Robots, skills and temporary jobs: evidence from six European countries

ABSTRACT

In our analysis of the impact of robot adoption on the use of flexible contracts in six European countries, we find that control for the type of innovation model that is dominant in an industry is crucial. In a ‘high knowledge cumulativeness’ innovation regime, robot adoption reduces the probability that high-skilled workers will receive temporary contracts, while no significant effect has been found for medium- and low-skilled workers. The rationale is: In a high cumulativeness regime, innovation depends on a firm’s internal knowledge sources, and high-skilled (rather than medium- and low-skilled) workers are crucial carriers of knowledge. The situation is different in ‘low-cumulativeness’ regimes. In the latter, firms are primarily using externally acquired knowledge in their innovation process. This makes workers more easily interchangeable and robot adoption significantly increases the probability to get temporary jobs for both medium- and high-skilled workers, but leaves low-skilled workers unaffected.

KEYWORDS:

• Robots
• temporary jobs
• innovation regimes
• knowledge cumulativeness
• skills

JEL:

• J3
• J5
• M5
• O3

Disclosure statement

No potential conflict of interest was reported by the author(s).

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Notes

¹ We also use additional data to normalise robots (EUKLEMS) or to perform robustness checks (Eurostat data for trade and national accounts, and UN Comtrade database for international trade in R&D).

² Although SES is conducted over many European countries, singling out our sample was the result of a trade-off among a number of data limitations. First, many countries included in the SES dataset do not report a sufficient industry breakdown for correctly mapping economic activities into High & Medium and Low-Knowledge Cumulativeness industries. This is the case for Denmark and Finland. Germany also falls in this group in 2006. The importance of this country in terms of robot adoption led us to retain it for 2010 and 2014. Second, in many Eastern European countries, robot adoption figures are close to zero. Third, several countries are not covered in Peneder’s study (2010) on knowledge cumulativeness. Unfortunately, for other countries such as Sweden, that is important in terms of robot adoption, SES does not report data for temporary workers. Fortunately, our sample includes the biggest five economies of the former EU-28 and, in 2014, they still account for 85% of all robots introduced in the EU-28.

³ Following Graetz and Michaels (2018), we do not use the IFR categories ‘all other manufacturing’, ‘all other non-manufacturing’, and ‘unspecified’. This is because the bulk of robots from the latter three industries is included in ‘unspecified’ and the risk of misallocation of these robots among industries is high. We use weights based on shares of employees to split robots in the R&D and in the Education sectors. The motivation is that Peneder’s taxonomy does not include Education, but covers R&D as an industry with high knowledge cumulativeness (see Table 1).

⁴ Differently from Acemoglu and Restrepo (2020), our dependent variable is not the cumulated days of employment over years, but the probability to have a temporary job in 2006, 2010 and 2014 respectively. We assume that 10 years of robot exposure is a sufficient time to shape the propensity of firms to employ temporary workers.

⁵ Eurostat, Statistics explained, online documents, https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Glossary:High-tech.

⁶ Since we have a hierarchical model with different potential levels for clustering, we follow the Abadie et al. (2022) suggestions to decide what to cluster over and assume that besides the assignment process, the sampling process and the level to which the heterogeneity of treatment emerges do matter. Despite robot exposure is at country-industry level, we have to consider that the Structure of Earnings Survey relies on a two-stages sample design, where a stratified sample of local units (establishments) is drawn in the first stage, and a simple sample of employees is taken within each of the selected local units, in the second stage. Further, almost all our key results stem from interactions between country-by-industry level robot exposure and individual level age, skills and education. It means that we should observe heterogeneity of the treatment (robot exposure) within the country-by-industry level cluster. Eventually, it is plausible to assume that, being other individual characteristics equal, the probability to get a temporary contract is correlated for employees within the same establishment/firm as result of a specific strategy conducted by companies. Even though we do not directly focus on the company level, we hypothesised different behaviour among companies in the same industry (see H.2a and H.2b) and control for their productivity dispersion (see discussion about its effects on section 3.2). For these reasons, we conjecture that correlation of errors at establishment level is more important than that at the country-by-industry level.

⁷ APE is the numerical derivative of the probability to get a temporary contract with respect to a variable of interest for each observation using the other covariates as they were observed, and then the average of all these individual marginal effects across the sample.

⁸ In high-cum industries (Table 3, column 10), robot exposure only reduces the probability to get temporary jobs for high-educated workers, whereas the APE is positive for individuals with both primary (0.001) and secondary educational attainments (−0.0005 + 0.001=+0.0005). We do not go through this result because it is not confirmed by robustness tests (see Tables A.2, A.3 and A.4, columns 10).

⁹ The total effect of robots for workers with secondary educational attainment (Rob x Sec_Educ) is the algebraic sum of the coefficients Rob_Exposure + Rob x Sec_Educ.

¹⁰ Graetz and Michaels find that, unlike ICT, automation does not polarise the labour market, since the negative effects of robots on hours worked by the least educated workers are no lower than those of the middle skilled.

¹¹ As already reported for results in Table 3, all the differences between the APEs of interest estimated across the two technological regimes (low- vs high-cum) in Tables A.2, A.3 and A.4, are statistically significant (see Table A.9, columns 6, 10 and 12).