The Future of Employment Revisited: How Model Selection Determines Automation Forecasts

04/28/2021 ∙ by Fabian Stephany, et al. ∙ University of Oxford 0

The uniqueness of human labour is at question in times of smart technologies. The 250 years-old discussion on technological unemployment reawakens. Prominently, Frey and Osborne (2017) estimated that half of US employment will be automated by algorithms within the next 20 years. Other follow-up studies conclude that only a small fraction of workers will be replaced by digital technologies. The main contribution of our work is to show that the diversity of previous findings regarding the degree of job automation is, to a large extent, driven by model selection and not by controlling for personal characteristics or tasks. For our case study, we consult experts in machine learning and industry professionals on the susceptibility to digital technologies in the Austrian labour market. Our results indicate that, while clerical computer-based routine jobs are likely to change in the next decade, professional activities, such as the processing of complex information, are less prone to digital change.



There are no comments yet.


page 18

page 21

page 24

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

The motivation behind our work is the discussion about technological unemployment, which has accompanied technological processes throughout the last 250 years. The debate about the susceptibility of human labour to digital technologies accelerated since a prominent study by frey2017future concluded that half of US employment will be automated within the next 20 years, which would pose a sizeable thread to societal stability 222In addition and interaction with other global dynamics, such as rising income inequalities (stephany2017your; stephany2019deepens) or climate change and mass migration (hoffmann2019quantifying).. Their estimations are the basis for several follow-up studies, which infer that the share of jobs at risk is much smaller. Our work examines the reason for the stark diversity in previous findings about job automation. We propose that differences in the degree of susceptibility emerge mainly from model selection. In order to test this assumption, we conduct a case study similar to Frey and Osborne with a survey among Austrian research and industry experts. Our model testing confirms that differences in previous findings on the automation of jobs are mainly driven by the design of the model, rather than heterogeneity among tasks within occupations. Our results indicate that, while clerical computer-based routine jobs are likely to change in the next decade, professional activities, such as the processing of complex information, are less prone to digital change.

Machines have both complemented and competed with human labour in the past. Inventive ideas and creative destruction, as schumpeter1942creative puts it, have competed with powerful social and economic interest over the technological status quo. Various movements, such as the Luddites, who destroyed new machinery in the 18 century textile industry, have tried to deter progress in times of rising unemployment. However, the Luddite fallacy has found its way into the literature, as employment has not been eradicated alongside fast technological development, but instead continued to expanded. Rather than eliminating human labour as such, technological advancements have changed a number of work profiles and led to the creation of new professions.

Whenever modern society experiences technological advancement, concerns about technologically induced unemployment arise. In recent history, technological progress has often been linked to a displacement in specific professions (bresnahan1999computerisation) or even entire industries (charles2013manufacturing; jaimovich2012trend). However, to date, technological progress has not caused mass unemployment. We have seen a shift in labour from the agricultural sector to manufacturing branches, and further into the service sector (david2015there). Overall employment has been steadily increasing worldwide, despite (or perhaps because of) technological progress. Hence, new technologies display two opposite effects on employment (aghion1994growth). On the one hand, technologies substitute human labour in order to decrease production costs and increase productivity. This displacement effect lowers employment. On the other hand, reduced production costs increase real income and hence demand. The latter effect fosters production and demand for labour.

According to goldin1998origins, technological progress led to the simplification of work processes in the 19th century. A combination of machines and unskilled labour substituted skilled labour and decreased demand in terms of skills. However, as technologies improved, technological job displacement shifted away from skilled to unskilled labour. acemoglu2017robots calculate that an increased use of robots in the US economy between 1990 and 2007 had a negative effect on the labour market. According to their calculations, an increase in the number of industrial robots by one, per 1,000 people employed, reduces the employment-to-population ratio by 0.18 to 0.34 percentage points.

Similar to signs of competition with rooters for physical work, mcafee2014second emphasize that computerization has now started challenging human performance in cognitive tasks. beaudry2016great, in an empirical analysis, find evidence that the demand for skilled labour has been declining in recent years. This is an indication that skills under pressure of substitution are altering as technological progress persists. david2013growth show that the implementation of computer-based technologies has put pressure on wages. As routine tasks are increasingly automated, displaced workers reallocate to the lower skilled service sector with deteriorating wages. According to goos2009job, this has resulted in the increased polarization of the labour market in a number of developed economies (see also dustmann2009revisiting). Increasing demand for well-paid jobs in which non-routine cognitive tasks are performed, as well as non-routine manual work at the lower end of the income distribution, in combination with the automation of repetitive cognitive skills, is forcing employment away from the middle of the income distribution (see also autor2003skill; david2013task; michaels2014has).

