Literature review

The SBTC hypothesis was originally formulated to explore and explain the changes effected by the increasing automation and computerisation of the workplaces (Katz and Murphy Reference Katz and Murphy1992). The hypothesis essentially assumes that the more educated are more likely to have jobs and fare better in the new, emerging economy. This is formalised as a two-class-worker model: the skilled and the unskilled. The emphasis on a more educated workforce, possessing the right skills and abilities able to match the new work requirements, is identified by several authors (Autor Katz and Krueger Reference Autor, Katz and Krueger1998; Acemoğlu Reference Acemoğlu2002) as the main driving force behind the structural changes in the occupational structure. Empirical evidence demonstrates that the most educated workers tend to be the most productive, earning higher wages, and experience greater employment growth (Acemoğlu and Autor Reference Acemoğlu, Autor, Shelterer and Card2011). However, this analytical framework makes the simplistic assumption of a direct correspondence between tasks and skills (Acemoğlu and Autor Reference Acemoğlu, Autor, Shelterer and Card2011).

Nevertheless, new empirical evidence suggested that the impact of automation on occupations is not solely determined by their level of training and education. Works by several authors (Autor, Levy, and Murnane Reference Autor, Levy and Murnane2003, Goos and Manning Reference Goos and Manning2007) have shown that evolution in employment tasks was no longer monotonous across earning levels. Shifts in occupational structure in favour of those at both ends of the wage distribution have defined structural changes in the labour markets (Acemoğlu and Autor Reference Acemoğlu, Autor, Shelterer and Card2011). This new reality has been formalised in a new theoretical model, which explains the dynamics of occupational change by adding the class of middle-skill workers. The model also introduces a task-based approach that models the influence of tasks on skills allocation and its impact on earnings and occupational structure (Acemoğlu and Autor Reference Acemoğlu, Autor, Shelterer and Card2011).

Changes in the occupational structure are thus explained by the impact of new technologies or new business practices (Acemoğlu and Autor Reference Acemoğlu, Autor, Shelterer and Card2011), the implementation of which enables the replacement of tasks deemed to be ‘routine’ or repetitive. The tasks under discussion are distinguished from those designated as ‘non-routine’ in that they demand the application of skills in contexts necessitating problem-solving, complex communication, creative content, or personal services (Haslberger Reference Haslberger2022; Autor Reference Autor2015). In many non-routine contexts, technology has an augmenting role, which is consistent with the findings of Frey and Osborne (Reference Frey and Osborne2017), who identify the ‘engineering bottlenecks’ as tasks that are not easily automatable: namely, manual tasks involving perception and manipulation, as well as creative and social tasks. In both approaches, the task and skill content of jobs is analysed and acknowledged as one of the key drivers which explain the structural shifts of employment towards the non-automatable/ hard-to-automate jobs (Pouliakas Reference Pouliakas2018).

A significant body of research has focused on defining the task content of occupations from the routinisation perspective. According to Haslberger (Reference Haslberger2021), the main approaches related to defining the task content are laid out by Autor and Dorn (Reference Autor and Dorn2013) and identify three task categories: routine, abstract, and manual tasks, and by Macias and Hurley (Reference Fernández-Macías and Hurley2017), who build three indexes based on survey responses aggregated at the job level.

These approaches have materialised as routine task intensity (RTI) indexes, which refer to both manual and non-manual tasks, in an attempt to explain the decline of routine jobs consistent with the RBTC hypothesis (Autor and Dorn Reference Autor and Dorn2013; Fernández-Macías and Hurley Reference Fernández-Macías and Hurley2017; Haslberger Reference Haslberger2022). Complexity indexes, on the other hand, seek to explain the above-average performance of jobs that are difficult, if not impossible, to automate (Fernández-Macías and Hurley Reference Fernández-Macías and Hurley2017; Haslberger Reference Haslberger2022) by summarising higher-level skills such as communication, abstraction, and decision-making associated with above-average employment growth (Caines et al Reference Caines, Hoffmann and Kambourov2017). These indexes serve as an alternative to the routine dimension of occupational change, rather than as a counterpoint to RTI indexes (Haslberger Reference Haslberger2021).

