The influence of robot applications on rural labor transfer

Against the backdrop of rapid advances in AI, big data, cloud computing, and IoT, robotics has diffused across many industries, reshaping production and potentially altering employment dynamics. China’s aging population and declining fertility are eroding the demographic dividend, increasing labor costs and accelerating robot adoption. Rural migrant workers, a large and crucial segment of China’s labor force, are concentrated in routine, low-skill tasks and face heightened employment instability as automation rises. While technology can both substitute for and complement labor, the net impact on spatial labor mobility—particularly rural migration—remains unclear. This study investigates whether and how robot applications influence the migration decisions of rural labor across regions, and unpacks the mechanisms—expansion versus substitution effects—that drive these outcomes. It adopts a micro-level individual perspective and differentiates between active and passive crowding-out channels to provide nuanced evidence and policy-relevant insights. The study posits that: (H1) robot applications increase the migration likelihood among the rural mobile population; and (H2) the impact is particularly significant among lower-skilled workers.

China’s dual economic and social structure creates distinctive urban-rural stratification, reflected in three strands of literature: (1) factor exchange inequality disadvantaging rural areas; (2) imbalanced allocation of public resources; and (3) impediments to urban-rural circulation. Broader rural issues studied include employment, land rights, poverty reduction, income gaps, and resource misallocation. Labor mobility research highlights technology, policy, migration costs, and urban characteristics: digitalization can optimize labor allocation; policy reforms (e.g., hukou) facilitate mobility; social networks reduce job search costs; infrastructure and environmental quality shape migration choices; and disasters affect mobility. The literature on robots and labor presents mixed evidence: substitution effects can reduce employment and wages, especially for routine and low-skill jobs; expansion effects can raise productivity, create new tasks, and increase labor demand; other studies find limited net employment effects but sectoral reallocation (e.g., from manufacturing to services). Skill heterogeneity is central: robot adoption tends to hurt low- and medium-skilled workers while boosting demand and wages for high-skilled labor. However, few studies directly examine how robot applications affect cross-regional labor mobility, especially for China’s rural migrants. This paper fills that gap.

Theory: The study adapts a task-based framework (Acemoglu & Restrepo, 2022), with a Cobb-Douglas structure over a continuum of tasks and three inputs: robots, low-skill labor, and high-skill labor. An automation frontier l* delineates tasks performed by robots versus labor. Under cost minimization, robot adoption reallocates tasks from low-skill labor to robots, affecting demand for each input. The framework shows two channels: (i) substitution—robots replace low-skill tasks, reducing their employment; (ii) expansion—productivity gains lower prices and expand output, increasing labor demand elsewhere. The net employment effect is theoretically ambiguous. Labor mobility is modeled by extending a spatial choice framework with worker skill draws (Fréchet) and iceberg migration costs, yielding migration probabilities that depend on city attributes, skill matching, and costs. Increased robot use raises skill requirements, potentially disadvantaging low-skilled workers and altering migration incentives.

Empirical design: Main specification is a heteroscedasticity-robust Linear Probability Model estimating the effect of city-level robot density on individual re-migration intention of rural migrants: Mobility_icht = β0 + β1 Robot_ct + X_i + μ_c + λ_t + δ_ht + ε_icht, where Mobility_icht equals 1 if the respondent does not plan to reside locally for more than five years (indicating intent to move), and 0 otherwise. X_i includes individual (age, age^2, gender, ethnicity, education), family (marital status, family size, accompanying family), and economic/locational controls (local health record, monthly rent, minimum wage, migration duration, city GDP, secondary industry employment, urban human capital). Fixed effects: city, year, and industry×year; standard errors clustered at the city level (and at broader levels for robustness).

Robot density: Constructed via a Bartik/shift-share approach combining IFR industry-level robot stocks with baseline (circa 2008) city-industry employment shares aligned to ISIC Rev.4 and the 2002 China industry classification. The city-year robot density aggregates exposure across industries using fixed pre-period employment weights. Alternative constructions use customs import data (2014–2015) for robot HS codes (84795010, 84795090, 84864031) aggregated to the city level and a variant formula with different labor weighting.

