Report - How Robots Change the World_watermark

Historical evidence from the US labour market suggests that manufacturing workers typically turn to service-sector employment when their old production jobs disappear. Analysing the job moves of over 35,000 US workers over the course of their careers (using longitudinal household data dating back to 1995), we find that more than half the workers who left production jobs were absorbed into just three occupational categories: transport; construction and maintenance; and office and administration work (see Fig. 12).

But according to our research, these three occupational categories are among

the most vulnerable to automation over the next decade. This highlights the

ongoing threat to workers who are at highest risk to automation as the service economy enters its own era of technology-driven job disruption.

Size represents the category’s share of new occupations: red boxes are shrinking occupations; green boxes are growing occupations

At the same time, we have demonstrated how robot investment drives economic growth and creates new jobs across each economy. Policymakers must anticipate the mix of forces this level of automation will unleash— including job creation as well as job displacement across the economy.

To better understand the skills challenges many workers will need to overcome to adapt to an automated future, Oxford Economics developed a Skills Matching Model

in partnership with Cisco (see page 52). This model simulates the more complex dynamics at play when labour markets are disrupted, as organisations fill vacancies with compatible workers, who then leave a vacancy in their previous sector. The model captures the many small moves involved in the labour market’s evolution around technological disruption, highlights the new skills the economy now demands, and

illustrates how the skills deficit might be best addressed.

almost 30% of the redundant cohort lacked the “interactive skills” future jobs will require. In addition, more than a quarter lacked necessary “foundational skills” such

as continuous learning, reading, and writing. Some skills challenges, such as ICT and technical skills, require formal education and regular refresher courses. Others, such as the softer skills of negotiation, persuasion,

and customer service, typically require on-the-job experience, and might be supported by more flexible, virtual training options.

In this report, we have outlined both the potential value that robots bring and the real fear that the displacement they cause could exacerbate existing economic inequalities. Robotization poses a fundamental dilemma for policy-makers, who must balance the potential gains of long-term growth with the short-term pain of social dislocation.

Our analysis took a coordinated, evidence-based approach to understand how robots are changing the world. The issues raised will

affect many different sections of business and society. The repercussions of robotization are interconnected and complex, but the growth in robotics is inevitable—these challenges must be embraced and addressed. Policy-makers, business leaders, technology companies, educators, and workers all have a role to play. With this in mind, we offer a framework for action so that different stakeholders can seek to navigate the challenges and opportunities that robotization will bring.

Business Leaders

•Do not hesitate to seek technological solutions to your business challenges; the pace of innovation

is increasing, and global competition continues to intensify.

•Get buy-in from your workforce when making technology investments. Communicate clearly your intentions regarding robots directly with your workers. Invest in the required training and education programs in parallel to your robotics strategy, to allow workers to anticipate and adapt.

•Recognise that technological changes will disrupt the lives of many workers and take responsibility collectively to help them onto a path towards future opportunities.

“Get buy-in from your workforce when making technology investments.”

Educators

•Be cognizant of the shifting demands for skills across the economy. Technology skills are critical, but the economy is evolving in ways that

require many non-technical and “soft” skills from workers, too.

•Develop flexible approaches to delivering skills training and education. In addition to keeping formal training curriculums fresh and relevant, this means investing more in “lifelong learning” and “on-the-job training” programmes.

•Invest in closer, more cooperative relations with local industries, to anticipate emerging workforce needs in the

local economy based on assessments of sectors that are growing and shrinking as a result of new technology. Anticipate which new skills will be needed for new jobs.

•Share data with local authorities to tailor education and training to local strategic skills

priorities and employment programmes. Intensive “boot camp” training is one option to jumpstart vulnerable or displaced workers into new fields of work.

Technology Companies

•Put steps in place to mitigate the job displacement and societal disruptions robots can create, to

avoid undermining the consumer market on which businesses depend.

•Collaborate around industry-wide initiatives to invest in human capital— for example committing to share a portion of profits.

•Partner with governments and educational institutions to take responsibility for job retraining and coaching. Recognise the range

of skills required in the workforce to optimise the potential of technology and to grow the market.

•Look for technological solutions for the problems robots create.

