Robot Application and Labor Force Employment: Substitution or Complementation? —An Empirical Analysis Based on the Data from 22 Economies

2.1 Impact of Robot Application on Total Employment

Previous researches are not consistent in their conclusions regarding to the impact of robot application on total employment. Some scholars believed that the efficiency advantage of robots will benefit non-automated jobs, thereby creating greater demand for labor (Autor and Salomons, 2018). This viewpoint has been supported by empirical evidence. For example, Gregory et al. (2018) studied the EU data, finding that the application of robots caused 9.6 million old jobs to be replaced from 1999 to 2010, but 21 million new jobs emerged. Based on transnational data research, Autor and Salomons (2018) found that robot application will reduce the proportion of workers’ labor value added in all industries, but in terms of employment alone, it can promote overall job growth through countervailing effects. Wei et al. (2020) reached similar conclusions based on data from China respectively.

Some scholars, however, held that despite the countervailing effects, substitution effects are dominant in most cases (Acemoglu and Restrepo, 2018). They further proposed that new technologies not only affect workers’ production efficiency in current tasks, but produce a “displacement effect” over the allocation of new tasks, resulting in a greater favor of capital in the allocation of task content and deepening the adverse influence over labor force (Acemoglu and Restrepo, 2019). Empirically, Acemoglu and Restrepo (2020) analyzed the data of the US, concluding that robotics has negative impact on employment, more so in states with a higher degree of robotics use. In addition, some scholars considered that, under the offset effect of substitution effect and complementary effect, the impact of robots on employment is not obvious. For instance, Arntz et al. (2016) researched transnational data and found that though robotics application reduces jobs for low-skilled labor force, its impact on overall employment is not significant.

In summary, robot application has a positive complementary effect and a negative substitution effect on employment. The working mechanism of complementary effect can be summarized into two aspects. First, the application of robots is conducive to improving production efficiency, stimulating the expansion of corporate scale, and bringing about greater labor demand (Autor and Dorn, 2013; Autor and Salomons, 2018). At the same time, the mechanism may produce a ripple effect in the entire supply chain and work in a wider scale (Wang and Dong, 2020). Second, the application of robots helps create new business formats, new models and new job demands (Wang et al., 2020), and the newly generated demand often outnumbers the jobs replaced (Gregory et al., 2018). The working mechanism of substitution effect can also be summarized into two aspects. First, robots enjoy comparative advantages in some jobs compared with labor force and therefore will replace this portion of labor force, leading to “technological unemployment” (Acemoglu and Restrepo, 2020). Second, robot application will intensify the Matthew Effect, in which case large enterprises, with scale and technical strength, compress the living space of small and medium-sized ones (SMEs), causing the labor force originally employed in SMEs to be “indirectly replaced.” To sum up, it is difficult to accurately predict the employment effects of robotics with sheer theoretical analysis. Accordingly, this paper proposes the following competing hypotheses to be verified:

Hypothesis H1a: When complementary effects are dominant, robot application brings up the labor demand.

Hypothesis H1b: When substitutive effects are dominant, robot application brings down the labor demand.

It is worth noting that, the existing literatures based on transnational samples and developing country samples tend to conclude that complementary effects are dominant or the impact is not significant (Arntz et al., 2016; Gregory et al., 2018), while those based on samples from countries such as the United States where robots were introduced early and used extensively are more likely to conclude that substitution effects are dominant (Acemoglu and Restrepo, 2018, 2019, 2020). This phenomenon is similar to the inverted U-shaped curve hypothesis about economic development and income gap in the course of industrialization proposed by Kuznets. If this hypothesis is introduced into the labor market, will the impact of robots on employment resemble the Kuznets curve for its connection with the development stage and degree? Given so, this paper puts forward the second research hypothesis:

Hypothesis H2: The impact of robotics on employment is not simply linear, but in an inverted U-shaped structure featuring the transition from complementation to substitution as robots are used to a higher degree.

2.2 Impact of Robot Application on Employment Spillover

Previous literatures examine the robot-induced employment spillover mainly in two dimensions. The first is the inter-industry spillover of labor. For instance, Autor and Dorn (2013) believed that automation technology will have a substitution effect on labor force in routine tasks, but the labor force is not squeezed out of the labor market because of the substitution, instead, they are pushed into services and other industries. Dauth et al. (2018) studied data in Germany and found that the substitution effect of robot application on manufacturing will countervail in the service industry, canceling out the substitution and complementary effects in general. Zhao et al. (2020) found based on China Labor-force Dynamics Survey data that robot application has not reduced the overall regional employment level, but accelerated the reallocation of labor between the secondary and tertiary industries. The second is the spillover of labor in the industrial chain. For example, Kong et al. (2020) studied the industry-specifi c data in China, finding that robot application is conducive to job growth in local downstream industries, but the impact on employment in upstream industries is not significant. Wang and Dong (2020) concluded based on data on Chinese listed companies that robot application has a negative impact on labor force employment in both upstream and downstream industries.

