Industrial robot applications and individual migration decision: evidence from households in China

China’s rapid adoption of industrial robots within the broader Fourth Industrial Revolution has raised concerns about human–machine substitution and its consequences for labor markets and migration. With China hosting over half of global robot installations in 2022 and experiencing large-scale internal migration (a floating population of 376 million in 2020), understanding how IRA shapes individual migration decisions (IMD) is critical. This study asks whether and how exposure to industrial robots in destination cities influences migrants’ location choices, and through what mechanisms (e.g., wages, costs of living, employment competition). It leverages city-level variation in IRA exposure and microdata on migrants to empirically test competing hypotheses: IRA may deter migrants via substitution effects (H1a) or attract them via skill complementarity and higher wages (H1b), with heterogeneous impacts across individuals and cities (H2), and mechanisms through wages (H3a), housing prices (H3b), and employment competition (H3c).

The literature on IRA and labor markets highlights competing substitution and creation effects. Robotics can displace routine, low-complexity tasks and reduce employment (Acemoglu and Restrepo 2018, 2020; Frey and Osborne 2017), yet also create new, often higher-skill jobs and increase productivity (Pissarides 2000; Aghion and Howitt 1994; Acemoglu and Restrepo 2019; Graetz and Michaels 2018). Net effects vary over time and context. Research on IMD identifies economic (employment opportunities, wages, growth, innovation), political (policy environment, governance), and environmental (quality of life, pollution) determinants of migration (Autor and Dorn 2013; Dahlberg et al. 2012; Banzhaf and Walsh 2008). Technological change (AI, automation, robotics) reshapes occupational structures and relocation incentives, potentially reducing migration via remote work or prompting migration due to job displacement or attraction to high-tech hubs (Sjaastad 1962; Acemoglu and Restrepo 2020; Wang et al. 2020). Emerging studies link technology adoption to migration flows, often via reduced in-migration to automated areas (Faber et al. 2019), though positive effects appear where creation effects dominate (Liu et al. 2023). Heterogeneity spans skill levels (polarization, differential impacts on low- vs high-skill workers), industries (larger effects in manufacturing), and regions (variation by industrial structure and policy environment). This study addresses a relative gap by directly connecting IRA exposure to individual-level migration choices and examining mechanisms.

Empirical strategy: A conditional logit model (McFadden 1973) grounded in individual utility maximization is used to estimate how destination-city IRA exposure affects migrants’ location choices among feasible alternative cities. For migrant i and city j at time t, utility depends on city IRA exposure, city-level characteristics, and individual characteristics; the probability of choosing city j is modeled via the conditional logit choice probability. Data structure enumerates, for each migrant, the set of alternative cities they plausibly consider (C_i), generating stacked choice observations per migrant-city pair with a binary indicator for the chosen destination. Measurement of IRA (shift-share/Bartik-style): Following Acemoglu and Restrepo (2020) and Goldsmith-Pinkham et al. (2020), city-level IRA exposure is constructed by interacting national/industry-level robot installations (IFR) with baseline (2012) city-industry employment shares from the China Industrial Statistical Yearbook and economic census, yielding a Bartik-style exposure measure at the city-year level. Additional variants include secondary-industry-only exposure and a 2010 base-year construction. Data sources: (1) Industrial robots: IFR industry-level robot installations, combined with employment data from the Industrial Statistical Yearbook and census; (2) City characteristics: China Statistical Yearbook for Regional Economy and China City Statistical Yearbook; (3) Migrants: China Migrants Dynamic Survey (CMDS) 2018, a large, nationally representative micro-survey (stratified multi-stage PPS) on demographics, employment, income, mobility, and health. Final matched sample: 85,761 migrants and 359,893 migrant-city observations. Variables: Dependent variable is choice_ijt (1 if migrant i chooses city j). Key independent variable is city-level IRA exposure (log). City controls include per capita GDP, industrial enterprise density, fiscal expenditure share, average wage (log), population size, education supply, medical services, industrial structure, fixed asset investment, PM2.5, housing prices, plus migration-cost proxies (same-province move, moving to better air quality, household income-to-city wage ratio). Individual characteristics used for heterogeneity include age, gender, education, health, marital status, household registration (hukou), employer type (SOE), self-employment, and routine task indicator. Addressing endogeneity: A two-stage least squares (2SLS) approach instruments Chinese city IRA with shift-share exposure based on U.S. industry-level robot adoption, leveraging convergence and import dependence, while assuming exclusion via lack of direct effects of U.S. IRA on Chinese city labor market composition. First-stage predicts city IRA; second stage plugs predicted IRA into the conditional logit for IMD. Additional estimators and tests: Poisson regressions at the city panel level (350 cities) to examine meso-level migration counts; IV-Poisson and Hilbe two-step methods; robustness checks replacing core variables (secondary-industry IRA; 2010 base year), including lagged IRA and dynamic specifications. Heterogeneity is analyzed via interactions between IRA and individual- and city-level characteristics (e.g., coastal, city size >5 million, marketization level, wage level, environmental quality). Mechanism analysis uses city- and individual-level panel models to test mediators: wages, PM2.5, housing prices, working hours, difficulty finding a job, and difficulty purchasing a house.

