Industrial robot applications and individual migration decision: evidence from households in China
IRA and labor market
Since the industrial revolution, the impact of Industrial Robot Applications (IRA) on the labor market has been a contentious topic among scholars, primarily focusing on the substitution and creation effects. The substitution effect posits that robotics primarily supplant low-complexity, repetitive tasks, thereby boosting productivity. However, this can also engender a rise in unemployment as workers are displaced from their roles (Acemoglu and Restrepo 2018). Technological advancements may lead to “technological unemployment,” with Frey and Osborne (2017) predicting that about 47% of U.S. occupations could be automated within the next 10–20 years. On the other hand, the creation effect suggests that technological advancements generate new jobs through increased production efficiency and output (Pissarides 2000; Berg et al. 2018). Aghion and Howitt (1994) noted that while technology replaces some jobs, it also creates new opportunities by increasing demand for high-skilled labor. Supporting this view, studies by Acemoglu and Restrepo (2019) and (Gregory et al. 2016) show that technological progress can positively impact employment, with new job positions emerging from technological advancements. The overall effect of IRA on the labor market hinges on the relative size of these two effects. In the short term, the substitution effect may dominate, potentially resulting in a contraction of employment opportunities as robotics displace certain job functions. However, over longer periods, this effect may diminish, and the creation effect can prevail, leading to a net positive impact on employment. Aghion et al. (2020) argued that concerns over “machine substitution” are unwarranted as job creation significantly outweighs job destruction. However, if the creation and substitution effects are balanced, the overall impact on employment remains uncertain (Dauth et al. 2017). Ultimately, the interplay between the substitution and creation effects not only reshapes the labor market but also propels industrial upgrading Graetz and Michaels (2018). This ongoing debate encapsulates the complex and dynamic impacts of technological advancements on employment, underscoring the need for nuanced understanding and adaptive policy responses to harness the benefits of IRA while mitigating adverse effects.
Influence factors of IMD
The result of labor mobility is essentially the result of individual utility maximization (Djajić 1989). The individual utility is shaped by a multitude of factors, encompassing both the elements that enhance living standards in the destination region and the unfavorable conditions in the origin region, all of which have a direct influence on IMD. The academic research on the influence factors of IMD includes three strands. The first strand is the economic factors. The level of urban economic development, industrial structure, and employment opportunities are the main economic factors of IMD. People tend to choose places with more employment opportunities and higher salaries to pursue better economic benefits and development opportunities (Autor and Dorn 2013; Lagakos 2020). Moreover, the city’s economic growth rate, industrial diversity, and innovation ability play an important role in attracting labor inflow. Cities with great economic prosperity and development potential are more likely to attract labor resources (Zhou et al. 2021). The second strand is the political factors. The policy environment and governance capacity of the government have a significant impact on the IMD of the floating population. The government’s policy measures, tax incentives, and land use policies can support the development of enterprises, thereby affecting IMD (Lavenex et al. 2016; d’Andria and Savin 2018). The stability, transparency, and honesty of the government also have an impact on IMD. A politically stable and clean government is conducive to attracting investment and migrant inflow and providing a better development environment (Dahlberg et al. 2012; Giménez-Gómez et al. 2019). The third strand is the environmental factors. The level of ecological conditions and quality of life have a significant impact on IMD. Banzhaf and Walsh (2008) found that people do “vote with their feet” on environmental quality. Factors that lead to local environmental deterioration, such as increased pollutant emissions, decreased rainfall, and global warming, all lead to the net migration of labor (Chen et al. 2022). People tend to choose places with good living environments and beautiful ecological environments to provide better living and working conditions. Further, urban infrastructure, transportation convenience, and public service facilities also affect IMD. Developed infrastructure and convenient transportation networks can provide more job opportunities and convenient living conditions (Wang 2022).