Essentially, the biased influence of robot application on work force engaged in routine tasks will twist the structure of the employment market and push labor force to transfer from production sectors to services (Autor and Dorn, 2013). However, according to the “push-pull” theory, population migration is susceptible to both pushing force from the origin and pulling force from the destination. When this theory is introduced into study on employment spillover, the impact of robot application on employment may include both inter-industry forward crowding-out effects and reverse siphon effects. As to the spillover effect of labor in the industrial chain, previous researches mostly focus on the conduction effect of the impact in the domestic industrial chain, without paying close attention to whether the rise of robotics is coupled with the anti-globalization wave in recent years and the return of industries worldwide. In theory, the advantage of robotics in efficiency helps bring down the labor cost, weaken the motivation of upstream economies in the industrial chain for offshore outsourcing of labor-intensive industries, and hence pose adverse influence on employment in downstream economies. Based on such analysis, this paper proposes the following hypotheses:

Hypothesis H3a: The impact of robot application on employment includes both inter-industry forward crowding-out effects and reverse siphon effects.

Hypothesis H3b: The impact of robot application on employment is transmitted not only in the domestic industrial chain, but also in the international industrial chain.

3.1 Data Sources

The empirical data in this paper mainly comes from the International Federation of Robotics (IFR), ^[1] the International Labor Organization (ILO), ^[2] the World Bank, ^[3] and the Research Institute for Global Value Chains of the University of International Business and Economics (UIBE GVC). ^[4] Specifically, the data on robot stock by industry comes from IFR, and this database counts the use of robots in about 75 economies. The industry-specific data on the number of employees, the average weekly working hours of the labor force, the gender ratio of industry practitioners and the proportion of highly skilled labor comes from the ILO. Industrial value added and competitive advantage data are sourced from the UIBE GVC. The dependency ratio data derives from the World Bank. In addition, because of the lack of some industry-specific data of China in the ILO and World Bank databases, it is sought from the China Statistical Yearbooks and the China Labor Statistical Yearbooks over the years.

Given the statistical starting years and missing data in the main databases, this paper selects 2008–2019 as the research sample. To ensure the completeness of data for each industry during the sample period and avoid regression errors caused by missing industries or missing statistical data, 22 sample economies with complete statistical data available for the same industry categories were selected. It should be noted that the IFR and ILO industry classification standards are not fully consistent. Based on the IFR industry classification, this paper aggregates some industries in the ILO and arrives at seven industry categories, namely agriculture, forestry, animal husbandry and fishery, manufacturing, production and supply of electricity, gas and water, quarrying and mining, education and research & development, construction, and other services. The analysis sample in the economy-industry-year three dimensions is eventually made available with 1,848 valid observed value in total.

3.2 Model

This paper uses data in the three dimensions of economy, industry and year to explore the impact of robot application on labor force employment, and constructs the following model:

Wherein, i means economy, j industry, and t year. Explained variable lnlab_ijt refers to logarithm of the number of employees, core explanatory variable lnrob_ijt is logarithm of robot stock, and control_ijt refer to controlled variables, including characteristic variables of industry and of economy. u_i , v_j and r_t respectively represent economy, industry and year fixed-eff ect, and ε _ijt is random error term.

3.3 Setting of Variables

4.1 Population Regression and Analysis

Table 2 reports the empirical results of how robot application affects total employment. Except for column (1), robot application has positive impact on labor force employment, which passes the 5% significance test. The comparison among columns (1)–(3) indicates that the estimated coefficient of including the core explanatory variable alone into the model is positive but not significant, while as industry characteristic variables and economy characteristic variables are successively included into the model, the regression coefficient value and significance of robot application are gradually increased. Given the possible endogeneity with the core explanatory variable, column (4) adopts 2SLS for estimation. The result shows that the impact of robot application on employment remains significantly positive and the estimated coefficient is substantially higher than that in column (3), indicating that the instrumental variable method can effectively address the estimation errors caused by missing variables and reciprocal causation. The comprehensive analysis of columns (3)– (4) suggests that in general, the impact of robot application on labor force employment is dominated by complementation.

