of O*NET task elements of each skill measure. Our two minor changes were necessary to correct one obvious problem of O*NET and one adaptation to the Brazilian labour market (see Data Appendix A.A1). The methodology for measuring skill rests on only one ad hoc choice: the percentile of occupations for each skill group that feeds the NLP algorithm that extracts the verbs matched to the CBO. The results presented above are based on the top 10 per cent of occupations. Selecting only the occupations closest to the top increases the risk of leaving out verbs representing each skill. On the other hand, too low a cut-off point would include occupations with verbs not characteristic of the four skills we considered. Notwithstanding, the qualitative results hold with different cut-off points at, for example, 5 per cent and 15 per cent.

The results are also robust to the choice of the English-to-Brazilian Portuguese translation algorithm. We ran initial tests with the “Google translate” function from Google Calc. Overall, the results are robust, but DeepL Pro yields the best translations because it considers the context of the occupation description more accurately.

Moreover, the regression results hold when: (a) the sample is restricted to full-time employees; (b) workers who change jobs more than three times over the observation period are excluded; (c) when instead of $?$Occ4, the wider definition of occupation change ($?$Occ2) is used; and (d) the definition of sector and occupation changes is conditional upon moving to a different firm. As mentioned previously, the findings are also robust against changes in the set of control variables, as well as to restricting the sample to male employees.

A.1 Task measurement

The skill requirements by occupation are estimated by means of a novel technique suited to the Brazilian labour market. Previous studies for Brazil (Ehrl and Monasterio 2019; Arellano-Bover 2020; among others), calculated skills by occupation based on a mapping between Brazilian and United States occupations and the task scores from the O*NET. Kubota and Maciente (2019) avoided CBO-O*NET mapping and discrete occupational categories. The authors have built an ad hoc dictionary of keywords associated with automation, based on the classification of activities according to their routine requirements and cognition as proposed for Germany in Spitz-Oener (2006).

Acemoglu and Autor (2011, p. 1163) is our starting point for converting O*NET task requirements to operational measurements. Their classification includes the element “4.A.3.a.3 Controlling Machines and Processes (not including computers or vehicles)” in the RM task measure. There are two major problems with replicating this in our study: (a) “A.3.a.3 Controlling Machines and Processes” explicitly excludes the handling of machines and computers, but the occupations that achieve the highest importance scores are airline pilots and computer programmers. 19 This obvious contradiction may have gone unnoticed by Acemoglu and Autor (2011); (b) in the Brazilian economy, RM tasks are much more physical than in the United States since its capital endowments are relatively lower. A direct application of Autor and Acemoglu’s classification would not take this into account. Therefore, we chose to change their measure for the routine tasks to correct these shortcomings and limitations. Our paper makes the following change to the manual routine activities measure: instead of "Controlling Machines and Processes", we have included the element 1.A.3.a.3 Dynamic Strength. Therefore, our measure of task requirements by skill is as follows:

1. Non-routine cognitive (NRC)

   • 4.A.2.a.4 Analysing data/information

   • 4.A.2.b.2 Thinking creatively

   • 4.A.4.a.1 Interpreting information for others

2. Routine cognitive (RC)

   • 4.A.4.a.4 Establishing and maintaining personal relationships

   • 4.A.4.b.4 Guiding, directing and motivating subordinates

   • 4.A.4.b.5 Coaching/developing others’ routine cognitive

   • 4.C.3.b.7 Importance of repeating the same tasks

   • 4.C.3.b.4 Importance of being exact or accurate

   • 4.C.3.b.8 Structured v. unstructured work (reverse)

3. Routine manual (RM)

   • 4.C.3.d.3 Pace determined by speed of equipment

   • 4.C.2.d.1.i Spend time making repetitive motions

   • 1.A.3.a.3 Dynamic strength

4. Non-routine manual (NRM)

   • 4.A.3.a.4 Operating vehicles, mechanized devices or equipment

   • 4.C.2.d.1.g Spend time using hands to handle, control or feel objects tools or controls

   • 1.A.2.a.2 Manual dexterity

   • 1.A.1.f.1 Spatial orientation

A.2 Building a dictionary of verbs and skills

After recalculating the task measures, we selected the occupations that reached the top 10 per cent in each skill. The elements associated with each of these occupations composed the file to be translated into Portuguese. This resulted in 6,100 tasks.

Table A1 lists the top ten most frequent verbs in each skill. A visual inspection of these verb lists suggests that the procedure we created for this study makes sense. Even without an ad hoc verb list, we have obtained verbs that intuitively correspond to skills.

Table A1. 10 most frequent verbs in each skill (in English)

Source: elaborated by the authors.

Table A2. A random sample of frequency of stems by skill in top_skills_verbs

Source: elaborated by the authors.

The same process of filtering verbs was done with respect to the CBO activities descriptions. This resulted in a list of more than 132,000 thousand verbs listed in task descriptions of all 2,487 CBO occupations. We named this file cbo_verbs_occupations. Table A3 presents a sample of this dataset. We selected only the verbs that both files cbo_verbs_occupations and top_skills_verbs had in common. There were 996 verbs that fulfil this restriction.