“Collaborate around industrywide initiatives to invest in human capital.”

Workers

•Audit your own job to better understand the balance between

unique, human skills and automatable skills, to compete in the right areas and make your job more “robot proof.”

•Adopt a “lifetime learning” mindset. Unlike in previous generations, there are no jobs for life. Retraining and upskilling will become a normal part of the employment landscape.

•Support programs that develop job flexibility, even in unionised work settings, to help develop cooperative job sharing and flexitime.

“Adopt a ‘lifetime learning’ mindset. There are no jobs for life.”

Government Policy-

Makers

•Explore and analyse the implications of robotization for the economy, the workplace and the wider society, and adapt policy programmes to the evolving landscape.

•Develop collaborative environments such as science parks, living labs, and other accessible innovation ecosystems to foster skills development for small and mediumsized enterprises.

•Map out the existing skillsets at the local level with labour force surveys and analyse alongside regional business trends and growth strategies, as a basis for strategic planning. Incentivise companies and workers with fiscal benefits to engage in local programs to retrain workers with locally relevant skills.

•Identify the areas most vulnerable to dislocation from the rise of robots and develop aggressive, forward-thinking programs to counteract those effects. Explore all policy options, from infrastructure investments to training initiatives and innovative welfare programmes (such as universal basic income).

LITERATURE REVIEW

To explore the implications of robot densification on manufacturing jobs and economic performance we developed two econometric models, grounded in academic literature. Our methodology built most prominently on four previous studies:

1.Graetz and Michaels (2015)23 drew on the data from the International Federation of Robotics (IFR) in an industry-country panel specification of 17 countries over the period 1993-2007. The researchers found that an increase in the use of robots per hour worked to boost total factor productivity and average wage levels, and to have a negative impact on hours worked by low-skilled workers, relative to middleand high-skilled workers.

2.Acemoglu and Restrepo (2017)24 used the same dataset but restricted their analysis to the period 1990-2007 (partly due to data limitations and partly to avoid the post-2007 crisis period) and explored localised effects of robot densification in the US economy, by exploiting heterogeneity in both local labour distributions across industries and national change in the use of robotics. They found that one additional robot per thousand workers reduces the US employment-to-population ratio by 0.37%, on average.

3.Dauth et al (2017)25 adopted a similar localised model in the German context with a timelier dataset, from 1994-2017. They found that while industrial robots had a negative impact on employment in the German manufacturing sector, there was a positive and significant spillover effect on jobs in non-manufacturing sectors to offset it.

4.Chiacchio et al (2018)26 also built on the Acemoglu approach in a study of six EU countries between 1995-2007. They found that one additional robot per 1,000 workers reduced the employment rate by 0.16-0.20 percentage points. They do not find robust and significant results on the impact on wage growth.

DATA

We used data on robot investment (per unit) from the IFR in Japan, the European Union, the United States, South Korea, and Australia. The IFR reports robot stock and investments for 50 countries over the period from 1994 to 2014. It is based on yearly surveys of robot suppliers and captures around 90% of the world market. The information is broken down at the industry level, but data availability differs across countries.

We augmented the IFR dataset with Oxford Economics’ data on GDP, GVA by sector, employment by sector, population by 5-year age band, and wages and compensation in total and by sector (where available) from our datasets. We also used data on trade from the COMTRADE database.

We used a sub-national unit of analysis for the modelling exercise, corresponding to European NUTS 2, US Metro areas, Japanese prefectures, Australian states, and South Korean districts.

We built a panel dataset for 29 countries (all of EU, plus US, Japan, Australia, Korea) over 11 years within the period (2004-2016), disaggregated by region and sector.

We used a dynamic panel approach (using the Generalised Method of Moments estimator,

or GMM) to account for secular trends. The GMM approach allowed us to use internally generated instrumental variables (i.e., using past values as instruments) that helped us establish a causal link between the growth in robot density and employment or productivity.

Model specification

The following section describes the two models used in our analysis.