The benchmark regression of the controlled variables and the 2SLS regression yield results are just as expected. Take column (4) for example. The estimated coefficient of the average weekly working hours ( ln weekwork ) of the labor force is significantly negative, indicating that under constant total workload, the longer average weekly working hours, the weaker demand for labor force. The estimated coefficient of gender ratio ( ln gender ) is significantly negative, indicating that an excessively high proportion of male employees is unfavorable to overall job growth, which coincides with the reality that increasingly more women are flooding into the labor market. The estimated coefficient of the proportion of highly skilled talents ( hskills ) is negative and passes the 5% significance test for two possible reasons. First, highly skilled talents yield higher productivity, while higher productivity of individual laborers drives down the corporate demand for labor force. Second, economies with a higher proportion of highly skilled talents are more likely to offshore outsourcing of non-core industries, mostly labor-intensive. Therefore, a higher proportion of highly skilled talents is possibly not helpful for overall job growth.

4.2 Heterogeneity Regression and Analysis for Different Industries

To explore industry heterogeneity, Table 3 reports the regression results by industry. ^[1] Specifically, Panel A and Panel B respectively introduce the benchmark regression and 2SLS regression. Given the mild discrepancies between their estimation results, to avoid possible endogeneity problems, Panel B is analyzed below as an example. ^[2] As shown in the table, the introduction of robotics into agriculture, forestry, animal husbandry and fishery, manufacturing, construction, and other services can all significantly boost employment, while the introduction into production and supply of electricity, gas and water and quarrying and mining can both significantly restrain employment. As to education and research & development, the impact of robot application on employment is not significant.

The impact of robot application on labor force employment across industries is not consistent for the possible reason that endowment characteristics of industries give them different capacity to accommodate robotics. For instance, the employment effect of robotics may appear heterogenous for different levels of industry concentration. In highly concentrated industries such as production and supply of electricity, gas and water and quarrying and mining, robots are more likely to replace human labor because the industries are able to bear the cost of using a large number of robotics equipment. In industries with less concentrated markets such as agriculture, forestry, animal husbandry and fishery and manufacturing, robotics-induced eficiency advantage for enterprises, not so much dif erent in their market positions, is more likely to stimulate expansion in enterprise size and therefore generate stronger demand for non-automated jobs (Wang and Dong, 2020). In addition, the ef ects of robot application are mainly manifested during production, mostly by liberating workers from physical labor, while those in education and research & development are mainly engaged in brainwork, making possible that robots will not have a significant impact on employment in the industry.

4.3 Heterogeneity Regression and Analysis for Different Economies

5.1 Robot Penetration Analysis

Scholars conclude differently as to the impact of robotics on employment. Based on analysis of the heterogeneity conclusions of the existing literatures, this paper believes such impact in a multi-country framework may not be linear, as stated by Acemoglu and Restrepo (2020), but in an inverted U-shaped structure featuring a shift from complementary to substitutive effects. Given so, this paper tries to validate the assumption in the three ways of introducing the perspective of time sequence, grouping by level of robot use, ^[1] and including into the model quadratic term of the level of robot use.

First, from the perspective of time sequence, the progress of time will drive a transformation of robots’ usage from low to high. After 2014, the global installation of robots experienced an explosive growth. This paper takes the year of 2014 as the dividing point and conducts grouped regression for the period 2008–2013 and 2014–2019, respectively. If the regression coefficients of the two groups are both positive and the coefficient value and significance of the former are higher than the later, it indicates that as time progresses, the level of robot use across the countries continues to increase and its complementary effect on employment weakens. Second, in the case of grouping by the level of robot use, different economies differ evidently in such level in the same period of time. Accordingly, this paper divides the sample economies into “low level of robot use” and “high level of robot use” for grouped regression. If the coefficient of the former is significantly positive while that of the latter is significantly negative, it further suggests the level of robot use may be the driver for the shift from complementary to substitutive effects. Third, the level of robot use and its quadratic term are included into the model. If the coefficient of linear term is significantly positive and that of quadratic term significantly negative, it can be verified that the impact of robotics on employment is not linear, but in an inverted U-shaped structure featuring a shift from complementary to substitutive effects.