The impact of technological advancements on IMD
With the rapid advancement of technology, particularly in the fields of digital technology, artificial intelligence (AI), and robotics, we have witnessed significant changes in the structure and dynamics of the labor market. These changes have not only altered the nature of work but have also reshaped individuals’ willingness to relocate and their migration decisions. According to (Sjaastad 1962), migration decisions are an evaluation process where individuals assess the expected benefits against the costs. Within this framework, the rise of digital work environments, such as the prevalence of remote work, has reduced the need for physical relocation, thereby potentially affecting the economic motivations for moving. Nazareno and Schiff (2021) further argue that the development of AI and automation technologies may lead to the disappearance of certain occupations, especially those highly susceptible to automation. This shift in occupations forces individuals engaged in these professions to consider relocating to other regions to seek new employment opportunities or to transition into industries less affected by technology (Li et al. 2020; Ling et al. 2024). Additionally, the introduction of robotics has replaced a significant number of low-skilled jobs in some areas, as demonstrated by (Acemoglu and Restrepo 2020), leading to a migration of labor towards regions or industries with less technological penetration. At the same time, the application of these technologies has also enhanced the productivity and attractiveness of certain areas, drawing skilled workers towards these high-tech regions (Wang et al. 2020; Ling et al. 2023). In summary, the development of digital technology, AI, and robotics has had profound impacts on the labor market and individuals’ willingness to relocate. These technological transformations not only change the nature of work but also affect individuals’ geographic and occupational mobility. Therefore, understanding how these technologies shape new trends in the labor market is crucial for developing effective labor and migration policies.
In summary, while the existing studies mostly concentrate on the impact of IRA and the influence factors of IMD, there is relatively little research on the direct relationship between them. As a transformative technological force, the IRA undoubtedly exerts profound influences on the labor market. However, the precise mechanisms through which it affects IMD have yet to be comprehensively examined. IRAs not only directly affect job opportunities through the creation effect and substitution effect, forcing the floating population to reassess the factors of IMD. IRA also indirectly affects IMD by influencing wages, economic growth rate, industrial structure, innovation vitality, and other aspects. Therefore, it is necessary to further explore the impact of IRA on IMD, which is the focus of this paper.
Research hypothesis
IRA and IMD
Incorporating established theories provides a robust framework for understanding the impact of industrial robots (IRA) on individual migration decisions (IMD). The Economic Theory of Migration by (Sjaastad 1962), which views migration as a cost-benefit analysis, helps explain how changes in labor demand influenced by IRA might affect migration decisions. Additionally, the dual labor market theory suggests that developed economies have a structural demand for labor that the native population does not fill, applicable in understanding how technological advancements might create new demands for migrant labor in tech-driven sectors (Piore 2018). Furthermore, the spatial mismatch hypothesis by Ihlanfeldt and Sjoquist (1998) provides a lens to analyze how IRA-induced changes in job locations might influence migration patterns, particularly for low-income individuals. Recent advancements in artificial intelligence (AI), automation, and robotics have begun to significantly alter the economy and labor demand patterns (Autor 2015; Acemoglu and Restrepo 2019). The floating population, a key driver of regional economic growth, enhances urban agglomeration and economic expansion through cross-regional labor mobility, which provides a crucial labor supply for cities (Docquier and Rapoport 2012; Feng et al. 2024b). Policymakers are increasingly focused on the capacity of local labor markets to adapt to the disruptions brought about by new technologies and the potential of labor mobility to address issues such as persistent unemployment and regional disparities. Although the body of literature examining the relationship between IRA and IMD is still emerging, it is beginning to provide insights into these complex interactions. Frank et al. (2019) observed that new technologies alter labor demand, leading to micro-scale employment redistribution and geographic redistribution, such as worker migration. Faber et al. (2019) found that robot adoption triggers migration, primarily driven by reduced in-migration rather than increased out-migration. Bessen (2019) noted that technological advancements might cause some industries to decline and others to grow, affecting labor inflows and outflows. Conversely, Liu et al. (2023) argued that robot importation has a positive effect on migrants by creating new jobs, suggesting that the creation effect outweighs the substitution effect. Based on this integrated theoretical framework and empirical evidence, the following hypotheses are proposed:
Hypothesis 1a (Technology Substitution Hypothesis): IRA has a negative effect on IMD; the floating population may be more inclined to move to areas less dependent on industrial robots.
Hypothesis 1b (Skill Complementarity Hypothesis): IRA has a positive effect on IMD; the floating population may be attracted to areas with dense industrial robots.
