The employment implications of additive manufacturing

ABSTRACT

In spite of the fast spread of Additive Manufacturing (AM) in several countries and industries, its impact on employment is still unexplored and theoretically ambiguous. On the one hand, higher product customisation and shorter time-to-market entail an expansion of the market, thus fostering labour demand; on the other hand, AM profoundly changes the way goods are produced and little evidence exists regarding the complementarity or substitutability between AM technologies and labour.

In this article, we contribute to fill this gap. We estimate labour demand functions augmented with a (patent-based) proxy of AM-related innovation in 31 OECD countries, across 21 manufacturing industries, over the 2009–2017 period. Our econometric findings show an overall positive relationship between AM technologies and employment at the industry level, due to both market expansion and complementarity between labour and AM technologies, while no labour-saving effect emerges. The importance of each mechanism, however, is heterogeneous across sectors.

KEYWORDS:

• Additive manufacturing
• 3D printing
• employment
• technological change
• industry-level analysis

JEL classification:

• J23
• O31
• O33.

Acknowledgments

The authors would like to thank two anonymous reviewers for their most insightful suggestions. The authors are especially grateful to Davide Castellani for his valuable comments and suggestions on the earlier version of the paper.

Supplementary material

Supplemental data for this article can be accessed here.

Disclosure statement

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

Notes

¹ See also Table A1 in the online Appendix.

² AM is becoming one of the main areas of study in the social sciences, from economics to business and management (Mariani and Borghi, 2019).

³ Indeed, this revolutionary manufacturing perspective, which involves adding and instantly joining layers of various materials in specific locations and creating objects from digital 3D data, has progressively gained attention in several fields, being used either as a complementary or mainstream manufacturing technology (Laplume et al., 2016).

⁴ For recent surveys, see Pianta (2006) and Vivarelli (2014).

⁶ An innovation can be labour-augmenting (i.e. increasing labour productivity) and at the same time labour-biased (requiring more labour) if the elasticity of substitution between labour and other factors in production is low enough (Acemoglu, 2009, pp. 500–503).

⁷ Several studies analysing the effect of product innovation report evidence on business stealing and cannibalization (recently, Harrison et al., 2014).

⁸ The pros and cons of different innovation proxies are well known, being extensively stressed in the literature (Archibugi and Pianta, 1996; Hagedoorn and Cloodt, 2003).

⁹ For instance, survey data from the International Federation of Robotics (used by several contributions cited above) do not allow disentangling manufacturing users of robots from integrators, i.e. the firms ‘integrating’ the robot into the manufacturing process. This challenges a correct attribution of robots to the sector/country adopting it.

¹⁰ Including a proxy for either adoption or production alone would imply a serious omitted-variable problem when dealing with the total employment effect at the industry level of the introduction of a new technology, since each of them taken alone overlooks one important part of the employment implications of a specific technology.

¹¹ Nonetheless, the use of data on adoption of the new technology is more questionable whenever the research question regards total employment, and therefore, the results are reported in terms of the ‘effect of ICT/robots on employment’. Indeed, by using proxies of pure adoption at the industry level, they capture the effect of ICT/robot ‘adoption’ on employment, i.e. not taking into account the effect on employment in industries producing the new machines and the related goods (e.g. software). The related bias can be more or less relevant depending on the type of technology under investigation, i.e. its relative impact in adopting and producing industries. For instance, while the impact of ICT has been radical in both producing and adopting industries, opening up entirely new markets (e.g. personal computers and mobile phones), the impact of industrial robots in producing industries is probably much more limited.

¹² We focus on applications to the USPTO as it is considered the reference patent office when seeking protection for innovative technologies (Cantwell, 1995).

¹³ The version of PATSTAT used is PATSTAT Online (2019 Autumn edition) V5.14, accessed between September and October 2019.

¹⁴ We followed guidelines from Pasimeni (2019) to improve the effectiveness of our SQL query in PATSTAT.

¹⁵ See Cappelli et al. (2021) on the relationship between economic resilience, measured by unemployment rate, and technological resilience after the 2008 crisis in 248 EU regions, highlighting the role of institutions and policies at the country level in affecting the interaction between the two and their dynamics.

¹⁶ FDM and SLS were invented and their patent applications first filed at the USPTO in 1989 and 1986, respectively. Patents were granted in 1992 and 1997. The core patent for FDM expired in 2009 and SLS’s one in 2014.

¹⁷ Since the variables include zeros, we added 1 before taking natural logarithms.

¹⁸ According to the perpetual inventory method, the initial stock is given by$F_{ij0}^{AM}/\left(\delta +G{R}_{ij}\right)$, with $G{R}_{ij}$ representing the average growth rate in AM patent families between 1989 and 2015. Having collected patent information over the 1989–2017 period, we build the stock using additional information on pre-sample years. We use 2015 as the last year to compute $G{R}_{ij}$ as $F_{ijt}^{AM}$ drops in absolute value after that year. This does not depend on declining interest in the technology per se, as explained above.

¹⁹ Panel A in Figure 4 shows a higher number of countries presenting no patenting activity as compared to panel A in Figure 2. This is due to constraints in the computation of the stock measure used in our estimations because of countries showing single or few patent applications, hence making it impossible to compute an average growth rate $G{R}_{i,j}\ge 0$. The same holds for similar sectoral cases reported in panel B.

²⁰ We omit the terms $SB$,$SS,$ and $SII$ in equation (2)(2) $+{\alpha }_{5}nonA{M}_{ijt-3}+{\alpha }_6{X}_{ijt-1}+{\gamma }_i+{\gamma }_j+{\gamma }_t+u_{ijt},$(2) as they are collinear with the sector FEs included in all specifications.

²¹ In order to capture potential between-sector interactions and general equilibrium effects, we also carried out the analysis at the macro level. Moreover, using macro data allows us to also undertake a preliminary exploration of the potential heterogeneous effects of AM on employment by education level. Results are reported in Table A6.1 in the online Appendix and details of the analysis are reported in the table’s notes. On average, the elasticity of employment to AM is about 0.12 and 0.06 in unconditional and conditional demand estimations, respectively, and both are statistically significant at the 1% level. The elasticity is larger for middle-educated workers compared to highly educated workers; it is very small and not significant for low-educated workers.

²² Our sample includes a few countries (i.e. Estonia, Greece, Latvia, and Portugal) and sectors (i.e. 19–Coke and refined petroleum products; 33–Repair and installation of machinery and equipment) that do not have AM patents at the USPTO over the period considered in our sample. Thus, we further control for the robustness of our main results when we drop observations related to these countries and sectors. Similarly, the results (reported in Table A5.4 in the online Appendix) are robust.