To measure the level of robot use, this paper refers to the method of “robot penetration” proposed by Wang and Dong (2020) and uses the following formula:

Wherein, PR_ijt means robot penetration, rob_ijt robot stock, and lab_{ij, t = 2008} the number of employees. To simplify the calculation, this paper leaves out industrial factors in the grouping by level of robot use and uses the Formula (4) alone to calculate the economy-level robot penetration. The median of the mean robot penetration of the economies in the sample period is then taken as threshold to divide the 22 economies into two groups for grouped regression.

Panel B in Table 6 reports the 2SLS estimation results. ^[1] From the perspective of grouping by period, there exists a complementary effect between robot application and employment. In the grouped regression of “2008–2013” and “2014–2019”, the regression coefficients of robot application are both significantly positive, but the coefficient value and significance of the latter are decreased compared with the former. The consistence between the empirical results and the expectations suggests as the level of robot use increases, the complementary effect weakens, but the effect remains dominant at the current stage since the level of robot use in most economies hasn’t reached the critical value. In the sense of grouping by robot penetration, the group with low penetration shows a significant complementary effect between robot application and employment; in the group with high penetration, the substitutive effect dominates, further verifying that higher level of robot use helps drive the shift from complementary to substitutive effects.

5.2 Industry Spillover Analysis

To comprehensively understand the impact of robot application on the entire labor market, this paper includes all the industries covering robot application into a unified analysis framework. Meanwhile, to simplify the model, it digs into the labor force reallocation mechanism driven by robotics from the perspective of three industries. Specifically, it first consolidates the data on the seven industry categories involved in the sample into three industries. Then, while exploring the impact of robot application in the secondary and tertiary industries on employment in the primary industry, to remove interference from the robots in the primary industry on the estimated results, it includes the primary-industry robot application into controlled variables. Lastly, the same method is applied to discuss the impact of robot application in the primary and tertiary industries on employment in the secondary industry as well as the impact of robot application in the primary and secondary industries on employment in the tertiary industry.

Table 7 explains the pertinent estimated results. Estimated results of the variables along the main diagonal are significantly positive in mutual verification with the aforesaid regression results of industry heterogeneity. To be specific, column (1) lists the impact of robot application in the secondary and tertiary industries on employment in the primary industry. The result shows the application in the secondary industry will produce a siphon effect to curb the labor flow into the primary industry, while the application in the tertiary industry has an impact on the primary-industry employment that is not significant. Column (2) reports the impact of robot application in the primary and tertiary industries on employment in the secondary industry. According to the result, introduction of robots into the primary industry will generate a crowding-out effect to boost employment in the secondary industry, while the introduction into the tertiary industry also produces a certain degree of siphon effect on the secondary-industry employment. Column (3) shows the impact of robot application in the primary and secondary industries on employment in the tertiary industry. The result indicates that robots introduced into the primary and secondary industries will form a crowding-out effect to promote the shift of labor force towards the tertiary industry. In general, the inter-industry spillover effect of robot application is manifested as the layer-by-layer crowding-out effect from the primary industry to the tertiary industry and the layer-by-layer siphon effect the other way around.

5.3 Transmission Effect of the GVC Position

Robot application affects not only the reallocation of labor force across industries, but also labor force employment across upstream and downstream economies along the global value chain (GVC). To identify whether the introduction of robots into GVC upstream economies produces any impact on downstream employment and whether it works the other way around, this paper constructs the following models:

Wherein, i-3 means the economy three positions behind the GVC position of economy i, and i+3 the economy three positions ahead of the GVC position of economy i; other variables and letters have the same meaning as in the previous context. This paper uses data disclosed in the World Input-Output Database (WIOD 2016) and the method proposed by Koopman et al. (2012) to measure the GVC position index of the economies in 2014, which is used to represent the average GVC position index of the economies in 2008–2019.

Table 8 summarizes the estimation results of transmission effects along the industrial chain. Columns (2) and (4) include local robot application ( ln rob ) into the regression model as a controlled variable. Before and after the inclusion, in the regression of the upstream-to-downstream transmission effect and the downstream-to-upstream transmission effect, the regression coefficient value and significance of the core explanatory variable haven’t changed fundamentally, verifying that the regression results are robust and reliable. Specifically, the regression coefficient of ln rob_{i + 3} in column (2) is significantly negative, indicating the introduction of industrial robots into upstream economies in the GVC inhibits the downstream job growth. A possible reason is that robot application eases the pressure of labor cost for upstream economies that therefore are less motivated for offshore outsourcing of industries, which goes against the job growth in downstream economies. The ln rob_{i − 3} regression coefficient in column (4) is negative but not significant, suggesting the indistinct impact of industrial robots introduced into GVC downstream economies on upstream employment.