Heterogeneity of IRA influencing IMD
The influence of IRA on IMD has different characteristics in different skill levels, industry structures, and geographical regions. The application of new technologies (e.g., robotics) usually requires highly technical and specialized knowledge and skills. Relatively speaking, uneducated workers are more vulnerable to the negative effects of robots than educated workers (Aghion et al. 2019). Some papers analyze directly from the perspective of skill levels. For example, (Acemoglu and Autor 2011) established a simple task model and found that automation technology can improve the employment demand for high-skilled labor, and also reduce the employment proportion of medium-skilled labor performing regular tasks. Since workers engaged in low-skilled work are not easy to be replaced by machines, automation technology has also increased the market demand for low-skilled labor, resulting in the phenomenon of employment polarization. Korinek and Stiglitz (2018) have identified that artificial intelligence tends to impede employment opportunities for low-skilled workers, attributing this to their lack of requisite skills for the AI era and an inability to swiftly adapt to the employment structure. For the differences in industry structures. IRA has a greater impact on the manufacturing industry, but less on the agriculture and service industry. The progress of automation technology reduces the employment proportion of the labor force with low and medium education levels in the manufacturing industry, which puts forward higher requirements of labor in related industries (Davoyan 2021). For the differences in geographical regions. Due to the different distribution of industrial structures and the different types of employment groups in different regions, the impact of IRA on employment is different. Differences in regional economic levels, economic policies, and systems all lead to regional differences in the employment effect of IRA (Miguelez and Temgoua 2020). According to the above analysis, the following hypothesis is proposed:
Hypothesis 2. IRA has a heterogeneous effect on IMD.
Mechanism of IRA influencing IMD
As previously mentioned, the application of robotic technology significantly affects both the availability and the nature of employment, which may subsequently influence IMD (Acemoglu and Restrepo 2020). The decision to migrate, however, is influenced by a multitude of factors, including economic opportunities, cost of living, educational resources, and the policy environment, all of which require further research and analysis (Jin et al. 2022). Therefore, a deep investigation into how these factors interact and collectively affect migration dynamics is crucial for understanding and addressing the social impacts of technological changes. This theoretical analysis explores how IRA can simultaneously generate both positive and negative effects on IMD, leading to the formulation of three distinct hypotheses (Fig. 2).
Firstly, the deployment of IRA often leads to an increase in productivity and efficiency within industries. According to economic theories such as the productivity-wage linkage theory, improvements in productivity are typically associated with higher wages (Caselli and Manning 2019). This is particularly true in sectors where high-skilled labor is required to manage and work alongside advanced robotics. The increased demand for such skilled workers often results in higher wage levels, making regions with a high concentration of IRA attractive to potential migrants. This scenario aligns with the economic pull factors described in migration theory, where individuals are drawn to areas offering better economic opportunities (Llull 2018; Monras 2020). Thus, we propose:
Hypothesis 3a: IRA has a positive effect on IMD by raising wage levels.
However, the influx of workers and the economic prosperity driven by the IRA can lead to an increase in housing prices. Regions experiencing technological booms often see a rise in the cost of living, particularly in housing markets (Rabe and Taylor 2012; Zhou and Hui 2022). This phenomenon can be explained by the supply-demand imbalance in the housing market, where an increased demand for housing from incoming workers drives up prices. High housing costs can deter potential migrants, especially those in lower-wage brackets or those who do not directly benefit from the high-tech economy (Hyatt et al. 2018). This leads to the second hypothesis:
Hypothesis 3b: IRA has a negative effect on IMD by driving the housing price up.
Lastly, the introduction of an IRA can intensify competition in the labor market. While IRA creates demand for high-skilled labor, they can also lead to job displacement, particularly among low-skilled workers (Goos et al. 2014; Beaudry et al. 2016). This displacement results from the automation of tasks previously performed by human labor. Furthermore, even among high-skilled sectors, the competition for specialized positions can become fierce, as more individuals qualify themselves for the limited number of high-tech roles. This increased competition can make it challenging for potential migrants to secure employment, thereby acting as a deterrent (Xu et al. 2020). This observation forms the basis of our third hypothesis:
Hypothesis 3c: IRA has a negative effect on IMD by intensifying employment competition.
Methodology and data
Empirical model
This paper tries to explore the impacts of IRA on IMD and quantitatively tests the actual effect of IRA on IMD. Drawing from previous studies on migration and availability of microdata, we adopt the conditional logit model to identify the influence of IRA on IMD (McFadden 1973). The conditional logit model has a sound microeconomic foundation, which is built on an individual utility maximization framework (Okamoto 2019). According to the principle of individual utility maximization, the floating population chooses the city with the greatest utility among potential alternative cities by combining their conditions and urban characteristic information. The utility function of the floating population is set as follows.