Recent publications, such as ford2015rise, raise concerns that "this time it could be different" and there will be no room for creating new jobs. frey2017future set the starting point for a series of papers that attempts to calculate the impact of digital technologies on the demand for human labour. Based on their original data, collected during a workshop involving machine learning experts, several papers about the susceptibility of jobs have been published.

Yet, transferring the data on susceptibility from the US labour market to European economies is challenging in many respects. Until now, there has been no piece of research that has analysed the impact of computer automation on the labour market by using newly collected data from European countries. This approach allows us to correct the shortcomings in transferring the original US data (O*NET) of frey2017future to the International Standard Classification of Occupations (ISCO). It also adjusts for regional particularities in labour markets, for example, differences in regulation or cultural particularities. Even though technological innovations have become market-ready, customers may hesitate to substitute them for human interaction. In addition, we analyse the possibility of a non-linear relationship between education and future digitalization, since both low- and high-skilled jobs are assumed to be less affected by digital technologies than medium-skilled professions dustmann2009revisiting.

Addressing previous limitations, we assume that the strong differences in the degree of susceptibility between frey2017future and follow-up studies is due to model selection. As case study, our investigation examines the degree of future digitalization of job profiles in Austria. We link expert opinions with individual data from the OECD’s PIAAC data, which in turn allow for heterogeneity among workers within the same occupation. Our results indicate that, models with a binary outcome, as applied by frey2017future result in a much higher share of jobs at risk than models with a fractional dependent variable, as used by the OECD. In both settings, clerical computer-based routine jobs are likely to change in the next decade, professional activities with the processing of complex information are less prone to digital change. The following section 2 describes the methodology and data, followed by the 3 section, which summarizes the results, while the last section concludes the paper.

2 Data and Methods


were the first to attempt to quantify the potential of computer-based job displacement in the near future. Based on the estimates of robotic experts, the authors calculated the susceptibility to computerization of different jobs, according to the O*NET database in the US. They conclude that 47% of the jobs in the US are at a high risk (>70 % probability) of being replaced due to computerization.

bowles2014computerisation applies the same method and transfers the results to European economies using the differences in the sectoral structure of each country. He concludes that 54% of jobs in Austria have a high risk of being displaced by computers.

arntz2016risk emphasize that the method used by frey2017future overstates the share of jobs susceptible to computerization. As frey2017future do allow for heterogeneity in tasks between different jobs, they do not allow for alterations in the tasks within one occupation. According to arntz2016risk, one profession may contain different sets of tasks, and thus the risk of computerization could vary within this profession. Using PIAAC survey data, they combine information about the composition of tasks within each job profile with information from robotic experts on the susceptibility of jobs for the US labour market. They further transfer the results to other OECD member countries, indicating that 9% of US workers and 12% of Austrian workers are at high risk due to computerization.333bonin2015ubertragung use a similar approach for Germany, pajarinen2014computerization for Finland, and nagl2017digitalisierung for the Austrian economy. According to nagl2017digitalisierung, 9% of Austrian workers have a high risk of being automated. Among OECD countries, Austria, as well as Germany, displays the highest share of the workforce at a high risk of computerization.

For the German labour market, dengler2015folgen

relate the risk of job automation to the tasks that are characteristic of each profession. They compute the share of tasks that can be classified as routine based, according to the classification by

spitz2006technical. According to their findings, 15% of German workers are employed in jobs with a high risk of automation. Likewise, for Austria, peneder2016osterreich find that 12% of Austrian workers primarily perform routine-based tasks.

Similar to the approach by frey2017future, we begin our analysis with expert opinions. Between December 2017 and January 2018, we consulted Austrian industry experts and machine learning professionals. The final data set contained 35 individual experts’ opinions, with 14 individuals representatives of Austrian companies in the fields of construction, consulting, insurance, investment, media, real estate and retail, and 21 responses were from industry and academic experts in machine learning and AI. Experts from both groups were individually requested to participate in an online survey. In comparison, the expert workshop by frey2017future, which was held in 2013 at Oxford University’s Engineering Sciences Department, included 70 machine learning experts (brandes2016opening). Together with their team of experts, frey2017future initially labelled 70 out of 703 US jobs. These binary labels were then used to predict risks of automation for all US professions. The resulting estimations formed the basis of the aforementioned studies in a European context. However, for the estimation of impacts of digital technologies on the Austria labour market, our expert opinions are better suited than the opinions stemming from the Oxford seminar. Machine learning experts are familiar with the scientific principles of the technologies disrupting the labour market, but they may not be fully aware of the social environments in which smart technologies could be implemented. For example, even when chatbots in the financial service sector become market-ready, from a technological point of view, some customers will still prefer interaction with a human. In addition, the gap between technological readiness and implementation varies to a sizeable extent between countries and cultural backgrounds. In order to address this aspect of the application of new technologies, we consulted Austrian experts from the field of machine learning/AI and professionals from various industry domains.