Analysing structural changes in the occupational structure often generates sweeping conclusions and predictions. Thus, SBTC usually predicts straightforward occupational upgrading (Goldin and Katz Reference Goldin and Katz2008, cited by Haslberger Reference Haslberger2022), while RBTC tends to predict straightforward polarisation (Autor et al Reference Autor, Levy and Murnane2003; Goos et al Reference Goos, Manning and Salomons2009). In many cases, the analytical framework and conclusions consider that a dichotomy exists between the two (Haslberger Reference Haslberger2022; Oesch and Picitto Reference Oesch and Piccitto2019; Oesch and Rodrigues-Menes Reference Oesch and Rodriguez Menes2011; Macias and Hurley Reference Fernández-Macías and Hurley2017), often driven by the fact that polarisation is the prevailing, mainstream analytical framework. This reality is more striking as it does not occur only in multi-country studies. Even in single-country studies for UK and Spain, divergent conclusions have been reached for similar periods (Sebastian Reference Sebastian2019; Salvatori Reference Salvatori2018).

This artificial dichotomy that shaped the literature on structural employment changes has overlooked several arguments. Acemoglu and Autor (Reference Acemoğlu, Autor, Shelterer and Card2011) state that SBTC is a ‘natural starting point’ in explaining structural changes in the occupational structure, whose findings are being refined following empirical evidence by RBTC. Autor (Reference Autor2015) also states that polarisation is likely to be less relevant in explaining the structural changes in the occupational structure in the future, as (middle-skill) job contents will increasingly incorporate non-routine tasks requiring higher levels of technical and interpersonal skills. This finding aligns with Sebastian (Reference Sebastian2018), who noted that in Spain, middle-skill workers transition to both low-skill and high-skill occupations.

New insights have shown that the realities of occupational change are more nuanced than the ‘hollowing out of the middle’ picture. Foster and Wolfson (Reference Foster and Wolfson2010) were among the first to show the unbalanced nature of polarisation, with the high-skill jobs showing significant expansion, whereas low-skill jobs posting only marginal gains. Similar findings are obtained for the US by Hunt and Nunn (Reference Hunt and Nunn2022). The strong expansion of high-skill jobs as a key driver of polarisation is formalised by Sparreboom and Tarvid (Reference Sparreboom and Tarvid2016), whose polarisation index factors in the ‘imbalanced nature of polarisation’, and also helps determine to what extent it is tilted towards job expansion at (usually) the higher end of the wage distribution.

These findings are confirmed by multi-country evidence. The diversity of structural change patterns across several European countries has been highlighted by several authors, including Fernández-Macías (Reference Fernández-Macías2012) and Martinaitis et al (Reference Martinaitis, Christenko and Antanavičius2021). They have demonstrated that polarisation is only one of several structural change patterns in these countries and is not even the prevalent one. Rather, research focusing on several countries arrives at strong conclusions in favour of upgrading (Oesch and Picitto Reference Oesch and Piccitto2019; Oesch and Rodriguez-Menes Reference Oesch and Rodriguez Menes2011). On the other hand, the performance of service workers has partially explained the polarisation patterns of the occupational structure (Tahlin Reference Tahlin2019; Acemoglu and Autor Reference Acemoğlu, Autor, Shelterer and Card2011). Gender, age structure, industry, and immigration are also used to explain occupational structure changes (Salvatori Reference Salvatori2018; Tahlin Reference Tahlin2019; Murphy and Oesch Reference Murphy and Oesch2018). While these factors have explained country-specific changes in the occupational structure, they have not been able to lead to a consistent model that can unequivocally explain the occurrence of either polarisation or upgrading.

The diversity of results has been summarised by Haslberger (Reference Haslberger2022), who concludes that most studies do not yield clear-cut results in favour of either upgrading or polarisation. This is evident in Appendix 2 Table A1 which focuses on multi-country studies. Although the list is not exhaustive, we observe different results for similar countries and time periods.

This diversity of conclusions is not confined to multi-country studies. Results for analysis sub-periods confirm the fact that changes in occupational structure have not been consistent across time. Hunt and Nunn (Reference Hunt and Nunn2022) show that structural occupational change patterns differ and that the shape of changes does not correspond to unambiguous polarisation, but to situations where low-skill jobs outperform the middle-skill ones in the sense that their decline is less pronounced. Tahlin (Reference Tahlin2019) observes that polarisation over a longer period comprises periods of mid-upgrading. Oesch and Murphy (Reference Oesch and Rodriguez Menes2017) show that polarisation patterns occurred in the 1980s in Switzerland and Ireland. Given this evidence, we adopt a critical stance on the validity of some single-period, long-term analyses, as the employment structure of the start year can significantly influence overall results, potentially overlooking major interim changes.