$$\beginarraylU_ijt=\alpha IRA_jt+\gamma ^\prime X_jt+\mu ^\prime Z_ijt\\\qquad+\,\varepsilon _ijt\left(i=1,2,\cdot \cdot \cdot ,N\rm;\,j=1,2,\cdot \cdot \cdot ,J\right)\endarray$$
(1)
Here, \(U_ijt\) represents the individual utility of migrant \(i\) at city \(j\) in time \(t\). \(\rmIRA_jt\) represents the IRA of city \(j\) in time \(t\). \(X_jt\) represents the economic characteristic control variables of city \(j\) in time \(t\). \(Z_ijt\) represents the individual characteristic control variables of migrant \(i\). \(\varepsilon _ijt\) is the random error term. Based on utility maximization, whether the migrant \(i\) chooses to flow into the city \(j\) meets the following conditions:
$$\rmchoice_ijt=\left\{\beginarrayc1,\,\forall\, k\,\ne\, j\,E\left[U_ijt\right] >\, E\left[U_ikt\right]\\ 0,\,\exists\, k\,\ne\, j\,E\left[U_ijt\right]\le\, E\left[U_ikt\right]\endarray\right.$$
(2)
When \(E[U_ijt]\, >\, E\left[U_ikt\right]\), \(\rmchoice_ijt=1\) means that the migrant \(i\) chooses city \(j\) as the destination. On the contrary, \(\rmchoice_ijt=0\) means that the migrant \(i\) dose not choose city \(j\) as the destination. We use the conditional logit model to estimate the parameters of Eq. (3). The estimated parameters reflect the impact of characteristic variables on the probability of IMD. When the estimated parameters are positive, it can be explained that the greater the probability of IMD. At the same time, the positive and negative coefficients also reflect the pull and resistance to the inflow of migrants.
$$Prob(choice_ijt=1)=\fracexp \left(\alpha Robots_jt^CH+\gamma ^\prime X_jt+\mu ^\prime Z_ijt+\varepsilon _ijt\right)\sum _j=1^Jexp \left(\alpha Robots_jt^CH+\gamma ^\prime X_jt+\mu ^\prime Z_ijt+\varepsilon _ijt\right)$$
(3)
Data structure
McFadden (1973) posits that migrants have the option to select any city as their destination. However, in reality, due to geographical and personal limitations, as well as the uncertainty of migration, migrants only choose their final destination from finite city samples. For this purpose, we calculate the number of cities where the migrants in each city flow into other cities and consider these cities as alternative cities for the floating population in each city. Based on this, each migrant has its corresponding set of alternative cities (represented by \(C_i\)) when selecting destinations. We create one dummy variable for each destination, resulting in \(\sum C_i\) dummy variables and \(N\) groups to be estimated (e.g., for migrant \(i=1\), he has \(C_1\) alternative cities, then we need to create \(C_1\) dummy variables). Therefore, the data structure of the conditional logit model is shown in Table 1.
Data sources
This paper mainly uses three data sources. The first data source is data for industrial robots and Industrial Statistical Yearbook. We use city-level over-time variations in employment patterns to construct Bartik-style measures of IRA. The application data for industrial robots at the industry level is acquired from the IFR. The number of industry employees and Bartik’s shift share are calculated using the industrial enterprise module of the national economic census data and provincial statistical yearbooks. The second data source is data for city characteristic information. The data are collected from China Statistical Yearbook for Regional Economy and China City Statistical Yearbook. The third data source is data for the floating population, which comes from China Migrants Dynamic Survey (CMDS) released by the National Health Commission. The data is sampled by stratified, multi-stage, and proportional PPS methods. The survey covers multiple aspects such as the family members, income and expenditure, mobility and employment, social integration and health status of the floating population. It is a nationally representative professional microdata that focuses on the economic and social statistical characteristics of the floating population in China. On this basis, we match CMDS data with industrial robot data and city characteristic data and finally select 85,761 migrants and 359,893 observations (\(N\) = 85,761, \(\sum C_i\) = 359,893).
Variable measurement
-
1.
Dependent variable: The dependent variable is \(\rmchoice_ijt\), which represents whether migrant \(i\) choose city \(j\) as the destination. It is a dummy variable. Figure 3 shows China’s geography distributions of migrants based on CMDS data in 2018. We can find that most of the floating population flows to provincial capital cities or developed cities, as these areas typically provide more employment opportunities and higher income levels. At the same time, provincial capital cities and developed cities have better infrastructure and public services, more education and training opportunities, and richer social resources. These factors have attracted the floating population to seek better employment and living conditions in these places. Moreover, the Central tendency trend of floating population in the east remains unchanged, while it spreads to the central and western regions.
