Research Policy 48 (2019) 103765

Contents lists available at ScienceDirect

Research Policy
journal homepage: www.elsevier.com/locate/respol

Is this time different? How digitalization influences job creation and T

destruction☆
Benjamin Balsmeiera,b, Martin Woerterb,
⁎

a
University of Luxembourg, Luxembourg
b
ETH Zurich, KOF Swiss Economic Institute, Switzerland

ARTICLE INFO ABSTRACT

Keywords: With the process of digitalization now in full swing, many are wondering how the adoption of new technologies
Digitalization influences job creation and destruction. Much hinges upon the specific tasks that machines take on and how
Employment many new tasks are created through the adoption of new digital technologies. Some argue that most tasks that
Innovation are at risk of automation are those performed by rather low- to medium-skilled employees, while most new tasks
Job creation
that emerge from the adoption of digital technologies complement high-skilled labor. We present evidence
Job destruction
Technological change
derived from representative survey data from Switzerland that is consistent with this view. Specifically, we find
that increased investment in digitalization is associated with increased employment of high-skilled workers and
reduced employment of low-skilled workers, with a slightly positive net effect. The main effects are almost
entirely driven by firms that employ machine-based digital technologies, e.g. robots, 3D printing or the Internet
of Things. We do not find any significant employment effects when non-machine-based digital technologies are
considered, e.g. ERP, e-commerce or cooperation support systems.

1. Introduction impact on jobs seems to be positive (Mastrostefano and Pianta, 2009;

Kogan et al., 2017).1 The crucial, yet unanswered, question is whether
There is widespread consensus among both academics and practi- this time is different.
tioners that innovation is the main driver of productivity, economic One way to think about potential job creation and destruction
growth and job creation. Whether this holds true with regard to in- conceptually is to differentiate between tasks that are regularly per-
novations that fall under the umbrella of ‘digitalization’ is a subject of formed by human beings but could be performed equally well or more
vigorous debate, though. This debate is particularly controversial when efficiently by machines, versus tasks that cannot be substituted by
it comes to the impact of digitalization on job creation and destruction. machine work, or new complementary tasks that are created because of
Indeed, for more than a century, bookshelves have been filled with machine adoption. Many politicians and employees feel challenged by
analyses on how technological development influences jobs. Many the ongoing digital revolution because the new technologies behind it
prominent academics have argued – and some still argue – that man and often seem to be direct substitutes for tasks that are regularly performed
machine are in a race, which machines will eventually win, leading to by human beings. Some technologies enable far-reaching changes of
mass unemployment (e.g. Leontief, 1952; Ford, 2015). Historical ex- established production lines, e.g. through intelligent platform designs
amples of jobs that were washed away by technological developments that render intermediaries obsolete or shift of tasks from employees to
seem to confirm this view, e.g. stoker that were replaced by electrical customers. Also, unlike many other technological developments of the
engines. At the same time, it is not a controversial statement that many past, digitalization is a general-purpose technology, i.e. it can be
new jobs were created in the wake of technological developments of the adopted across a wide range of industries, including the service sector
past (e.g. Landes, 1969; Chandler, 1977; Mokyr, 1990). After all, con- (Brynjolfsson and McAfee, 2016). In a frequently cited study on tasks
sistent with Schumpeter’s concept of creative destruction, the net that are susceptible to automation through digital technologies,

☆
The authors wish to thank Andrin Spescha and Gilles Aubert for invaluable research assistance. We would also like to thank Gudela Grote, Toni Wäfler, and
Spyros Arvanitis for their contributions to the design and funding of the digitalization survey for 2016. We gratefully acknowledge financial support from the MTEC
Foundation at ETH Zurich. Any errors and omissions remain the authors'.
⁎
Corresponding author.
E-mail address: woerter@kof.ethz.ch (M. Woerter).
1
For a comprehensive overview and broader discussion of the relationship between innovation and employment see Pianta (2006).

https://doi.org/10.1016/j.respol.2019.03.010
Received 1 December 2017; Received in revised form 28 February 2019; Accepted 7 March 2019
Available online 13 April 2019
0048-7333/ © 2019 Elsevier B.V. All rights reserved.
B. Balsmeier and M. Woerter Research Policy 48 (2019) 103765

robotics and artificial intelligence, Frey and Osborne (2017) conclude digitalization and on the adoption of a rich list of specific digital
that 47% of jobs in the U.S. are at risk of becoming redundant.2 technologies that ranges from well-known technologies such as ERP,
However, these numbers are unlikely to hold true (e.g. Arntz et al., customer relationship management (CRM), supply chain management
2017; Autor, 2015; Dengler and Matthes, 2015) and, even if they would (SCM) systems, e-commerce and robots to fairly recently adopted
turn out to be true, it is by no means clear that the world would be technologies such as social media, cloud computing, 3D printing, au-
facing a future of prevalent mass unemployment. Coinciding with the tonomous vehicles, and the Internet of Things.
evolution of information and communication technologies (ICT) over Consistent with the literature and recent advances in economic
the last few decades, plenty of new tasks, such as programming and theory (Acemoglu and Restrepo, 2017a), we find that investment in
data analysis, have been created that are still hard to automate via digitalization is associated with increased employment of high-skilled
machine work (Acemoglu and Restrepo, 2017a). These ICT-induced labor whereas low- and medium-skilled labor tends to decline or remain
technological developments, which might well be viewed as early ad- unaffected. Separate examinations of firms that adopt complex ma-
vances of digitalization, contributed to the emergence of whole new chine-based digital technologies and firms that use non-machine-based
industries that nowadays employ millions of workers performing tasks digital technologies reveal heterogeneous effects. Only machine-based
that formerly did not exist. Few had foreseen such developments at a technologies driving the current industrial revolution are responsible
time when companies like Facebook and Amazon had not even been for significant employment effects. The estimated impact of investment
founded. in digitalization is distinct from investment in general research and
Conceptually it is crucial to note two things. The tasks that become development (R&D) and, while important limitations remain, all results
obsolete due to digital technologies are typically of a different kind than survive controls for firm fixed effects and several robustness and sen-
the tasks created, and different types of digital technologies may have sitivity checks.
heterogeneous effects on skill requirements. Robots, for instance, ty- The investigation contributes to at least two strings of literature.
pically compete quite directly with manual labor that requires a low to First, we complement research in economics on the impact of innova-
medium skill level. Tasks that are created, on the other hand, typically tion on employment growth. Most prior studies on the topic have found
require a rather high skill level and are relatively few in total. Since jobs positive influences of innovation on employment at the firm level (e.g.
are defined by the tasks to be fulfilled, a change in the demand for Blanchflower and Burgess, 1999; Van Reenen, 1997; Entorf and
certain tasks may simultaneously lead to the creation and destruction of Pohlmeir, 1990; Smolny, 1998; Balsmeier and Delanote, 2015). Dif-
jobs with heterogeneous skill requirements. Over the last decade, for ferent types of innovation, e.g. product vs. process innovations, can
instance, research has found an increase in jobs for high-skilled and have differing impacts, though, and the net influence depends on the
low-skilled workers, while jobs for medium-skilled labor were in de- chosen strategies for innovation and varies across industries
cline (e.g. Autor et al., 2003; Goos and Manning, 2007; Goos et al., (Mastrostefano and Pianta, 2009). In the long run, the overall positive
2009; Michaels et al., 2014). This job polarization can be attributed to impact holds at the industry level; even after taking complex dynamics
technological change, but it is not clear which types of technologies of the relationship into account (Mastrostefano and Pianta, 2009).
caused these employment effects and whether technologies summarized While this literature achieved quite a deep level of understanding,
under the umbrella of digitalization lead to similar effects. empirical analyses often had to rely on rather broad measures of in-
So far, the literature has been unable to empirically link within firm novation, e.g. R&D expenses or innovation counts, which inhibit more
changes in investment in digitalization with within firm changes in fine-grained analyses of specific technological developments in a sys-
employment across different skill groups. The present paper aims to tematic way. Our data help us to overcome this limitation by enabling
address this gap in the literature in two steps. First, we will estimate us to link within firm changes specifically in investments in digitali-
standard models of employment changes in different skill categories on zation to within firm changes in jobs with differing skill requirements.
changes in investment in digitalization. In a second step, we propose Related studies analyze changes in the casualization of work and work
and empirically confirm a differentiation between the impact of non- contracts that are caused by the process of digitalization (e.g. Aubert-
machine-based digital technologies, such as ERP, social media, or e- Tarby et al., 2018) or provide in-depth analyses of a single technology,
commerce, and rather complex machine-based digital technologies, e.g. 3D-printing (Rayna et al., 2015). Arntz et al. (2018), on the other
such as robots, 3D printing or the Internet of Things. Technically, ma- hand, estimate the net impact of digitalization on employment for the
chine-based technologies are characterized by a powerful combination whole German economy, using a rigorous structural model. Consistent
of data access, computation and communication technologies with with our findings, they report a small positive net effect. Both their and
acting hardware. This gives machine-based digital technologies their our result run counter to the overly negative prediction by Frey and
disruptive power, and it is these technologies in particular that are Osborne (2017), which has received much attention by politicians and
supposed to drive the current industrial revolution (Rifkin, 2013). the popular press although it does not hold up to more rigorous ap-
Our empirical analysis is based on representative data from two proaches (Arntz et al., 2017). Apart from a contribution the scientific
surveys.3 We exploit firm-level information on investment in literature, our study thus contributes to the public debate on how di-
gitalization influences job destruction and creation. In line with prior
literature on job polarization (e.g. Autor et al., 2003; Goos and
2
While the proportion of jobs at risk is even higher (57%) if all Western Manning, 2007; Oesch and Rodrıguez Menes, 2011) our study suggests
economies are taken into account (Schwab, 2016) the numbers reflect only to expect heterogeneous employment effects across skill groups. It
theoretical potentials and appear to be strongly upward biased. Once taking supports former findings of increased demand for high-skilled labor and
into account workers’ ability to adapt and switch tasks within jobs, these figures lower demand for medium-skilled labor but differs with respect to low-
decline strongly, see e.g. Arntz et al. (2017). Second, these figures solely reflect skilled labor. While former studies showed that demand for low-skilled
theoretical technological potentials. The employment effects of these theore- workers increased in response to technological changes of the past
tical technological potentials are probably very different, as discussed by Autor decades, we find that firms’ investment in digitalization is associated
(2015).
3 with a decline in firms’ low-skilled workers. Due to the restriction of our
Both surveys are based on the same representative sample of firms with at
empirical setup to a single country and rather small dataset of firms
least 20 full-time equivalent employees and were carried out by the KOF Swiss
Economic Institute in cooperation with the Chair of Work and Organizational
Psychology at the Department of Management, Technology, and Economics at
the ETH Zurich, and the University of Applied Sciences and Arts Northwestern (footnote continued)
Switzerland, School of Applied Psychology in 2016 and by the KOF Swiss analyzed with a different focus elsewhere (e.g. Arntz et al., 2018; Lehmer and
Economic Institute in 2015. Similar data exists for Germany, and has been Matthes, 2017; Arntz et al., 2016).