The participants in our survey were asked about their opinion on the 100 most common professions in Austria, as listed in Table 1. In contrast to the focus on the susceptibility to computerization (frey2017future), we asked our experts:

"Do you think that the tasks, which are characteristic of this profession today, will be substituted, to a significant degree within the next 10 years, by algorithmic technologies (such as machine learning, computer vision and natural language processing) or mobile robotics?" (Yes=1/No=0)

. This question analyses the degree to which the nature of certain professions is going to change due to technological advancement. Answers to this question do not necessarily reflect the risk of occupations being fully substituted by technologies.

Experts were allowed to avoid answering the question in relation to as many jobs as they wished. However, in the end, only a small minority of jobs remained unlabelled. In order to extract an indicator of future digitalization that is unique to each profession, we calculated three measures: the mean and mode of all expert opinions, as well as an indicator of the experts’ consensus on each profession. The consensus is equivalent to the mode, but only for those professions to which at least 75% of all experts attributed the same label. With this definition of consensus, 45 professions remained and received a binary label, as shown in Table 1.


In the second step, the profession labels were matched with profession groups from the Austrian and German samples of the 2015 OECD survey of the PIAAC. The PIAAC survey supplied our analysis with individual characteristics, as well as job- and firm-level indicators. In addition, the survey contains information about the frequency of specific tasks performed by interviewed individuals during their average working routine. These tasks, as listed in Table 2, include human interaction, IT usage, physical work, problem-solving, reading or understanding, and writing or calculating. As the individuals provided answers about the frequency by which they undertake a given task, we normalized the answers according to the value of the working hours as follows: ’on a daily basis’ (value=1), ’less than daily, but more than once a week’ (value=1/2), ’less than once a week, but more than once a month’ (value=1/7), ’less than once a month’ (value=1/30), or ’never’ (value=0). This labelling is likewise applied by arntz2016risk, since it reflects the differences in scale between days, weeks and months.


Thirdly, the expert opinions about the future change of professions are related with the PIAAC data. These opinions about professions are matched via the ISCO-08 classification for each individual’s job444Only the German PIAAC sample contains the respective ISCO-08 Level 4 job classifications. Hence, the fitting of the inferential models is performed only with the labelled subset of the German employees.

. The PIAAC survey is conducted in a way that it contains a representative sample of the population. However, not all observations within the survey contain answers to all questions. Thus, the specification of the model leads to a loss in observations due to non-responses. There is no reason to assume that the loss in observation systematically changes the sample. We perform a mean imputation for the non-response values, which increases the model’s sample size by 55%, but does not lead to a significant difference in results. Compared to the 2012 labour force survey, our sample displays a slight shift towards younger age groups. Furthermore, the sample shows a higher share of female employees (for details, see Table

3). Nevertheless, the impact of technological change on job profiles stays unchanged.

In order to relate the above-mentioned characteristics to the given expert opinions about the individual’s job, we test three inferential models. The consensus indicator serves as the dependent variable, while various combinations of personal-, job- and firm-level controls, as well as task frequencies, are included in the model (Table 3). The correlation analysis in Table 6

across all characteristics only indicates a sizeable association between the three test score variables. All measures are considered at the individual level with a sample of 507. The extrapolated sample contains 4,438 individuals: 2,051 from Austria and 2,387 from Germany. In a first round, we apply a logit model. This stepwise procedure is illustrated in Columns (1)-(6) in Table

4. The Akaike information criterion indicates that Model (6), with all controls, yields the best model fit. In the second round of the model selection, we test a linear discriminant analysis (LDA) with a Bayesian estimation of the dependent variable (james2013introduction, Chapter 4)555The probability of belonging to class k, given characteristics X, is described by ,  , while describes the probability of , given that , while

is the prior probability of observing

, which is similar to the approach chosen by frey2017future. In order to compare the logit and LDA models, we apply a cross-validation method (40% training sample). The comparison of the in-sample predictions shows that the logit model (area under the curve (AUC)666The AUC measures the area under the receiver operating characteristics (ROC) curve. The AUC is a measure of prediction accuracy, since the ROC curve plots the true positive rate against the false positive rate of a prediction model.: 0.94) slightly outperforms the LDA model (AUC: 0.92). The estimations of the LDA model are very similar to the results of the logit model, as summarized in Table 5. Lastly, we compare the results of the logit model with a fractional response model (papke1993econometric)777,  , while . In this model, the mean of the experts’ opinions is considered as the dependent variable. Accordingly, the fractional model refers to a larger sample size. However, the results in Table 4, Columns (6) and (7), show that the logit model still yields a significantly better model fit.