Baseline effects: Across conditional logit specifications, IRA significantly reduces the probability that a migrant selects a given city. In the fully controlled model, a 1 percentage point increase in IRA is associated with a 0.58 percentage point decline in the selection probability; a one standard deviation increase in IRA reduces the probability by 0.23 times. Representative coefficients include IRA = −0.5798 (p<0.01) and standardized IRA = −0.2276 (p<0.01). This supports Hypothesis 1a (technology substitution dominating the creation effect in the short run). Controls: Better medical services, advanced industrial structure, and higher fixed asset investment significantly increase city attractiveness; higher PM2.5 and higher housing prices significantly deter migration. Migrants are more inclined to cross provinces toward cities with better environmental quality, and they respond to relative wage levels. Endogeneity and robustness: 2SLS estimates using U.S.-based robot exposure as an instrument confirm negative and significant effects of IRA on IMD (e.g., coefficients around −0.03 to −0.034, p<0.01 in later-stage models). Poisson city-level regressions show IRA effects are negative and robust (e.g., −0.0856, p<0.01; IV-Poisson −0.0556, p<0.01; Hilbe two-step −0.0879, p<0.01 with first-stage residuals significant at 10%). Robustness checks using secondary-industry IRA (−0.1192, p<0.01), a 2010 base year (−0.0931, p<0.01), and lagged IRA (−0.1226, p<0.01) uphold the findings; a dynamic IRA term is small and not significant in one specification. Heterogeneity (individual): The negative IRA effect is attenuated for more recent migrants and those with higher education and better health (positive interactions), but stronger for older, male, and unmarried migrants (negative interactions). Migrants with fewer family members are more likely to choose high-IRA cities (negative interaction with family size). Agricultural hukou exhibits a positive interaction (more tolerant of high IRA). SOE employees are less deterred (positive interaction), whereas self-employed/individual businesses are more deterred (negative interaction). Routine-task workers face stronger negative effects (negative interaction), consistent with substitution risk. Heterogeneity (city): Negative IRA effects are stronger in coastal regions and in high-marketization environments (negative interactions), and mitigated in mega-cities (>5 million population) and where wage levels are higher (positive interactions). Poor environmental quality (IRA × PM2.5) further deters migration (negative interaction). Mechanisms: City-level IRA significantly raises average city wages (coef ≈ 1.424, p<0.01) and housing prices (≈ 0.507, p<0.01) but shows no significant effect on PM2.5. Individual-level regressions show IRA increases individual wages (≈ 0.719, p<0.01), working hours (≈ 2.004, p<0.05), difficulty finding a job (≈ 0.037, p<0.10), and difficulty purchasing a house (≈ 0.084, p<0.01). These results confirm H3a (wage channel), H3b (housing cost channel), and H3c (employment competition channel).

The results indicate that higher IRA exposure in destination cities deters migrants, primarily via short-run substitution pressures, higher housing costs, and intensified job competition, even as wages rise. This reconciles the mixed theoretical expectations by showing that wage gains do not fully compensate for higher living costs and employment frictions for many migrant groups. The heterogeneity patterns highlight that more educated, healthier, younger, unmarried women, agricultural hukou holders, and SOE employees are less sensitive to IRA risks, while routine-task workers, older, male, and unmarried migrants are more vulnerable. City characteristics moderate effects: large cities and high-wage environments can offset some deterrence, while coastal, high-marketization, and polluted cities amplify it. These findings inform labor market and urban policy by revealing how automation reconfigures migration incentives and spatial labor allocation, suggesting targeted policies to mitigate inequality and facilitate adaptation in highly automated regions.

This study links city-level industrial robot exposure to individual migration choices in China using a conditional logit framework with rich controls, addressing endogeneity and conducting extensive robustness, heterogeneity, and mechanism analyses. Main contributions include: (1) establishing that IRA reduces the likelihood of migrants selecting highly automated cities; (2) documenting systematic heterogeneity across individuals and cities; and (3) identifying wage, housing cost, and employment competition as key mechanisms. Policy implications emphasize combining skills training and social protection with measures to manage living costs, working hours, and environmental quality to attract and retain human capital in automated cities. Future research should improve micro-level alignment of robot adoption data with individual outcomes (e.g., firm/occupation-level IRA), incorporate dynamic adjustment paths, and expand beyond China to cross-country comparative analyses to assess generalizability across institutional settings.

Two primary limitations are noted: (1) Data granularity mismatch—IRA is measured at the industry and city levels and matched to micro-level migrant decisions, which may obscure sector- or occupation-specific effects. Finer-grained data on robot adoption at firm or occupation levels would improve identification. (2) External validity—evidence is from China; results may not generalize to countries with different institutions, labor markets, and migration regimes. Comparative or multi-country studies are needed to validate and refine the conclusions.