2
B. Balsmeier and M. Woerter Research Policy 48 (2019) 103765

with at least 20 employees in 2015, it is too early to safely conclude that 3. Empirical investigation
digitalization will have different effects on low-skilled labor than
technological developments of the past. The present study provides an 3.1. Data
early snapshot of a potentially important change though, which de-
serves further attention by future studies. It also points to a largely In order to investigate the consequences of investment in digitali-
neglected heterogeneity across technologies that is often masked by zation for job creation and destruction we draw on two datasets.
more aggregated data. First, we use data from a unique representative survey on the di-
gitalization activities of Swiss firms.5 The main goal of the survey was
to provide a comprehensive overview of the adoption of fairly complex
2. Institutional background - a snapshot of the Swiss economy machine-based and non-machine-based digitalization technologies in
the private sector. The survey includes detailed information on the
Before jumping to the analysis section, it is worthwhile looking at educational attainment of firms’ employees along other firm char-
the institutional background of our empirical setting. Switzerland is a acteristics, e.g. research and development expenditures. The survey was
technologically advanced, small and open economy. It is the most in- carried out in 2016. It followed the same professional standards (sam-
novative country in Europe according to the European Innovation pling techniques, non-response control etc.) that are regularly used for
Scoreboard 2017 and it is also ranked among the most competitive the well-known Community Innovation Surveys, which have been ex-
countries in the world (WF Ranking, 2016–2017, Bris and Cabolis, tensively studied in the past.6
2017). Related to this performance, we observe a rather stable industry Second, we use data from the Swiss Innovation Survey (SIS) 2015,
structure, which is strongly exposed to international competition. which is equivalent to the Community Innovation Survey in the EU
About 53% of manufacturing firms with more than five employees are member states. The data proved to be reliable and robust against sta-
exporters, and the average sales share of exports amounted to 36% in tistical biases in terms of ‘type of questions’ or ‘common response bias’
2014. Also, the service sector is significantly exposed to international (Podsakoff et al., 2003; Keupp and Gassmann, 2013). Although the
competition; 33% of service firms are exporters, and the average sales questionnaire includes qualitative questions, we only use quantitative
share amounts to 12%. While we have observed an increase in the information for the study at hand. Potential imprecision in qualitative
economic importance of the service sector in countries like the U.S., the answers or ‘subjective’ answers should thus not affect this investigation.
Netherlands, and some Scandinavian countries, Switzerland’s manu- Both surveys are based on the KOF enterprise panel, which is a re-
facturing sector has remained a mainstay of the country’s economy; its presentative sample of Swiss firms, comprising 5700 firms in total. The
share in value added only slightly decreased from 28% in 1991 to 25% panel is a stratified random sample of firms with more than five em-
in 2010 (Arvanitis et al., 2014). ployees. Stratification refers to 34 industries and within each industry
The international competitiveness of the Swiss economy is partly to three firm-size classes. It covers the manufacturing, construction, and
due to the effectiveness of the education sector. It contributes to the service sectors. The response rate amounted to 30% in both surveys.7
high employment rate and the large proportion of well-educated For the present study we restricted the sample to firms with at least 20
people. The total employment rate of 80% is the second highest in employees in 2015, because only for these firms were information
Europe (Eurostat, 2016). Switzerland has a comprehensive vocational about their digital investments requested in both surveys.8 Finally, we
training system that is often seen as the backbone of its manufacturing merged both surveys at the firm level based on the unique time-in-
excellence. It provides workers with specialist skills, which makes them variant firm identifier of the KOF enterprise panel, allowing us to as-
a highly sought-after workforce in several sectors. Overall, the Swiss semble information on two points in time, 2014 and 2015, per firm.
employees are well educated. The share of high-skilled employees This approach improves identification (see below) at the expense of a
(tertiary degree) amounts to 29%, the share of employees with a mid- loss in observations of firms with fewer than 20 employees and those
skilled level (secondary school for general education and vocational
education) and the share of low-skilled labor (compulsory school)
amounts to 61% and 10% respectively (FSO (Federal Statistical Office), 5
The digitalization survey was conducted by the ETH Zurich, KOF Swiss
2018a). Economic Institute, the Chair of Work and Organizational Psychology at the
A high employment rate with well-educated staff has its price. Department of Management, Technology, and Economics at the ETH Zurich,
Switzerland ranks second among European countries in terms of sal- and the University of Applied Sciences and Arts Northwestern Switzerland
aries per working hour. The average Swiss employee earned 62.01 CHF (FHNW), School of Applied Psychology (APS).
6
(CHF 1 ≅ USD 1.02 ≅ EUR 0.90) per hour in 2012. Only Norway has For more information about the survey and comprehensive descriptive
higher salaries.4 Although labor costs are very high, the labor share has statistics, see Arvanitis et al. (2017).
7
The authors are part of the team that collected the data and have in-depth
remained stable since the 1980s (Siegenthaler and Stucki, 2015).
knowledge of the questionnaires and the survey process. The response rate is
Given its large manufacturing sector, its technological expertise and
quite satisfactory, given that both surveys are not compulsory. A comprehen-
its well-educated labor force, the Swiss economy should be very re- sive non-response control and targeted recalls significantly increase the survey
ceptive to the adoption of digital technologies. Indeed, 29% of firms use response rates and guarantee sufficient coverage in all cells (34 industries times
robots, 12% use 3D printing, and 10% use Internet-of-Things applica- three firm-size classes). Due to item non-response we lose 145 observations. In
tions (see Appendix, Table A1 for details). Owing to the increasing in- the first-difference setting we do not detect any item non-response for em-
vestment in these and other digital technologies over the last few years, ployment and R&D activities, however, there are missing values of the ex-
many expect the demand for skills to change, which is an important penditures for digital technologies.
8
field of investigation and the main subject of the paper at hand. The company size distribution in the population is much skewed. Like in
many other countries, very small companies are dominating. According to the
enterprise census in Switzerland (provisional data for 2016) 90% of all com-
panies have between 1 and 9 employees (full-time equivalent). 8% have be-
tween 10 and 49 employees, 1.5% have between 50 and 250 employees, and
0.3% are large companies with more than 250 employees (FSO (Federal
Statistical Office), 2018b). It is important to note however, that in total 50% of
4
If we correct for purchasing power, Switzerland – with an hourly wage of the employees work in companies with more than 50 employees (FSO (Federal
36.65 CHF – remains among the three countries with the highest compensation, Statistical Office), 2018c). Hence, our data covers a large share of the economic
ranking third after Norway and Belgium (FSO (Federal Statistical Office), activity in Switzerland, although it is representative only for firms with at least
2016). 20 employees.