Model 1: Manufacturing employment

The model specification, as shown below, isolates the impact of robot density (robots per 1,000 workers) on the manufacturing employment-to-population ratio in each region (r) and year (t), having controlled for secular trends (lagged employment to

population ratio), economic performance (GDP per capita), globalisation (trade with China, trade with the rest of the world), wage levels (compensation per capita), region-specific factors (using panel data), and other trends/ year-specific events (year dummies).

Model 2: Local labour productivity

The model specification, as shown below, isolates the impact of the log of robot density on the log of GVA per worker in each region and year, having controlled for the pre-existing trend (lagged productivity), globalisation (trade with China), wage levels (compensation per capita), region-specific factors (using panel data), and other trends/year-specific events (year dummies).

Our approach embodies one key difference with the existing literature: we use a dynamic panel method exploring year-on-year variations, while previous studies focused on the cumulative change in employment between two distinct years (e.g., Acemoglu between 1990 and 2007). Acemoglu and Restrepo justified this approach as a way of avoiding the potentially confounding effects of the recession post-2007. While this is a valid concern, we mitigated this risk by using

more recent years of data (up to and including 2014) and dynamic panel methods. In addition, the cumulative change method would have been inadequate for the period 2004-2014 due to the changes in employment over that period (which declined in the recession years and increased in the subsequent years) and the change in robot stock (which increased consistently over time). By using appropriate control variables, including time dummies, in

a dynamic panel setting, our model accounts for the effects of the recession and isolates the impact due to robots.

To test the relative impact of robot densification in higherand lower-income regions, we categorised regions by average compensation per worker. Those that were higher than the national average over the sample period were labelled higher-income regions; those lower than average were labelled lower-income regions. We used interaction terms (also known as partitioned variables) in the same model specifications as above to identify the relative differences. As a sensitivity check, we tested alternative control variables to establish that the coefficients on the control variables were significant and of the right sign.

ANALYSING SHORT-TERM AND LONG-TERM IMPACTS

To control for the fact that the impact of robotization may take time to impart its full effect on employment and productivity, we used lag variables. This involved investigating the relationship between robot installations in one period, and employment and productivity in subsequent periods, with various degrees of time lag.

ESTIMATOR SELECTION

In order to produce robust estimates, we used various tests to select the most appropriate estimator. We ran the Wooldridge test for autocorrelation28 to determine the need for a dynamic panel approach. The data was found to be autocorrelated, meaning employment and productivity in current periods might be affected by past trends in the same variables.

We used the Arellano Bond/Blundell Bond estimator (“System GMM”) to account for the presence of such “dynamic effects” in the data. This method augments the Arellano–Bond estimation (“difference GMM”) by allowing

the introduction of more instruments and can dramatically improve the efficiency of the model.29 Using a dynamic panel model also enables us to identify overall coefficients for explanatory variables, corresponding to the long-term effects as well as contemporaneous ones corresponding to the short-term.

Even where a dynamic model is the preferred specification, it still comes with potential risks. The most prominent risk is omitted variable bias. For example, the growth in ICT services corresponds broadly with the rise in industrial robots but because our unit of analysis is at the regional level, there is insufficient data to control for ICT spend at local levels across the 29-country sample. We attempted to mitigate this risk by using a proxy for local productivity growth (wage growth) and past values of the dependent variable as instruments, which reduces the need to identify all alternative instrumental variables. Instruments in dynamic models also account for serial correlation between past and current values of employment and productivity.

ECONOMETRIC RESULTS

Model 1: Manufacturing employment

The regression output for our three preferred model specifications are presented in Table A1.

1.1Our preferred model for the average employment impact across all regions.

1.2Our preferred model to estimate the variable impacts in higherand lowerincome regions.

1.3Our preferred model to estimate the productivity impact across all regions.

In all three models, the lag dependent variable is low but significant at the 5% level. The coefficients on the dependent variables are significant and of the expected sign in all three models. All three models include year dummies. We have also tried estimating other specifications across all three models but were unable to obtain a satisfactory model.

Table A1: Change in the number of jobs due to one additional robot

The employment model implies that a 1 unit increase in robots per 1,000 workers reduces the average employment ratio by 1.5 in the short run and 2.0 in the long run. The impact is lower for high skilled regions, at 1.2 and higher for low skilled regions, at 2.0.