-
2.
Independent variable: The dependent variable is \(\rmIRA_jt\), which represents the exposure to robots variables of city \(j\). Following Acemoglu and Restrepo (2020), we use city-level over-time variations in employment patterns to construct Bartik-style measures of IRA (Bartik 1991; Goldsmith-Pinkham et al. 2020). The IFR collects data on industrial robots by industry. We interact IFR data with the China Industrial Statistical Yearbook data to obtain the installation density data of industrial robots in various industries. Then, we select 2012 as the benchmark year to calculate the allocation share of industrial robots in each industry in each city. Based on this, we obtain the city-level exposure to robots variables. The specific measurement methods are as follows:
$$IRA_jt=\mathop\sum \limits_h=1^H\fracEmp_jh,t=2012Emp_h,t=2012* \fracIRA_htEmp_h,t=2012$$
(4)
Here, \(H\) represents industry numbers. \(IRA_ht\) represents the industrial robots installation of industry \(j\) in time \(t\). \(Emp_h,t=2012\) represents the employment of industry \(j\) in the benchmark year. \(\fracEmp_jh,t=2012{Emp_h,t=2012}\) is Bartik’s shift share. After the measurement, it is possible to obtain China’s geography distributions of industrial robots. Figure 4 shows the geography distributions of IRA in 2012 and 2019. We can find that the development of IRA has made significant progress in recent years, with more and more cities introducing robot technology and applying it to industries such as industrial manufacturing, and service industry. Additionally, the distributions of IRA exhibit regional heterogeneity over time, especially between southeastern and northwestern regions.
3. Control variables: To control other factors that may affect the IMD of the floating population. We select control variables from two aspects: economic characteristic variables and individual characteristic variables.
Generally speaking, city-level economic factors are the most important factors affecting IMD. Therefore, we add economic characteristic control variables to the empirical model, including per capita GDP, industrial enterprise intensity, government financial expenditure scale, average wage of employees, population size, education, medical services, industrial structure, and fixed assets investment level. In addition, more and more scholars are paying attention to the impact of environmental quality and housing prices on IMD (Banzhaf and Walsh 2008; Zhou and Hui 2022). Therefore, we also add PM2.5 concentration and annual average housing price level to the empirical model to control the impact of environmental quality and housing price level on IMD. The floating population also considers the transportation costs, psychological costs, living environment expectations, and quality of life expectations when choosing the city as the destination with a consideration of all potential alternative cities. Thus, we add three variables to the empirical model, including whether to move within the same province, whether move to a city with better environmental quality, and the ratio of household income to the average wage of the chosen city. The specific definition and descriptive statistics of economic characteristic variables are shown in Table 2.
The individual characteristics of the floating population sample also affect IMD. We collect individual characteristic variables such as age, gender, education level, health status, marital status, household registration type, nature of employment unit, whether engage in individual business, and whether engage in routine tasks for heterogeneity analysis. The specific definition and descriptive statistics of individual characteristic variables are shown in Table 3.
Empirical results and discussion
Baseline regression
Table 4 demonstrates the benchmark estimation results of the conditional logit model. Column (1) shows the results that control only fixed effect without adding any control variables. The coefficient of IRA is significantly negative at the 1% level, which means migrant workers consider industrial robot applications when choosing the destination. While columns (2) to (4) are the estimated results with control variables added. Based on column (1), columns (2) to (4) control sequentially economic characteristic variables and individual characteristic variables. We can see the results are still significant in IRA at the 1% level. Intuitively, for every 1 percentage point increase in IRA, the probability of the floating population choosing the city significantly decreases by 0.58 percentage points. Column (5) is the regression results after standardizing the variables, which can reflect the relative impact of IRA on IMD. The results indicate that for every standard deviation increase in IRA, the probability of the floating population choosing the city decreases by 0.23 times, indicating that IRA inhibits the inflow of labor. The possible explanation is that the substitution effect of IRA is greater than the creation effect, along with a phenomenon of “machine replacement” (Qian et al. 2023). As reported in the World Robotics 2023 Report, China holds the distinction of being the world’s largest market for robots, experiencing an annual increase of 13% in robot installations from 2017 to 2022. According to data from the National Bureau of Statistics and the Ministry of Industry and Information Technology, annual industrial robot density has grown by 32.22% each year (2017–2022). However, employment in the manufacturing sector has shown a declining trend, with an annual decrease of 0.6%. This suggests a strong correlation between the rise in automation and robotics in manufacturing and the reduction in human labor within the same sector. The increase in robot density, particularly at such a high annual rate, indicates a shift towards more automated production processes, which likely contributes to the decrease in manufacturing jobs. This trend may reflect the broader impact of automation on labor markets, where technological advancements are leading to a transformation in the nature of work and employment opportunities in the manufacturing industry (Frank et al. 2019). For the floating population, whether they can find satisfactory jobs is one of the main factors considered for their IMD. The improvement of IRA level reduces employment opportunities and increases employment difficulty, which leads to a decrease in the probability of the floating population choosing this city (Dixon et al. 2021). The findings verify that IRA has a negative effect on IMD, thus confirming Hypothesis 1a.