3
B. Balsmeier and M. Woerter Research Policy 48 (2019) 103765

that responded only in one survey to all questions of interest.9 Com- Table 1
paring the descriptive statistics of the final sample with the non-merged Descriptive statistics.
cross-sectional data revealed that the restriction to firms observed twice Variable count mean median sd min max
did not cause a sample selection bias (see Tables A2a, A2b and A3 in the
Appendix).10 Investment in 447 6.20 0.68 57.82 0.00 1,204.50
digitalization
Total employees 447 272.80 98.00 1,057.28 20.00 18,965.00
3.2. Measurement of educational attainment High-skilled emp. 447 36.70 7.84 238.62 0.00 4,930.90
Mid-skilled emp. 447 46.45 13.30 130.90 0.00 1,455.52
In order to disentangle potential differences in job creation and Low-skilled emp. 447 175.55 51.20 763.18 0.00 12,896.20
R&D investment 447 11.38 0.00 52.08 0.00 763.02
destruction across jobs with heterogeneous skill requirements we dif-
Machine-based tech. 447 0.62 na na 0.00 1.00
ferentiate between three levels of formal education. The category ‘high-
skilled’ contains workers with professional tertiary education (ISCED97 Notes: This table provides descriptive statistics on all variables used in this
4), including graduates from universities of applied sciences (ISCED 5b) study, reference year 2015. Investment in digitalization and R&D investment
and conventional universities (ISCED97 5a and 6). The category ‘mid- are measured in CHF 100,000 (CHF 1 USD 1.02 EUR 0.90). Machine-based
skilled’ refers to workers with completed upper secondary education tech. indicates firms that employ at least one of the following technologies:
(ISCED97 3) and workers who have finished vocational education and Computerized automated control systems, programmable logical controllers,
training (VET) rather than a general upper secondary education. The rapid prototyping, CNC/DNC machines, autonomous vehicles, 3D printing,
third category (‘low-skilled’) comprises untrained workers and dual radio-frequency identification (RFID), or the Internet of Things. High (mid,
low)-skilled is the number of employees with a high (medium, or low) level of
VET students (ISCED97 1 and 2). VET refers to upper secondary edu-
educational attainment, as described in detail above. Total employment is total
cation that prepares students for labor market entry. Dual VET refers to
amount of full time equivalent employment. Data source: KOF Swiss Economic
upper secondary education that combines classroom education and Institute at ETH Zurich. The descriptive statistics for the single cross-sections
workplace training. These students typically work in the firm for about are presented in the Appendix Table A2a and Table A2b. There are a few firms
three days per week and spend the remaining time in classroom edu- with less than 20 employees in 2014 (see Table A2b). If we drop these ob-
cation. This education usually takes three to four years to complete. servations, the main results do not change.
Since these students cannot be employed like fully trained workers, we
follow the literature and combine them with untrained workers in the by machine-based technologies in contrast to the tasks that are reg-
category ‘low-skilled’ (Bolli et al., 2017). ularly performed by non-machine-based digital technologies. We con-
The descriptive statistics show that firms created on average about sider the following technologies to be complex machine-based: com-
14 high-skilled jobs, while they reduced mid-skilled jobs by about four puterized automated control systems, programmable logistic
employees and low-skilled jobs by about nine employees. In total the controllers, rapid prototyping, computerized numerical control (CNC)
average firm created two jobs in the period under investigation. and direct numerical control (DNC) machines, robots, autonomous
vehicles, 3D printing, and the Internet of Things. Technically, the cru-
3.3. Measurement of digitalization cial difference between these and non-machine-based technologies is
their powerful combination of data access, computation and commu-
Our main explanatory variable is firms’ investment in digitalization. nication technologies with acting hardware. This often gives machine-
For the study at hand we look at changes in investment between 2014 based digital technologies their disruptive power, and it is these tech-
and 2015. The descriptive statistics show that the average firm spent nologies in particular that are supposed to drive the current industrial
about CHF 620,000 (CHF 1 ≅ USD 1.02 ≅ EUR 0.90) on digitalization revolution. Given the technical complexity of machine-based digitali-
and increased their digital investment from 2014 to 2015 by around zation, their implementation is often capital intensive, requires high-
CHF 84,000, i.e. 13.5%. The average firm is of medium size, has 273 skilled labor, and is only profitable if it allows firms to produce at much
employees, and invests about CHF 1.138 million in R&D (Table 1 pro- lower cost per unit or higher quality standards than otherwise possible.
vides further descriptive statistics in levels of 2015. Table A3 in the Here lies another crucial difference compared to non-machine-based
Appendix shows descriptive statistics of the differenced variables and technologies that might sometimes be adopted with the same intentions
corresponding correlations are presented in Table A4). but are in fact often ‘just’ used to facilitate extant single processes rather
In order to get a grasp of the heterogeneous influence of different than to disrupt or significantly change the whole production process.
digital technologies on job creation and destruction we differentiate Consider, for instance, ERP, social media or e-purchasing. For most
between the adoption of fairly complex machine-based digital tech- firms, these technologies are rather easy to adopt and complement
nologies, which are supposed to drive the current industrial revolution, extant processes instead of substituting them. They typically do not
and other non-machine-based technologies, which are regularly change the core production process or set new industry standards of
adopted by a wide range of firms across all industries. It is assumed that product production.
the complexity and challenges of digital technologies and also their Finally, we argue that especially machine-based digital technologies
consequences for employment differ between both categories. This has can be seen as powerful tools that enable workers to raise their pro-
much to do with the complexity of their adoption, but also especially ductivity as long as they are qualified and skilled enough to bring them
with their disruptive power and the tasks that are regularly performed to a fruitful usage. High skilled workers will not only more likely use
these technologies as complements, but will also more likely develop
9
new unforeseen solutions, while low-skilled workers are more likely to
See Appendix A12 for a complete list of questions used in this study. lack the technical know-how required to use these technologies as
10
Switzerland shows a low amount of firm foundations and closures per year.
fruitfully. Further, the latter group of workers has a natural interest not
The latest available figures show that there were 694 closures in 2014 and 845
to propose any change where machine-based technologies substitute for
in 2013 in total; and there were about 700 to 900 firm foundations with more
than 5 employees in the same years. The annual closure rate of firms with at tasks regularly fulfilled by low skilled workers even if it would en-
least 10 employees was 0.3%. Given these very low numbers in the complete counter it. Together with the higher complexity of machine-based as
Swiss economy, we would expect to find 1-2 firm closures in our sample, if we opposed to non-machine based digital technologies the relevance of
could observe them. Although, we cannot rule out that these could have a advanced skills increases while medium to low skill levels face lower
significant impact on the results, the likelihood seems rather low. Since newly demand, leading in turn to heterogeneous impacts on job creation and
founded firms are not included in our sample, the results are at least not directly destruction across skill levels. Although this differentiation is not
influenced by newly founded firms.