After identifying the appropriate model environment, the logit model (1) is used to predict the digitalization probabilities, P(y=1|X), for all individuals in the sample, based on their set of characteristics (). Here, individuals with professions, which have not been judged by our experts, also obtain a probability. The average estimated probabilities of future digitalization are shown in Figure 2, and are aggregated for ISCO-08-Level 1 (Figure 3) and ISCO-08 Level 2 (Figure 4) professions in Austria.

,     (1)

Based on the consensus of our experts, we are able to specify a degree of future digitalization for 47 occupations. More than 75% of our experts agreed that the characteristic tasks of these professions will change to a significant degree with the development of digital technologies and mobile robotics. With the use of the PIAAC data set, we are able to relate the degree of digitalization to personal characteristics and occupation-specific tasks. Based on this relationship, we estimate the degree of digitalization for all professions in the data set. In contrast to the work by frey2017future, we apply local experts’ opinions and perform our estimations on the basis of individual characteristics.

3 Results

For some tasks we see a clear relationship with the consensus of our experts. In Figure 1, the frequencies of the 39 tasks are compared to the consensus of our experts. On average, some tasks, such as coding (itusage_code), are, on average, performed less than once a month, while others, such as sharing information with others (human_share), are carried out on an almost daily basis. For some activities, prevalence does not differ significantly between the two consensus job groups, for example, itusage_code or human_share. However, for most of the activities, a clear separation between the consensus groups is visible. Activity involving long physical work (physical_long) is less commonly performed in professions that are expected to change during digitalization, according to our experts. Other activities show the exact opposite pattern. Calculating (wricalc_calculator) or the use of computer software Excel (itusage_excel), for example, is much more prevalent in professions that are expected to change. This observation, confirmed by the findings of the inferential model, is a first indication that professions with a high degree of computer-based office routines are more likely to change in light of digital technologies.


In addition to the 39 tasks, individual-, job- and firm-specific characteristics can help explain the consensus opinions of our experts, as shown in Table 4. The final and full model (6) indicates that, apart from work activities, education, firm sector, job responsibility and training are related to the degree of future digitalization. Individuals with a high level of education, who work in a job that requires training or responsibility, are typically less likely to be employed in an occupation that is going to change significantly. Interestingly, our results indicate a non-linear relationship with education. Individuals with a medium level of schooling are employed in jobs with a higher level of future digitalization than workers with high or low levels of education.

Our model indicates that certain work activities are strongly related to the degree of digital change in the workplace. Tasks such as extracting complex information by reading books (reading_book) or writing non-routine content (wricalc_report) are related to professions with a low degree of technological change. On the other hand, activities such as calculations (wricalc_calculator) or extracting simple information (wricalc_news) are associated with a stronger change in the job profile in the near future. mcafee2014second, for example, show that news stations have begun implementing algorithms that are able to write simple pieces in the context of sports or weather forecasts. Moreover, for professions that predominantly rely on physical labour, impacts of technological change are also low.


Among occupations, there is a clear trend (Figures 3 and 4): clerical support workers, who perform simple computer-based office routines, are highly susceptible to technological changes. This is in line with previous findings. On the other hand, professionals, who work with complex and unstructured information, and skilled workers in agricultural fields, who perform physical work, are less likely to experience major changes in their job profile. Professional occupations involving teaching and healthcare within legal, social or cultural environments (Figure 4) exhibit particularly low probabilities of digital transformation. This finding is consistent for individuals working in a job that requires an academic degree, as well as for those without such a qualification. On average, most occupations show a probability of change between 40% and 60%.


When comparing our model findings, clear differences emerge with regard to the degree of susceptibility in employment to digital technologies. However, our model testing suggests that these difference are mainly driven by model selection, rather than heterogeneity among tasks within occupations. Table 5 compares the set-ups of our research and previous studies.