We consider that for the purpose of analysing structural changes in European labour markets, data limitations are extremely important. In many cases, analysis has been carried out for several sub-periods (Eurofound 2015; Fernandez-Macias and Hurley Reference Murphy and Oesch2018; Sebastian Reference Sebastian2018), and due to major data breaks, results are explicitly reported over smaller periods in order to avoid reaching overall conclusions based on inconsistent data. A similar problem has occurred in the US (Hunt and Nunn Reference Hunt and Nunn2022), where the handling of a data break practically reversed the well-known polarisation result of Autor (2015) considered ‘an artefact of occupation code redefinitions’ (Hunt and Nunn Reference Hunt and Nunn2022).

Data and methods

Employment data come from European Union Labour Force Survey (EU-LFS), available at the two-digit ISCO-08 occupation level, which is the most appropriate for a ‘job approach’. The analysis period spans from 2011, the first year in which data were reported using the ISCO-08 classification, to 2023.

Significant data issues have contributed to our choice. While a longer time period may help to better assess changes in the occupational structure, major data breaks between 2008 and 2011 (Fernandez-Macias and Hurley Reference Fernández-Macías and Hurley2017; Sebastian Reference Sebastian2018) can significantly alter the results of the analysis. The transition between ISCO-88 and ISCO-08 data is not properly documented by Eurostat, as no 2011 data were reported using the ISCO-88 classification. Thus, the best option was to choose a period for which data is fully consistent.

The perceived impact of COVID-19-related measures on labour markets (OECD 2021; Livanos and Ravanos Reference Livanos and Ravanos2024) which resulted in major changes in work arrangements was also considered, leading to the choice of 2019–2023 as the maximum available time period to assess structural changes that persisted after the lockdown restrictions ended.

The first period of analysis, 2011–2019, corresponds to the recovery phase following the Great Recession, with 2019 representing a peak year in terms of employment progress and economic growth (European Commission 2019). While this affects the results for the first period, valid results on structural change can be obtained even for non-standard periods (Verdugo and Allègre Reference Verdugo and Allègre2020).

Wage data come from the Structure of Earnings Survey (SES), with waves of 2014 and 2018, representing the earliest and most reliable comparable data based on the ISCO-08 classification.

Country-specific changes are presented as distributional trends/patterns using occupational shares by wage-ranked employment quintiles, consistent with several cross-country analyses (Fernández-Macías and Hurley Reference Fernández-Macías and Hurley2017; Salvatori Reference Salvatori2018). Data availability and reliability issues for some countries, and outcomes observed by several authors (Fernández-Macías Reference Fernández-Macías2012; Oesch and Piccitto, Reference Oesch and Piccitto2019) made us choose/use a two-digit occupation-based wage structure, which provides a strong focus on job quality.

A scatter plot and a local polynomial fitting curve are employed to model changes in employment shares, using a rank-based ordering of 2-digit occupational wages as implemented by Dauth (Reference Dauth2014), starting from the lowest (1) to the highest (38) one to accommodate a wide variety of wage levels and distributions across the EU member states.

The Job Polarisation Index (JPI) (Sparreboom and Tarvid Reference Sparreboom and Tarvid2016) is used to explore the existence and unbalanced nature of job polarisation. As an analytical approach that complements the analysis of distributional patterns, the JPI uses the changes in employment shares over a six-year period, noted with the ${{\overline {{\Delta _{\rm{5}}}} }}$ operator (Sparreboom and Tarvid Reference Sparreboom and Tarvid2016), for high (h), medium (m), and low (l) level job tiers/categories using the following formula:

$${p\! = {{\rm{1}} \over {\rm{2}}} \cdot \left( {{{\overline \Delta }_{\rm{5}}}l + {{\overline \Delta }_{\rm{5}}}h} \right) \cdot \left( {{\rm{1}} + \left| {{{\overline \Delta }_{\rm{5}}}h - {{\overline \Delta }_{\rm{5}}}l} \right|} \right) \cdot {\rm{100}}}$$

The index is positive if polarisation occurs while a negative value shows a pattern of mid-upgrading (Fernández-Macías Reference Fernández-Macías2012). It also helps to distinguish, to some extent, cases of ‘true polarisation’ with lower positive values, from cases of occupational upgrading, with higher positive values. For consistency with the country-specific job quintiles, country-specific income terciles based on ISCO08 2-digit occupational wages are used in order to accommodate the diversity of wage-based national occupational hierarchies (Haslberger Reference Haslberger2021).