4
B. Balsmeier and M. Woerter Research Policy 48 (2019) 103765

always applicable and contradictory examples exist, we believe it is an characteristics by estimating the standard model in differences. Hence,
important distinction within digital technologies and their various we regress the difference of the total number of employees (in a certain
possible impacts on job creation and destruction. skill category) in the second period minus the total number of em-
Table A1 in the Appendix gives an overview of digital technologies ployees (in a certain skill category) in the first period on the corre-
used by the Swiss firms in our sample. Non-machine-based technologies sponding difference in investment in digitalization. It is easy to show
like ERP, e-purchasing, and CRM are used by 79%, 61%, and 51% of the that this strategy is immune to any time-invariant omitted variable bias
sample firms, respectively. The diffusion rates are significantly higher like the regular firm fixed effects estimator (Wooldridge, 2002). For-
in large firms than in medium-sized and small (SME) ones. A note- mally, we estimate the following equation:
worthy aspect is that many firms are combining different types of di-
Skilli = 0 + 1 diginvest_chfi + 2 rndi + 3 empi + i (1)
gital technologies – for instance, the use of ERP to improve resource
planning, and CRM to support customer relationships with specialized where
information. They might also allow employees to access the technolo-
Skill { highskill, midskill, lowskill}
gical infrastructure of the firm via remote access (telework) and share
information within the firm using cloud services.11 In our analyses and i is the firm index. We explicitly control for the change in R&D
below we separately examine firms that have adopted at least one expenses along the change in investment in digitalization as the former
machine-based digital technology and those that only use at least one may well coincide and have similar effects on job creation and de-
non-machine-based technology.12 62% of the firms in our sample had struction. We also control for the total change in jobs to prevent our
used at least one machine-based technology by 2016, while 38% had coefficient of interest 1 from capturing general firm-size changes.
only used non-machine-based technologies (Table A1 in the Appendix). Owing to the estimation in differences we implicitly control for firm age
High-tech manufacturing industries such as machinery, electronic and and unobserved heterogeneity across industries.13 The results do not
optical instruments, and vehicles show slightly higher penetration rates change when we control for the initial level of employment instead of
of machine-based technologies than low-tech manufacturing industries the absolute change. We also find consistent results, if we use the share
such as food/beverages/tobacco and basic metals (see Table A2 in the of workers in a given category instead of the absolute change (see
Appendix for details). There is no firm in the sample that uses no di- Appendix, Table A5). For U.S. firms, it has been argued that imports
gitalization technology at all. This might be a typical outcome when and exports may still confound the estimation because changes in that
firms have at least 20 employees and labor costs are high, which en- regard might trigger investment in digitalization and employment
courages firms across all sectors to automate their production processes. changes alike. Fortunately, we can control for both of these factors but
Generally, we observe a positive correlation between firm size and the only for a smaller group of firms. We thus show these results as part of
rate of adoption of digital technologies. other sensitivity checks in the Appendix, Tables A6 and A7. Since all
results presented below remained qualitatively unchanged after the
inclusion of these controls, and other controls like foreign ownership
3.4. Methodological remarks and technology age (see Table A8a), we focus in our main analyses on
the short model (results on all the extended models are available from
The standard approach to investigate if investment in digital tech- the authors upon request). In alternative regressions we excluded emp
nologies increases or reduces the creation of jobs would be a classic from our framework, or exchanged it with firm sales as a measure of
ordinary least squares (OLS) regression of employment on investment firm size, and find similar results (see Appendix, Table A10a and Table
in digitalization. The main issue with this approach is that we can A10b).
hardly control for all kinds of firm characteristics that may simulta-
neously influence the creation of jobs and investment in digitalization,
4. Results
e.g., firms’ strategy or managerial quality. A straightforward commonly
used improvement over the classic OLS framework is the usage of panel
Table 2 presents the main results. We find broad support for het-
estimators, where one can include firm-specific indicators that control
erogeneous effects of changes in investment in digitalization on changes
for potential omitted variable biases from unobserved time-invariant
in employment of workers with different skills measured by their
firm characteristics. All those estimators require a strong panel com-
educational attainments. Specifically, we estimate a statistically sig-
ponent, though, i.e. information on firms observed over a long time
nificant (p < 0.01) increase in high-skilled employment and a statis-
span. In the absence of such a panel, but with information on two points
tically significant decrease in medium- to low-skilled jobs per CHF in-
in time, we can effectively control for unobserved time-invariant firm
vested in digitalization. The magnitudes of the estimated effects are
economically meaningful. Column 1, Table 2, suggests that a CHF
11
Technological complexity likely increases with the use of different types of 100,000 increase in investment in digitalization is associated with
technologies. Problems with interfaces, difficulties in programming/adapting about 5.8 more jobs for highly educated workers, four fewer jobs for
standard software to the specific technological infrastructure of a firm, or mid-skilled workers (Column 2) and about 2.3 fewer jobs for the low-
training employees to efficiently use new software is likely to increase the skilled (Column 3). In total, Column 4 suggests an increase of 1.6 jobs
technological complexity within a firm and might require new skills. In un-
reported analyses, we tried to measure technological complexity with the
13
number of digital technologies used by a firm. We found that the average firm A classical labor demand function also envisages controls for labor costs
in our sample uses seven different digital technologies and that most firms that and capital costs. Since we use first differences in our estimations, we can drop
use machine-based technologies also use others. We do not differentiate be- such controls for the following reasons. First, we look at a very short time
tween machine-based firms that use no non-machine-based digital technology period (2014 and 2015) with very stable economic conditions in Switzerland.
and those firms that use only machine-based technologies, because both groups There have been no changes in capital costs; the relevant interest rate (mean
of firms do not show significant differences in terms of job creation and de- level of mortgages) changed by only 0.03%-points (from 2.69% to 2.66%) be-
struction. This is consistent with the finding that investments in digitalization tween January 2014 and December 2015 (see Swiss National Bank, 2017). Also,
by firms without machine-based digital technologies show no significant re- labor costs hardly changed, increasing by 0.4% between 2014 and 2015 (see
lationship with job creation and destruction. Results are available from the FSO (Federal Statistical Office), 2017). Second, since capital costs do not vary
authors upon request. much among industries, we address them with our first-difference approach.
12
All results presented below are robust to excluding any single technology Wage data might cause serious endogeneity problems in single-equation set-
while keeping the others. Information on the amount of capital invested in tings; hence, some authors even propose omitting wage data (see e.g. Machin
specific digital technologies is unfortunately not available. and Van Reenen, 1998; Arvanitis and Loukis, 2015).