Table A3. Random sample of frequency of stems by CBO code in cbo_verbs_occupations

Source: elaborated by the authors.

A.2.1 Introducing TF-RR

Our aim is to use the data from top_skills_verbs to evaluate the importance of each skill in each CBO occupation. A simple count of the frequency of verbs associated with each skill is not a good choice. After all, the verbs in the top_skills_verbs contain a count of the frequency of verbs per skill, and therefore assigning a verb to a single skill would lose information.

In general terms, our measure of skill by occupation of CBO is the result of the following matrix product.

$Skill=W\mathrm{*}Occup$

Where:

• $W:$ Matrix $I$ skills by $J$ verbs based on top_skills_verbs dataset;

• $Occup:$ Matrix $M$ verbs by $N$ CBO occupations based on cbo_verbs_occupations dataset;

• $Skill$: Matrix with measure of the intensity I skills by N occupations.

In our case:

• $J=M=996$: number of verbs j and m;

• $I=4$: i $?$ { rc,nrc,rm, nrm};

• $N=2487$: number of CBO occupations.

${\text{tf}}_{ij}=\frac{\text{raw count of verb j in skill i}}{\text{raw count of verbs in skill i}}$

${\text{rr}}_j=\underset{i=1}{\overset{I}{?}}{\left(\frac{\text{raw count of verb j in skill i}}{\text{raw count of verb j in all skills}}\right)}^2$

$1/I?rr?1$

${\text{tf-rr}}_{ij}=t{f}_{ij}\mathrm{*}{rr}_j$

Our matrix $W_{ij}$ in eq A1 is:

$\mathbf{W}=\left[\begin{array}{cccc}{\text{tf-rr}}_{11}& {\text{tf-rr}}_{12}& ...& {\text{tf-rr}}_{1J}\\ {\text{tf-rr}}_{21}& {\text{tf-rr}}_{22}& ...& {\text{tf-rr}}_{2J}\\ {\text{tf-rr}}_{31}& {\text{tf-rr}}_{31}& ...& {\text{tf-rr}}_{3J}\\ {\text{tf-rr}}_{41}& {\text{tf-rr}}_{41}& ...& {\text{tf-rr}}_{4J}\end{array}\right]$

where tf-rr_ij is the tf-rr of skill i, term j. If skill i has no verb j, then tf-rr=0.

The elements of the $Occup$ (Eq. A1) matrix are much simpler.

$\mathbf{O}\mathbf{c}\mathbf{c}\mathbf{u}\mathbf{p}=\left[\begin{array}{cccc}{\text{Occup}}_{11}& {\text{Occup}}_{12}& ...& {\text{Occup}}_{1N}\\ {\text{Occup}}_{21}& {\text{Occup}}_{22}& ...& {\text{Occup}}_{2N}\\ \text{...}& \text{...}& ...& \text{...}\\ {\text{Occup}}_{J1}& {\text{Occup}}_{J1}& ...& {\text{Occup}}_{JN}\end{array}\right]$

where Occup_jn is the raw count of verb j in the description of CBO occupation n.

A.3 Definition of discrete skill groups

Our classification of discrete skill groups according to CBO 2002 1- and 2-digit occupations is similar to the one by Cortes et al. (2020) with minor adaptions according to the skill values from the O*NET:

1. Non-routine cognitive (NRC): managers and professionals, science technicians encompassing the 1- and 2-digit occupations [1, 2, 30-32)];

2. Routine cognitive (RC): teachers, advanced technicians in administration and others, administrative services and sales [33, 35, 39, 4, 52)];

3. Non-routine manual (NRM): technicians in transport services, in cultural services, service workers, agricultural producers [34, 37, 51, 61)];

4. Routine manual (RM): workers in agriculture, production, craft, repair [62-64, 7-9)].

Additional figures

Figure B1. Skill scores over the wage distribution 2003–2018

Source: elaborated by the authors.

Additional tables

Table C1. Skill scores by skill categories

Note: the cells show the mean and standard deviation (in parenthesis) for the skill scores (rows) over the entire observation period 2003–2018 for the skill categories (columns) non-routine cognitive (NRC), routine cognitive (RC), non-routine manual (NRM), and routine manual (RM). Calculations and based on the entire sample and apply weights for employment spell duration and hours worked.

Source: elaborated by the authors.

Table C2. Transitions Matrix between skill groups by subperiod 2003–2018

Note: the cells show the percentage of workers from a given skill category (non-routine cognitive (NRC), routine cognitive (RC), non-routine manual (NRM) and routine manual (RM)) in the initial year (column 1) that are either out of the formal labour market (OFLM), employed in the same or in another skill group five years later, or retired in or before that year. Per definition, the row sum is equal to 1. Calculations are based on the entire sample and apply weights for employment skill duration and hours worked.

Source: elaborated by the authors.