The RTI and complexity indexes are calculated in order to verify both RBTC and SBTC hypotheses using the methodology proposed by Haslberger (Reference Haslberger2022). Considering the analysis period, only Eurofound Working Conditions survey data from waves 5 and 6 (Eurofound 2023) are used, which correspond to the 2010 and 2015 years. ^{Footnote 1} The analysis of the wage schedules for RTI and complexity indexes is carried out using local polynomial fitting curves, which can provide an accurate understanding of the extent to which the polarisation or upgrading patterns are confirmed. Panel regression models with HAC covariance matrices are used to assess whether either, or both indexes, have statistical significance in explaining changes in occupational structure, for both periods taken separately. Following the results of Hausman tests, and because they refer to all countries in our panel, estimations are carried out using fixed effects models, for all specifications.

Most of the calculations and graphs are performed using R software. Apart from R base functions, the following packages used are: ggplot2 for graphs, and the plm package for estimating unbalanced panel regression models with country fixed effects.

Research questions and conceptual issues

Our main goal is to enhance the understanding of job polarisation and occupational upgrading, in terms of their compatibility and ability to explain the changes in the occupational structure observed in the recent past for the EU-27 countries.

To this end, our analysis starts from the generally accepted definitions used in the literature. Thus, job polarisation is defined as a decline in the share of middle-skill jobs accompanied by growth in both the high-skill and low-skill occupations and (occupational) upgrading when the share of high-skill jobs expands at the expense of both middle-skill and low-skill jobs (Fernández-Macías et al Reference Fernández-Macías and Hurley2017; Oesch and Piccito Reference Oesch and Piccitto2019) as an addition to the original two-class canonical model as described in Acemoğlu and Autor (Reference Acemoğlu, Autor, Shelterer and Card2011). ^{Footnote 2} Based on the findings of several authors (Fernández-Macías Reference Fernández-Macías2012; Oesch and Piccito Reference Oesch and Piccitto2019; Tahlin Reference Tahlin2019; Haslberger Reference Haslberger2021), the fact that national labour markets can exhibit a variety of occupational change patterns is also taken into account.

Thus, using the RBTC three-class framework (Acemoğlu and Autor Reference Acemoğlu, Autor, Shelterer and Card2011), a distinction is made between unambiguous polarisation, where both low-skill and high-skill jobs show increase at the expense of the middle-skill jobs, and dominant polarisation, where the decline of middle-skill jobs occurs while the share of low-skill jobs, in most instances, or high-skill jobs, in rare cases, actually decreases at a lower rate than that of middle-skill jobs. Similarly, a distinction is drawn between unambiguous upgrading, where higher skill levels are (monotonically) associated with higher increases, or lower decreases in employment shares, than those for lower skill levels, and dominant upgrading, where the decline in low-skill jobs is less pronounced than that of middle-skilled jobs, a situation similar to that described by Hunt and Nunn (Reference Hunt and Nunn2022) for the 1989–1999 period. ^{Footnote 3} The fact that polarisation and upgrading are not complete opposites, as shown by Haslberger (Reference Haslberger2021), is also analysed. Considering the three job tiers, it can be easily be shown that, if upgrading is present, polarisation can still occur if the gains of high-skill jobs are large enough to offset the decline in the low-skill jobs, decline which is lower than that of middle-skill jobs. Otherwise put, the share of the middle-skill jobs would shrink, as the decline in low-skill jobs is more than offset by the expansion of high-skilled jobs. This is explained in detail in Appendix 1.

Given the focus of this article, which is to explore whether polarisation and upgrading are both compatible explanations of occupational change and can both describe the current changes in the occupational structure of the labour market, the analysis will only examine the patterns of polarisation and upgrading, as defined by most authors, which predominate among the EU countries. Further information on structural change in several countries is provided in Appendix 2, so that interested readers can examine patterns that differ from those that are the focus of this article.