The estimated results of control variables are also consistent with economic intuition. According to the economic characteristic variables, the regression coefficients of MS, IS, and IFA are significantly positive at the 1% level, which shows that the level of medical services, industrial structure, and fixed assets investment improve the attractiveness of cities to the floating population. The regression coefficients of PM2.5 concentration and annual average housing price level are significantly negative at the 1% level. The higher air pollution and housing price levels, the lower the probability of the floating population choosing the city. This indicates that migrants also consider urban environmental quality and housing prices when choosing the destination, and are increasingly paying attention to physical and mental health and quality of life. In addition, according to the individual characteristic variables, the regression coefficients of TPM and BEQ are significantly positive at the 1% level, indicating that the floating population is more inclined to cross-province flow to cities with better environmental quality. The regression coefficient of RHI is significantly negative at the 1% level, indicating that the floating population is more inclined to move to cities with higher relative wage levels.
Endogenous discussion
To obtain an accurate estimation of the research, it is necessary to address potential endogeneity issues that may exist in the empirical model. Endogeneity problems typically arise from omitted variables and reverse causality. For example, the response of different cities and enterprises to IRA often varies, which not only affects the development of industrial robots but also affects decision-making and employee employment. These factors are difficult to measure. IRA and IMD mutually influence each other. The changes in the inflow of the floating population may affect IRA, subsequently impacting the progression of industrial robotics. Furthermore, the availability of data at the city level makes it impossible to control for all significant factors in the model. We use the two-stage least square method (2SLS) to deal with the endogeneity problem.
Following Acemoglu and Restrepo (2020), current manufacturing countries have shown a high convergence in the application scale of new production technologies and equipment, so it is reasonable to use the installation of industrial robots in other countries in the same industry as the instrumental variables. Based on this, we use the exposure to robot variables in the United States to construct Bartik-style instrumental variables. On the one hand, although China currently has the world’s largest installation of industrial robots around the world, the import dependence on industrial robots used in China is relatively high, with over 70% of industrial robots still supplied by foreign robot companies. As one of the main sources of imports, the shipment volume of industrial robots in the United States affects the installation of industrial robots in China. On the other hand, the research and development of robotics technology have become a global phenomenon, and the current manufacturing countries are showing a high degree of convergence in IRA. The development trend of industrial robots in the United States can reflect the technological progress trend of the industry. Therefore, it can be considered that the IRA in the United States has a strong correlation with the IRA in China. Its impact on China’s labor market mainly reflects the technical characteristics of similar industries, which meets the correlation requirements of instrumental variables estimation. At the same time, there is no evidence that the IRA in the United States affects the changes in the labor market structure of China’s manufacturing industry, which meets the exclusive constraints of instrumental variables estimation.
We conduct 2SLS regression analysis in two stages. In the first stage, we use instrumental variables to explain the IRA in the candidate city and obtain the predicted value of dependent variables. In the second stage, the predicted variables are used as the independent variables to analyze the impact of IRA on IMD through the conditional logit model. The 2SLS regression result is shown in Table 5. The coefficients of IRA are still significantly negative, which means the results are robust after considering endogeneity.
Poisson regression
We further investigate the impact of IRA on IMD from a meso-level perspective through Poisson regression, as the meso-level dependent variable is dichotomous (Carlsen et al. 2021). Specifically, we collate CMDS data and obtain city panel data from 350 prefecture-level cities.