5
B. Balsmeier and M. Woerter Research Policy 48 (2019) 103765

Table 2 (models 2 and 5) are insignificantly different though (p = 0.264), and

Investment in digitalization and job creation and destruction. the coefficients in the low-skilled employee regressions (models 3 and
(1) (2) (3) (4) 6) only pass the one-sided test (p = 0.071).16 This result is broadly
Variables High-skilled Mid-skilled Low-skilled Total consistent with recent findings from the U.S., where the impact of the
employees adoption of robots has been analyzed (Acemoglu and Restrepo, 2017b),
but it differs in terms of the net impact our results imply.17
Investment in 5.841*** −4.001** −2.307*** 1.597***
digitalization
These results fit the view that the strongest impact of digital tech-
(1.812) (1.902) (0.399) (0.401) nologies on job creation and destruction materializes where machines
Investment in R& 0.033 −0.033 0.001 0.003 are empowered by data access, computation and communication
D technologies. This is also the area where technological advances have
(0.024) (0.028) (0.006) (0.032)
been the most dynamic over the last few years, and this is supposed to
Total employees 0.274** 0.032 0.660***
(0.129) (0.121) (0.101) be the main driver of the current industrial revolution. The fact that we
Observations 447 447 447 447 do not find significant employment effects for firms that have adopted
Firm fixed effects Yes Yes Yes Yes only non-machine-based technologies may not be surprising, given that
R2 0.689 0.348 0.398 0.117
previous studies have found that ICT investment does not have major
impacts on either employment or productivity (e.g. Acemoglu et al.,
Notes: All models are estimated using OLS. All variables are measured in ab-
solute differences between 2015 and 2014. High (mid, low)-skilled is the ab- 2014; Chen and Xie, 2015).18
solute change in employees with a high (medium, or low) level of educational The change in total employment is significant positively associated
attainment, as described in detail above. Investment in digitalization is the with an increase in high- and low-skilled labor, suggesting that most of
absolute change in investment in digitalization, measured in CHF 100,000. the general hiring of employees happens in these categories.
Investment in R&D is the absolute change in investment in R&D, also measured
in CHF 100,000. Total employment is the change in total employment.
Heteroscedasticity-robust standard errors are shown in parentheses. ***, **, * 4.1. IV regression
indicate statistical significance at the 1%, 5%, and 10% level, respectively. Data
source: ETH Zurich, KOF Swiss Economic Institute. While we effectively control for unobserved time-invariant hetero-
geneity plus the main suspects of time-variant covariates in our pre-
per CHF 100,000 invested in digitalization. The difference between this vious models, some unobserved time-varying factors might still con-
estimate and the total of the estimates from models 1 to 3 is due to the found the estimation. Potential candidates are unobserved changes in
fact that we control for the total size increase in models 1 to 3.14 It firm strategies, e.g. changing to a more open innovation strategy that
suggests digitalization has a small net positive impact on employment enlarges the collaboration network. This might increase high-skilled
in Switzerland at least in the short run, which is consistent with recent jobs and reduce lower-skilled jobs and trigger additional investment in
findings derived from a structural model based on German data (Arntz digital technologies to improve communication within the extended
et al., 2018). network. If such changes were common, the previously estimated po-
Depending on the specific situation and technology employed, the sitive effect of investment in digitalization on high-skilled job creation
employment effects likely reflect the need for new personnel to install and the negative effect on low-skilled jobs would be overestimated.
and implement the acquired assets, change the production process, or To address these concerns we turn to instrumental variable regres-
to bring new services and products to the market. As individuals who sions. In order to derive consistent estimates from an IV regression, our
work on these tasks typically need to be trained in advance, the hiring instrument needs to fulfill the validity requirements, i.e. it needs to
of new workers may start before the actual adoption of the digital explain the change in investment in digitalization while being un-
technologies, resulting in a rather quick response of employment fig- correlated with changes in the number of jobs created in different skill
ures to investments in digitalization. The decline in low and medium categories. In the present empirical setting we exploit the fact that the
skilled employment is unlikely to reflect dismissals due to the strong main ICT network provider in Switzerland was in the middle of its
employment protection in Switzerland. Negatively affected employees, expansion of the high-speed broadband network right before the sample
however, will either learn about negative career prospects or will be starts. We considered an internet connection to be high speed if it was
informed in advance of the actual investment. Those workers may thus based on fiber-to-the-home technology, vectoring technology, or al-
search for new jobs with better career prospects rather quickly, such lowed download speeds of more than 100 Mbit/s. We obtained data on
that negative employment effects materialize within a short time frame the regional expansion of the Swisscom19 fiber and vector network for
and despite strong institutional job protection. 2014, which by then covered about 60% of the firms in our sample.
If we separately consider firms that employ at least one machine- Since it is the predominant – but not the only – way to get a fast internet
based digital technology (Table 3, Columns 1–3) and the other firms connection in Switzerland, we supplemented it with survey data on
(Table 3, Columns 4–6), we find that our main results are entirely whether firms would actually have access to high-speed internet even
driven by the group of firms that employ at least one machine-based
technology.15 The impact of digital investment on high-skilled em- 16
The two-sided test reveals a p-value of 0.1423.
ployees is not only insignificant for firms without machine-based digital 17
Since we take into account not only robots but also other applications like
technologies but also statistically significantly different from the coef-
the Internet of Things, it may not be surprising that the results differ somewhat.
ficient derived for firms that adopted at least one machine-based digital All results presented in Table 3 are even more pronounced in terms of statistical
technology (two-sided t-test, p < 0.003). This suggests that different significance if we only consider firms that use robots among those that have
types of digital technologies do indeed exert different impacts on em- adopted a machine-based digital technology. We also note differences in the
ployment. The coefficients in the mid-skilled employee regressions institutional environments of both settings, which might play an important role
too (see also the discussion below).
18
Arvanitis and Loukis (2009) found small positive effects of ICT investment
14
In alternative unreported regressions we used the share of skilled, on the productivity of Swiss firms in 2004 but did not consider employment
medium-, and low skilled labor as alternative dependent variables. The size of effects. They did not control for unobserved time-invariant heterogeneity ei-
the differences in shares resembles the estimated change in absolute employ- ther, which might have revealed different findings.
19
ment for the average firm but have a weaker statistical significance, see Table Swisscom is the incumbent in the Swiss telecommunications market and
A5 in the Appendix. the dominant provider of fixed-line telecommunications infrastructure. Sunrise
15
The fully interacted model is presented in the Appendix, Table A9. and Cablecom are the main, albeit significantly smaller, competitors.

6
B. Balsmeier and M. Woerter Research Policy 48 (2019) 103765

Table 3
Investment in digitalization - machine-based versus non-machine-based.
(1) (2) (3) (4) (5) (6)
Machine-based Non-machine-based

Variables High-skilled Mid-skilled Low-skilled High-skilled Mid-skilled Low-skilled

Investment in digitalization 6.247*** −4.046** −2.728*** 0.147 −1.178 1.048

(1.581) (1.840) (0.377) (0.174) (1.307) (1.327)
Investment in R&D 0.032 −0.034 0.002 0.226 0.238 −0.530
(0.022) (0.024) (0.005) (0.299) (1.077) (1.235)
Total employees 0.368** −0.124 0.734*** −0.102 0.585*** 0.458***
(0.167) (0.149) (0.111) (0.093) (0.063) (0.133)
Observations 277 277 277 170 170 170
Firm fixed effects Yes Yes Yes Yes Yes Yes
R2 0.762 0.458 0.601 0.165 0.189 0.136

Notes: All models are estimated using OLS. All variables are measured in absolute differences between 2015 and 2014. Firms are considered to be using machine-
based digitalization technologies if they employ at least one of the following technologies: Computerized automated control systems, programmable logical con-
trollers, rapid prototyping, CNC/DNC machines, autonomous vehicles, 3D printing, radio-frequency identification (RFID), or the Internet of Things. Firms are
considered to be merely using non-machine-based digitalization technologies if they do not employ any of the former machine-based technologies but use at least one
of the following technologies: ERP, CRM, SCM systems, collaboration support systems, social media, e-commerce, or telework. High (mid, low)-skilled is the absolute
change in employees with a high (medium, or low) level of educational attainment, as described in detail above. Investment in digitalization is the absolute change in
investment in digitalization, measured in CHF 100,000. Investment in R&D is the absolute change in investment in R&D, also measured in CHF 100,000. Total
employment is the change in total employment. Heteroscedasticity-robust standard errors are shown in parentheses. ***, **, * indicate statistical significance at the
1%, 5%, and 10% level, respectively. Data source: ETH Zurich, KOF Swiss Economic Institute.