Two types of model settings are prevalent. frey2017future start with binary opinions of experts and extrapolate them via a classification model for all occupations. bowles2014computerisation directly transfers these estimations to European labour markets. Both studies conclude that a high share of workers (47% in the US and 54% in Austria) share a high risk of computerization. arntz2016risk and nagl2017digitalisierung, on the other hand, begin with discrete probabilities and apply a fractional model in order to extrapolate. In comparison, they show that only about 12% and 9%, respectively, have an automation risk of more than 70%. In light of these contradictory findings, our model testing suggests that the different estimations are mainly due to the choice of model. Binary models yield a bimodal distribution of predicted probabilities with large high-risk groups. Fractional models lead to a bell-shaped distribution of probabilities with relatively low levels of high-risk individuals. Our own estimations for a fractional model (Figure 5) confirm this assumption. The ranking of occupational classes does not change significantly after the fractional model (Figure 6) has been used. However, predicted probabilities converge towards the mean.


When comparing the outcome of the binary and fractional model, the results of the latter contain a lower number of covariates, which are statistically relevant to the degree of digitalization. The fractional model, however, does not show any statistical significance concerning the covariates that have not been relevant in the binary model. In the fractional model, education and job responsibility show no statistical significance. Likewise, the tasks of speaking in front of humans, reading books, using words, and coding are not significant in the case of the fractional model environment. This general observation is not surprising from a statistical point of view, since the formally strict binary outcome in a small sample has now been changed to a smooth continuous scale in a sample twice the original size. However, it becomes clear that some covariates, such as physical work, writing reports, performing calculations or firm characteristics, are still aligned with the distribution of the fractional model. The distribution of other covariates has been polarized by the truncation of the binary model. The unconditional distributions of the binary and fractional models are shown in Figures 7 and 8.


Similarly, when moving the threshold of consensus from our final value towards 0.5, the outcome of the binary model starts to slightly approach the results of the fractional model. However, no significant changes appear, except for a deterioration in statistical significance.

4 Conclusion

Our model explicitly diverges from the approach taken in previous contributions to this field. We assume that the diversity of previous estimations of job susceptibility stems from model specification. In order to test this assumption we conducted a case study with local expert opinions about near-term changes in occupations in Austria. This is a significant conceptual improvement in contrast to prior investigations (arntz2016risk; bowles2014computerisation), which studies rely on the judgement of machine learning experts concerning the US labour market, stemming from the workshop organized by frey2017future. However, the authors do not allow for heterogeneity within the same profession. This limitation is ruled out by our model approach. Past findings are, in part, contradictory. 47% of jobs in the US (54% in Austria) share a high risk of automation, according to frey2017future and bowles2014computerisation, while arntz2016risk and nagl2017digitalisierung estimate this share to be 12% and 9%, respectively, for Austria. Our findings show that these differences are mainly driven by the selection of the model, and not so much by controlling for personal characteristics or tasks.

Our findings show that the tasks that humans perform during their typical working day are of significant importance when determining the impact of digital technologies on the future workspace. Activities such as extracting complex information by reading books or writing non-routine content reduce the impact of technologies. On the other hand, tasks such as calculations or extracting simple information will lead to a stronger change in job profiles in the next decade. Furthermore, as the current generation of technological progress has a stronger impact on cognitive and routine tasks than on physical labour, the extent of physical work within a job profile reduces the effect of digital change. Although the future of work will most likely be a complementary partnership between humans and computers, workers performing computer-related routine activities, such as spreadsheet calculations or Internet usage (stephany2021does; stephany2021one; stephany2020coding), are under stronger pressure to adapt. Our findings about the "inverse U-shaped" relationship between education and digitalization support previous hypotheses about the skill-based polarization of the labour market (goos2009job). This suggests further polarization in the near future.

Our results indicate that some jobs can expect to change more than others during the current phase of digital progress. This is surely not the first time in history that this has happened. During the Industrial Revolution, technological advancements made manufacturing jobs less intensive in terms of monotonous physical labour. In contrast to the age of the steam engine, today’s technologies, such as algorithms, unfold their potential in disciplines that require routine cognitive effort. Typical computer-backed office tasks, such as in the clerical professions, are more exposed to digital transformation than occupations marked by physical labour. Likewise, jobs in which complex information is processed and that require a high level of education and training are less prone to digital change in the near future. Teaching and health-care professionals working within in legal, social or cultural environments belong to occupations with the lowest level of technological pressure. In the near future, these disciplines can be regarded as a sustainable choice for future generations seeking job security in unsteady times.

In addition, while most research focuses on human labour that can be replaced by technology, little attention has been given to the effect that digital technologies have on job creation. As our findings improve the understanding of the displacement effect of technologies, more research should be conducted in order to incorporate the effect of job creation, and in turn appreciate the full impact of the technological change on the labour market.