Table 6 reports the results of the Poisson regression. The coefficient of IRA in column (1) is significantly negative at the 1% level. This result is consistent with the research conclusion of benchmark regression based on micro-level data. Additionally, we use instrumental variables estimation to conduct poisson regression, the results are shown in Table 6 column (3). We find that the coefficient of IRA is still significantly negative, indicating that IRA maintains a robust and negative impact on IMD. Column (4) shows the estimation result of Hilbe’s two-step method (Hilbe 2011). The estimated coefficient of IRA is still significantly negative, and the estimated residuals in the first stage pass the 10% significance test. The conclusion is robust after considering Poisson regression.
Further robustness tests
Three further robustness tests are conducted to ensure the validity of our empirical findings. First, replace core variables. We use the data of industrial robots in the secondary industry to recalculate IRA. In addition, we have expanded the sample size to recalculate the IRA in various cities using 2010 as the benchmark year. The results are shown in Table 7 columns (1) and (2). Second, consider the lag of impact. We investigate the impact of IRA with a first-order lagged on IMD in the model, and the estimated results are shown in Table 7 column (3). Third, consider the dynamic process. We include first-order lagged term of IRA in the model for estimation. The results are shown in Table 7 column (4). Upon conducting additional robustness tests, we find that the coefficients and significance levels of the core variables are largely consistent with the benchmark regression model.
Heterogeneity in individual characteristic
We focus on the heterogeneity in the year of migration, education level, health status, age, gender, marital status, number of family members, whether move within the same province, household registration type, nature of employment unit, and whether they engage in routine tasks for floating population affected by IRA. Given the constraints of the conditional logit model, which does not readily accommodate the inclusion of individual characteristic variables, we have devised an approach that incorporates interaction terms between IRA and these individual characteristics. By analyzing the interaction term, we investigate the heterogeneity effect of IRA on IMD. Table 8 panels A and B report the results of individual heterogeneity regression.
The coefficients of IRA still present a negative correlation, indicating the robustness of the model. From the perspective of individual heterogeneity, column (1) reports the heterogeneity influence of migrant years. The interaction coefficient is significantly positive, indicating that the negative impact of IRA on IMD has shown a decreasing trend in recent years. Columns (2) and (3) report the heterogeneity influence of education level and health status. The interaction coefficients are significantly positive, indicating that IRA has a relatively small impact on highly educated and physically healthy migrants, which reflects the competitive advantage of high-skilled level and physical health. Columns (4) to (6) report the heterogeneity influence of age, gender, and marital status, with significantly negative interaction coefficients, indicating that IRA has a significant impact on the IMD of older, male, and unmarried floating populations. The floating population of the younger generation of unmarried women does not over-think the impact of IRA during the migration process. This type of floating population has the ability to quickly learn and adapt to new environments, and strengthen their skills to find suitable jobs. The interaction coefficient in column (7) is significantly negative, indicating that the families with fewer members tend to have a greater tendency to migrate to cities with higher IRA. The interaction coefficient in column (8) is not significant. The interaction coefficient in column (9) is significantly positive, indicating that the agricultural registered residence labor force tends to flow to cities with high IRA. Columns (10) and (11) report the heterogeneity influence of the nature of the employment unit. The interaction coefficient of state-owned enterprises is significantly positive, while the interaction coefficient of individual businesses is significantly negative. This indicates that IRA has a smaller impact on the IMD of state-owned enterprise labor, but a greater impact on the IMD of individual businesses. A possible explanation is that state-owned enterprises have stable job positions and require employees to possess a certain level of skills. The labor force of state-owned enterprises is less likely to consider the impact of IRA. Column (12) reports the heterogeneity influence of whether engage in routine tasks and the interaction coefficient is significantly negative. Industrial robots have a great substitution effect on routine tasks, which leads to a negative impact on such workers engaging in routine tasks. The findings verify that IRA has an individual heterogeneous effect on IMD, thus confirming Hypothesis 2.