without Swisscom network coverage. Based on this information, we bit compared to the OLS results is not surprising because of the two step
defined our instrument as being a dummy that has a value of one if a procedure, which makes estimates less precise compared to OLS, all else
firm is covered by the Swisscom network or has another type of high- equal. The magnitudes of the coefficients estimated with OLS and IV are
speed internet access, and is zero otherwise. statistical insignificant, although the sizes of the coefficients nominally
We would argue that a fast internet connection should have trig- differ quite a bit with respect to medium and low-skilled labor. Under
gered investment in digitalization because most digital technologies the assumption that our IV estimates are not biased, the nominally
require such infrastructure in order to be implemented and used ef- lower effect of investment in digitalization on medium skilled labor
fectively. Indeed, we find this view supported by the first stage of our IV might be the result of unobserved training programs that are offered to
regression (see Table A11 in the Appendix). The instrument is in- medium skilled employees at the same time when changes caused by
dividually significant, but the strength is lower than expected. The F- the adoption of digital technologies apply. The seemingly under-
test on individual instrument strength reveals a value of 3.02 (3.83 in estimated effect of changes in low-skilled labor in the OLS case might
the sub-sample of machine-based digital technology adopters), while it stem from unobserved severance payment offers to this group of
is generally recommended to have a strength of ten or more to be on the workers, potentially because it is harder to retrain low-skilled workers
safe side (Stock and Yogo, 2005). The rather weak instrument calls for a than medium skilled works to fulfill new tasks created in the wake of
very cautious interpretation because IV regressions can be even more the digital transformation.
biased than OLS regressions in such cases (see Young, 2018, for an Overall, our results suggest that investment in digitalization does
extend discussion of weak instrument biases and violation of the ex- indeed have a significant impact on job creation and destruction. The
clusion restriction). Since we also have only one instrument, we cannot impact is, however, heterogeneous across skill levels and technologies –
formally test the exclusion restriction, i.e. whether the instrument is something that has received little attention in the literature so far.
indeed uncorrelated with the error term of the second stage. The fact
that it is not directly related to our depended variable, or other firm 4.2. Limitations and discussion
characteristics, suggests that it may still work. We acknowledge that it
is not perfect, and it could be correlated with some unobserved regional Although we carried out an extensive set of robustness and sensi-
characteristics that are in turn correlated with firms’ investment in di- tivity checks, some limitations that warrant attention remain, partly
gitalization and/or job creation. One noteworthy aspect is that such because they give rise to fruitful future research venues. First of all,
correlations may formally violate the exclusion restriction but do not while we link investment in digitalization to employment across dif-
necessarily cause huge biases (Conley et al., 2012). Overall, a very ferent skill levels within firms, unfortunately we cannot observe how
careful assessment of the IV results is warranted although our results much has been invested in specific digital technologies and when those
based on the previously estimated models generally hold up in this technologies were actually adopted. Appendix Tables A8b to A8d ac-
setup, as Table 4 shows. The coefficients are in all models of high- to tually show that our findings are more pronounced the more technol-
low-skilled jobs similar to those estimated previously. The statistical ogies are used for more than two years, indicating that the long term-
significance drops quite a bit in the full sample (models 1 to 3), where effects might be stronger than the short-term effects and that tech-
only the positive impact of investment in digitalization on high-skilled nology age might be a crucial factor driving our results. Further, as our
jobs remains significant at conventional levels. In the sub-sample of results suggest that differentiation among digital technologies is crucial,
firms that have adopted machine-based digital technologies we still find we would not be surprised to see heterogeneous effects even within the
the impact to be statistically significant (p < 0.05) on high- as well as machine-based technologies and different kinds of non-machine-based
low-skilled employment.20 That the statistical significance drops quite a technologies. Heterogeneity in employment effects may also result from
the fact that technologies are applied for different purposes, at different
stages of the production process, and at different stages of their ma-
20
Similar to the standard OLS results presented above we do not find any turity. For a deeper understanding it would thus be invaluable to ex-
significant effects in the sub-sample of firms that do not use any machine-based amine where and how exactly each digitalization technology is applied.
digital technology. Results are available from the authors upon request. We conjecture that the mechanisms through which each technology

7
B. Balsmeier and M. Woerter Research Policy 48 (2019) 103765

Table 4
IV regressions – full sample and machine-based digital technology adopters only.
(1) (2) (3) (4) (5) (6)
Full sample Machine-based

Variables High-skilled Mid-skilled Low-skilled High-skilled Mid-skilled Low-skilled

Investment in digitalization (est.) 6.539** −4.184 −2.587 6.449*** −2.525 −4.269**

(2.550) (3.406) (2.628) (2.398) (3.119) (1.987)
Investment in R&D 0.032 −0.033 0.001 0.032 −0.035** 0.003
(0.027) (0.028) (0.006) (0.023) (0.018) (0.008)
Total employees 0.224 0.046 0.681*** 0.350 −0.257 0.868***
(0.168) (0.240) (0.199) (0.214) (0.272) (0.150)
Observations 447 447 447 277 277 277
Firm fixed effects Yes Yes Yes Yes Yes Yes
R2 0.682 0.347 0.395 0.762 0.409 0.490

Notes: The table presents results of IV regressions. The first stage is an OLS regression of Δinvestment in digitalization on a dummy indicating whether the firm had
access to high-speed internet, and zero otherwise. The F-value of individual instrument strength is 3.02 in models 1–3 and 3.83 in models 4–6. All other variables are
measured in absolute differences between 2015 and 2014. Firms are considered to be using machine-based digitalization technologies if they employ at least one of
the following technologies: Computerized automated control systems, programmable logical controllers, rapid prototyping, CNC/DNC machines, autonomous ve-
hicles, 3D printing, RFID, or the Internet of Things. High (mid, low)-skilled is the absolute change in employees with a high (medium, or low) level of educational
attainment, as described in detail above. Investment in digitalization is the estimated absolute change in investment in digitalization, measured in CHF 100,000.
Investment in R&D is the absolute change in investment in R&D, also measured in CHF 100,000. Total employment is the change in total employment.
Heteroscedasticity-robust standard errors are shown in parentheses. ***, **, * indicate statistical significance at the 1%, 5%, and 10% level, respectively. Data source:
ETH Zurich, KOF Swiss Economic Institute.

contributes to job creation and destruction will be quite heterogeneous. education system to prevent further employment losses, especially if it
As such, industry-, technology- or even firm-specific analyses that focus turns out that the observed reduction is not temporary.
on one specific technology such as robots or autonomous cars would Fourth, it should be emphasized that our empirical setting refers to a
appear to be fruitful avenues for future research. When such research is highly developed country that is often considered to be among the
carried out, attention should also be given to measuring the market technological leaders in the world, while also having one of the best-
impact of digital technologies. educated workforces with quite a unique educational system. The latter
Second, we need to keep in mind that our sample covers a specific in particular should give us pause for thought when extrapolating the
time period when the business cycle was in an upswing and digital presented results to other countries of the world. Because of the high
technologies were considered to be much more disruptive than during wages and liberal migration policy it should be easier for Swiss firms to
previous years. Recent advances in economic theory suggest that the hire highly educated workers than for many firms situated outside
observed spread in demand for high-skilled and low-skilled employees Switzerland. The expected returns to firms’ investment in digitalization
is part of the digitalization process and is expected to be most pro- may thus be much higher than at other places in the world.
nounced in the middle of the transition phase from a non-digitalized Furthermore, the specific institutional background may cause Swiss
economy to a fully digitalized one (Acemoglu and Restrepo, 2017a). firms to invest in different kinds of digital technologies and with dif-
This means that our estimates may change over the course of the next ferent aims than firms in other countries. Given the high salaries in the
few years towards a more mature technological environment. The low-skilled sector, for instance, investment in automation may generate
sample is further limited by its restriction to firms with at least 20 larger pay-offs than in countries with relatively cheap labor.
employees. The relationship between investment in digitalization and Although we have no systemic evidence, it appears that the current
job creation might thus look different for very small firms in general rise of digital technologies may exert similar effects on job creation and
and the increasingly popular entrepreneurial firms in specific. It should destruction as previous technological disruptions. It is striking that,
also be mentioned again that our IV regression is far from perfect, such consistent with our results, most modern jobs require a higher skill level
that a very careful interpretation is warranted. There might still be than many jobs of the past, especially those commonly found before the
unobserved time-varying factors that are correlated with investment in second industrial revolution. One of the reasons why the second in-
digitalization and employment in different skill groups, which might dustrial revolution did not cause mass unemployment in the long run
cause our estimates to be biased up- or downwards, respectively. Po- (contrary to prominent assessments of the time) may be that the ca-
tential candidates are changes to a more open innovation strategy that pacities of the educational sector increased accordingly, ensuring that
enlarges the collaboration network. This might increase high-skilled the new demand for high-skilled labor is met. There might be reasons to
jobs and reduce lower-skilled jobs and trigger additional investment in believe this time could be dramatically different21, but the slightly
digital technologies to improve communication within the extended positive net effect that we have identified in the present paper might
network. If such changes were common, the previously estimated po- well be taken as a reason to feel less pessimistic about the changes to
sitive effect of investment in digitalization on high-skilled job creation come. We acknowledge, though, that it is in fact not possible to em-
and the negative effect on low-skilled jobs would be overestimated. pirically identify the true net effect without a structural model.
Third, vocational (dual) education is an important part of the Swiss
education system, and officials quite frequently assume that it accounts
for the country’s low unemployment rates. The presented results and 5. Conclusions and policy implications
other empirical evidence suggest that the digitalization process reduces
employment not only for low-skilled workers but also, at least to a The emerging digitalization of the economy is expected to have a
lesser extent, for mid-skilled employees as well (Siegenthaler and tremendous impact on several economic parameters, particularly on the
Stucki, 2015). Given that the private sector is strongly involved in the
design of courses, ensuring great flexibility and practical relevance for 21
One could argue that this time is dramatically different because machines
training, the digital challenges might be beyond the system’s capacity to pose a challenge for intelligent work rather than mainly manual labor as in the
adapt (Renold et al., 2016). It may require a redesign of the vocational past.

8
B. Balsmeier and M. Woerter Research Policy 48 (2019) 103765

number of jobs and the skills required to perform those jobs. The pre- There are three main objectives that policymakers could pursue to
sent paper has addressed this topic with econometric analyses based on address these challenges. First, it is of utmost importance for the
unique firm-level microdata from Switzerland, one of the technological competitiveness of a technologically advanced country to ensure that
leaders of the world. We found robust evidence on the relationship skilled labor is efficiently allocated to growing, productive sectors. This
between changes in investment in digital technologies within the firm implies flexible labor markets and workable product market competi-
and employment changes within the firm, explicitly considering het- tion to provide incentives to invest in innovation and the adoption of
erogeneity in terms of different skill levels and different types of digital new technologies. For small countries in particular it is important that
technologies. the worldwide free movement of talent is guaranteed and the hiring
Specifically, we found that investment in digital technologies is process is accompanied by low administrative costs. Second, training
positively associated with employment of high-skilled workers and and continuing professional development is necessary to improve the
negatively associated with employment of low-skilled workers, with an match between skills and job requirements. This includes training not
overall positive net effect on employment. These effects are entirely only in technical skills but also soft skills and emotional skills to
driven by machine-based digital technologies (robots, 3D printing, the manage the digital transition of labor markets. Since the ability to learn
Internet of Things) that are supposed to fuel the current industrial re- and retrain skills is positively correlated with educational attainment,
volution. We find no significant impact of investment in digitalization governments should increase the attractiveness of tertiary education.
when only firms that adopt non-machine-based digital technologies, Third, some parts of society might not be able (or willing) to success-
such as ERP, e-commerce or cooperation support systems, are con- fully manage the transition to a digital age. This could be addressed by
sidered. the development of collaborative artificial intelligences as mentioned
This evidence improves our understanding of the potential eco- above or innovative social measures such as an unconditional basic
nomic implications of the digitalization process, where digital tech- income.
nologies are constantly changing methods of production and constantly
generating new products and services. The breakthrough of machine- Appendix A. Supplementary data
based digital technologies required better technological infrastructure
(faster internet connections, ubiquity of internet access, cloud tech- Supplementary material related to this article can be found, in the
nologies, etc.), the skills to develop such technologies, and sufficiently online version, at doi:https://doi.org/10.1016/j.respol.2019.03.010.
skilled workers to use them effectively. Meeting these requirements
over recent years has accelerated the transformation of the economy References
into the digital age. This transformation process does not come without
frictions, though. While our estimations imply that total employment Acemoglu, D., Restrepo, P., 2017a. The Race Between Man and Machine: Implications of
will slightly grow in the current transition period, the average masks an Technology for Growth, Factor Shares and Employment. Working Paper.
Acemoglu, D., Restrepo, P., 2017b. ). Robots and Jobs: Evidence From US Labor Markets.
increase in employment of skilled labor and a decrease in employment Working Paper.
of unskilled labor. The result runs counter to the unconditional state- Acemoglu, D., Dorn, D., Hanson, G.H., Price, B., 2014. Return of the Solow paradox? IT,
ment that workers are in a race against machines – at least in the short productivity, and employment in US manufacturing. Am. Econ. Rev.: Papers Proc.
104, 394–399.
run and for technologically advanced, open economies with a large Arntz, M., Gregory, T., Lehmer, F., Matthes, B., Zierahn, U., 2016. Arbeitswelt 4.0 - Stand
share of high-skilled labor, like Switzerland. This has implications for der Digitalisierung in Deutschland: Dienstleister haben die Nase vorn. IAB-
corporate as well as political leaders. Kurzbericht Nr. 2.
Arntz, M., Gregory, T., Zierahn, U., 2017. Revisiting the risk of automation. Econ. Lett.
The global war for talent is likely to intensify further. Companies 159, 157–160.
should therefore prepare to spend more time and resources on re- Arntz, M., Gregory, T., Zierahn, U., 2018. Digitalisierung und die Zukunft der Arbeit:
cruiting and retaining skilled employees. In times of high demand and Makroökonomische Auswirkungen auf Beschäftigung, Arbeitslosigkeit und Löhne von
morgen. Bundesministerium für Forschung und Entwicklung (BMBF), Mannheim.
an increasingly mobile workforce, even well-situated companies might
Arvanitis, S., Loukis, E.N., 2009. Information and communication technologies, human
face strong competition for their most talented employees in the future. capital, workplace organization and labour productivity: a comparative study based
Many firms in Switzerland, a country with one of the highest propor- on firm-level data for Greece and Switzerland. Inf. Econ. Policy 21, 43–61.
tions of well-educated people, argue already that a shortage of highly Arvanitis, S., Loukis, E., 2015. Employee education, information and communication
technologies, workplace organization, and trade: a comparative analysis of Greek and
qualified workers would limit their growth. Swiss firms. Ind. Corp. Chang. 24 (6), 1417–1442.
In combination with a decrease of jobs for low-skilled workers, in- Arvanitis, S., Veseli, K., Woerter, M., 2014. Die Tertiärisierung der Volkswirtschaft setzt
equality within the population is likely to increase. From an economic sich fort. Die Volkswirtschaft Nr.3/2014.
Arvanitis, S., Grote, G., Spescha, A., Wäfler, T., Woerter, M., 2017. Digitalisierung in der
perspective, it is thus crucial to develop and apply instruments that Schweizer Wirtschaft: Ergebnisse der Umfrage 2016. KOF Studie Nr. 93, Juni. ETH
minimize potential negative impacts while nurturing positive effects. As Zürich, KOF Konjunkturforschungsstelle, Zürich.
long as medium- and low-skilled workers can be trained to learn new Aubert-Tarby, C., Escobar, O.R., Rayna, T., 2018. The impact of technological change on
employment: the case of press digitisation. Technol. Forecast. Soc. Change 128,
skills that enable them to take over new tasks created in the wake of the 36–45.
digitalization process, the promotion of such professional training Autor, D., 2015. Why are there still so many jobs? The history and future of workplace
programs could be very helpful for the affected individuals and the automation. J. Econ. Perspect. 29 (3), 3–30.
Autor, D., Levy, F., Murnane, R.J., 2003. The skill content of recent technological change:
economy alike. Professional training and educational attainment pro-
an empirical exploration. Q. J. Econ. 118, 1279–1333.
grams should not be seen as an all-purpose medicine, however, since Balsmeier, B., Delanote, J., 2015. Employment growth heterogeneity under varying in-
physical or mental limitations may constrain their effective usage. In tellectual property rights regimes in european transition economies: young vs. mature
innovators. J. Comp. Econ. 43, 1069–1084.
such cases different instruments need to be provided, and digital
Blanchflower, D., Burgess, S., 1999. New technology and jobs: comparative evidence from
technologies themselves might actually offer a solution. Collaborative a two country study. Econ. Innov. New Technol. 5, 109–138.
artificial intelligences, for instance, might enable medium and low- Bolli, T., Renold, U., Woerter, M., 2017. Vertical educational diversity and innovation
skilled workers to focus on tasks where they have a comparative ad- performance. Econ. Innov. New Technol (in print).
Bris, A., Cabolis, C., IMD World Digital Competitiveness Ranking 2018, 2017.
vantage over machines, e.g. any work that requires personal emotional Brynjolfsson, E., McAfee, A., 2016. The Second Machine Age. W. W. Northon Company
interaction, while machines substitute for personal weakness, e.g. in Inc.
predicting prices and recognizing patterns. This would increase the Chandler, A.D., 1977. The Visible Hand: The Managerial Revolution in American
Business. Harvard University Press, Cambridge, MA.
individual productivity of medium and low-skilled workers, which Chen, S., Xie, Z., 2015. Is China’s e-governance sustainable? Testing Solow IT productivity
should in turn create new job opportunities. If such efforts fail, in- paradox in China’s context. Technol. Forecast. Soc. Change 96, 51–61.
equality among workers is likely to rise, which poses a great challenge Conley, T.G., Hansen, C.B., Rossi, P.E., 2012. Plausibly exogenous. Rev. Econ. Stat. 94 (1),
260–272.
for public institutions and policymakers.