Heterogeneity in city characteristic
In the heterogeneity analysis of city characteristics, we focus on the differences in geographical location, city size, marketization level, wage level, and environmental quality (Shen et al. 2024; Wu et al. 2024). We also construct an interaction term between IRA and city characteristic variables. Table 9 reports the results of city heterogeneity regression. Column (1) reports the heterogeneity influence of geographical location (coastal = 1, others = 0), and the interaction coefficient is significantly negative. A possible explanation is that coastal areas are generally economically developed areas, with more complete competition mechanisms between enterprises and higher levels of IRA, thereby suppressing the inflow of labor. According to an IFR report, the density of industrial robots in China stood at 246 units per 10,000 workers in 2020. Notably, the density escalated in Shanghai (coastal area) is 383 units per 10,000 workers, while the density in Xi’an (hinterland area) is less than 150 units per 10000 people. The higher concentration of robots in places like Shanghai not only reduces the number of available jobs but also shifts the job market towards more specialized roles that require specific skills not possessed by the majority of the labor force. Consequently, workers who lack these specialized skills may find fewer employment opportunities in these regions, leading them to seek jobs in less automated areas where traditional labor is still in demand (Frank et al. 2019). Moreover, the intense competition among enterprises in coastal areas, fueled by technological advancements, can lead to a more volatile job market. This environment might discourage workers from migrating to these areas due to perceived lower job security and stability, further inhibiting the inflow of labor to coastal regions (Ronzoni et al. 2021). Column (2) reports the heterogeneity influence of city size (city with a permanent population exceeding 5 million = 1, others = 0), and the interaction coefficient is significantly positive. This indicates that IRA has a small impact on the IMD in mega-cities, which reflects the scale effect and aggregation level of big cities that are more likely to attract labor inflows. Column (3) reports the heterogeneity influence of marketization level (high marketization = 1, others = 0), and the interaction coefficient is significantly negative. With the advancement of the marketization process, the more intense the job competition, the more it inhibits the inflow of labor. Column (4) reports the heterogeneity influence of wage level, and the interaction coefficient is significantly positive. This indicates that the floating population chooses cities with high IRA for higher wage levels. Column (5) reports the heterogeneity influence of environmental quality, the IRA and PM 2.5 interaction coefficient is significantly negative. This indicates that the floating population considers environmental quality when making migration decisions. The high cost of living, fierce competition in the job market, and high environmental pressure in cities further limit the inflow of labor through the IRA (Jin et al. 2022). The findings verify that IRA has a city heterogeneous effect on IMD, thus confirming Hypothesis 2.
Mechanism analysis
According to the previous theoretical analysis, we use city panel data and individual sample data to explore the influence mechanism of IRA on IMD. City-level and individual-level mediating variables are considered to construct the following two models (Feng et al. 2024a):
$$City_jt=\alpha IRA_jt+\gamma ^\prime X_jt+\upsilon _t+\varepsilon _jt$$
(5)
$$Person_ijt=\beta IRA_jt+\gamma ^\prime X_jt+\mu ^\prime Z_ijt+\upsilon _t+\varepsilon _ijt$$
(6)
Here, Eq. (5) is a city-level panel regression model. \(\rmCity_{{jt}}\) are mediating economic characteristic variables, including city wage level, PM2.5 concentration, and housing price level. Equation (6) is an individual-level panel regression model. \(\rmPerson_{{ijt}}\) are mediating individual characteristic variables, including individual wage level, working hours, difficulty in finding a job, and difficulty in purchasing a house. Other settings are consistent with the benchmark model.
Table 10 reports the results of city-level panel regression. We can find that IRA significantly improves the city wage level and housing price level but has no significant impact on PM2.5 concentration. Table 11 reports the results of individual-level panel regression. We can find that IRA significantly increases individual wage level. At the same time, IRA also increases working hours, difficulty in finding a job, and difficulty in purchasing a house. It is not difficult to find that IRA affects the average city wage level and individual wage level after labor inflows. The effective wage income of individuals is the core issue that the floating population mainly considers when making migration decisions. Thus, hypothesis 3a is confirmed. However, the floating population also considers other factors, such as the housing price level of alternative cities, the difficulty in finding a job, and the difficulty in purchasing a house. The increase in these indicators reduces the probability of labor inflow, which requires a comprehensive evaluation of the utility among potential alternative cities. Thus, hypothesis 3b and hypothesis 3c are confirmed.
Indeed, government managers can attract the inflow of high-quality human capital by increasing wages and benefits. At the same time, government managers should consider the impact of environmental quality, working hours, cost of living, housing prices, and other factors on the labor force. In addition, with the continuous improvement of wage income, the demand for quality of life is increasing. In other words, when the wage level reaches a certain level, the attraction of high wage levels to labor is less than the effect of factors such as environmental quality and working hours. Therefore, government managers should properly handle the impact of new technologies on modern cities and the labor market. Find effective resource allocation points to more effectively attract high-quality human capital.
link