9
B. Balsmeier and M. Woerter Research Policy 48 (2019) 103765

Dengler, K., Matthes, B., 2015. Folgen der Digitalisierung für die Arbeitswelt. (8), 729–741.
Substituierbarkeitspotenziale von Berufen in Deutschland. IAB - Forschungsbericht, Michaels, G., Natraj, A., Van Reenen, J., 2014. Has ICT polarized skill demand? Evidence
11-2015. Nürnberg, Germany. . from eleven countries over twenty-five years. Rev. Econ. Stat. 96 (1), 60–77.
Entorf, H., Pohlmeir, W., 1990. Employment, innovation and export activity. In: Florens, Mokyr, J., 1990. The Lever of Riches: Technological Creativity and Economic Progress.
J., Ivaldi, M., Laffont, J.J., Laisney, F. (Eds.), Microeconometrics: Surveys and Oxford University Press, New York.
Applications. Oxford Blackwell, pp. 394–415. Oesch, D., Rodrıguez Menes, J., 2011. Upgrading or polarization? Occupational change in
Eurostat, 2016. Employment and Activity, Total Employment Resident Population Britain, Germany, Spain and Switzerland, 1990-2008. Socioecon. Rev. 9 (3),
Concept. Labour Force Survey, Last Update 11.10.2017. Luxembourg. . 503–531.
Ford, M., 2015. The Rise of the Robots. Basic Books, New York. Pianta, M., 2006. Innovation and employment. In: Fagerberg, J., Mowery, D.C. (Eds.), The
Frey, C.B., Osborne, M.A., 2017. The future of employment: how susceptible are jobs to Oxford Handbook of Innovation. Oxford University Press, Oxford, United Kingdom,
computerization? Technol. Forecast. Soc. Change 114, 254–280. pp. 568–598 chapter 21.
FSO (Federal Statistical Office), 2016. Arbeitskosten je geleistete Stunde im Produktions- Podsakoff, P., MacKenzie, S., Lee, J., Podsakoff, N., 2003. Common method biases in
und Dienstleistungssektor – Vergleich Schweiz – Europäische Union (EU) – EFTA- behavioral research: a critical review of the literature and recommended remedies. J.
Staaten; in Euro und KKS. Schweizer Lohnstrukturerhebung, Neuchâtel, Schweiz. Appl. Psychol. 88 (5), 879–903.
FSO (Federal Statistical Office), 2017. Schweizerischer Lohnindex: Index und Rayna, T., Striukova, L., Darlington, J., 2015. Co-creation and user innovation: the role of
Veränderungen auf der Basis 2010(=100) (NOGA 2008). Neuchâtel, Schweiz. . online 3D printing platforms. J. Eng. Technol. Manag. 37, 90–102.
FSO (Federal Statistical Office), 2018a. Educational Attainment of the Population by Renold, U., Bolli, T., Caves, K., Buergi, J., Egg, M.E., Kemper, J., Rageth, L., 2016.
Occupation Pursued (ISCO Main Occupational Group) and Highest Level of Education Feasibility Study for a Curriculum Comparison in Vocational Education and Training
Completed, 2017. Swiss Labour Force Survey (SAKE). Intermediary Report II: Education-employment Linkage Index. KOF Studies No. 80.
FSO (Federal Statistical Office), 2018b. STATENT Ergebnisse 2005 - 2016, Beschäftigte ETH Zurich, Zurich.
und Vollzeitäquivalente nach Wirtschaftszweigen auf Ebene Arbeitsstätten und nach Rifkin, J., 2013. The Third Industrial Revolution: How Lateral Power Is Transforming
Kantonen. Federal Statistical Office, Neuchâtel 2018. Energy, the Economy, and the World. Macmillan, USA.
FSO (Federal Statistical Office), 2018c. Struktur der Schweizer KMU 2016. Neuchâtel, Schwab, K., 2016. The Global Competitiveness Report 2016–2017. World Economic
2018. . Forum.
Goos, M., Manning, A., 2007. Lousy and lovely jobs: the rising polarization of work in Siegenthaler, M., Stucki, T., 2015. Dividing the pie: firm-level determinants of the labor
Britain. Rev. Econ. Stat. 89 (1), 118–133. share. Ind. Labor Relat. Rev. 68 (5), 1157–1194.
Goos, M., Manning, A., Salomons, A., 2009. Job polarization in Europe. Am. Econ. Rev.: Smolny, W., 1998. Innovation, prices and employment: a theoretical model and an ap-
Pap. Proc. 99 (2), 58–63 2009. plication for West German manufacturing firms. J. Ind. Econ. 46, 359–381.
Keupp, M.M., Gassmann, O., 2013. Resource constraints as triggers of radical innovation: Stock, J., Yogo, M., 2005. Testing for weak instruments in linear IV regression. Andrews
longitudinal evidence from the manufacturing sector. Res. Policy 42 (8), 1457–1468. DWK Identification and Inference for Econometric Models. Cambridge University
Kogan, L., Papanikolaou, A., Seru, A., Stoffman, N., 2017. Technological innovation, re- Press, New York.
source allocation and growth. Forthcoming: Q. J. Econ. 132 (2), 665–712. Swiss National Bank, 2017. Published Interest Rates for Variable Mortgages 2014/
Landes, D., 1969. The Unbound Prometheus. Cambridge University Press, New York. 1–2015/12. Zurich. Switzerland. .
Lehmer, F., Matthes, B., 2017. Auswirkungen der Digitalisierung auf die Van Reenen, J., 1997. Employment and technological innovation: evidence from U.K.
Beschäftigungsentwicklung in Deutschland. Institut für Arbeitsmarkt- und Manufacturing firms. J. Labor Econ. 15, 255–284.
Berufsforschung, Aktuelle Berichte 05/2017. Wooldridge, J.M., 2002. Econometric Analysis of Cross Section and Panel Data. The MIT
Leontief, W., 1952. Machines and Man. Scientific American. Press, Cambridge, MA.
Machin, S., Van Reenen, J., 1998. Technology and changes in skill structure: evidence Young, A., 2018. Consistency Without Inference: Instrumental Variables in Practical
from seven OECD countries. Q. J. Econ. 18, 1215–1243. Application. Unpublished manuscript. London School of Economics and Political
Mastrostefano, V., Pianta, M., 2009. Technology and jobs. Econ. Innov. New Technol. 18 Science, London.

10