Data: Individual-level CMDS (2014–2018) sampling migrants age ≥15 without local hukou, covering all 31 provinces and XPCC. The main analysis focuses on rural hukou individuals aged 17–64 who moved for work or business; excludes special regions/cross-border moves; trims top 1% monthly income. City-level controls come from official statistical yearbooks and bulletins; minimum wages from provincial/municipal announcements. Robot exposure data: IFR robot stocks by industry; Second National Economic Census for baseline employment structure. Robustness and auxiliary analyses use 2010 Census and 2015 1% sample for migration rates, CFPS (2014/2016/2018), and the 2018 input-output table.

Identification and robustness: Addresses endogeneity (reverse causality, omitted variables) via multiple instruments and checks: (i) Bartik IV using robot trends in major robot-producing countries (e.g., Germany, South Korea, US, Japan, Sweden, UK); (ii) instruments based on historical post/telecom (1984 landlines per 100 persons × global mobile connections at t−1) and historical treaty-port status × global mobile connections at t−1; (iii) alternative clustering at broader (provincial) levels; (iv) alternative models (Probit/Logit), exclusion of “not well considered” responses, and alternative dependent and explanatory variable constructions; (v) policy timing control for the 2014 hukou reform; (vi) customs-based robot import values/quantities as core variables; (vii) job-type transitions (agriculture vs non-agriculture) as outcomes. A DEA–Malmquist index evaluates macro-level factor allocation efficiency using inputs (water/electricity use, capital stock via perpetual inventory, period-end employment) and outputs (real GDP, invention counts) for cities (2014–2018).

• Baseline effect: A 1% increase in urban robot density raises the probability of rural labor re-migration by about 0.249% (LPM with full controls and FE). Results are robust across specifications and samples; alternative estimates range around 0.239–0.280%.
• Urban migration rate: Robot diffusion is associated with a decrease in urban migrant labor migration rates at the prefecture level, robust to controls.
• Direction of movement: Each 1% increase in urban robot density decreases the probability of rural labor moving into cities by 0.109%.
• Return to agriculture: Robot adoption increases the probability of rural workers shifting toward agricultural work/returning to agriculture (ordered Probit coefficient 0.649; LPM 0.485 on job type category transitions).
• Active vs passive crowding-out: Robot density increases migration via both channels, with stronger passive effects (active: +0.228%; passive: +0.267% per 1% density).
• Wages and hours: A 1% increase in robot density associates with a 0.086% increase in rural labor wages and a 0.093% increase in weekly working hours, indicating expansion effects dominating substitution at the current stage.
• Heterogeneity by age: Active extrusion effects are significant for younger (17–30: 0.285**) and middle-aged (31–43: 0.178***) workers; passive extrusion is strongest for older workers (44+: 0.319***), consistent with skill mismatches as tasks upgrade.
• Heterogeneity by education/skill: Overall, low-skilled workers exhibit stronger migration responses to robot exposure, particularly for passive extrusion; for active extrusion, high-skilled workers show larger effects, reflecting new high-tech job creation.
• Heterogeneity by occupation: Effects are more pronounced in low-skilled manufacturing jobs (passive: low-skill 0.284*** vs high-skill 0.223***; active: low-skill 0.200*** vs high-skill 0.393***).
• Mobility scope and family accompaniment: Without accompanying family members, the effect is larger (0.316***) than with family (0.117***). Strongly mobile (cross-province) workers show higher sensitivity (0.250***) than weakly mobile (insignificant 0.053).
• Regional development: Effects are strongest in developed cities (0.317***), moderate in generally developed cities (0.163**), and insignificant in less developed cities (0.017), reflecting faster technological upgrading and higher skill thresholds.
• Gender and family status: Effects are significant for both men (0.259***) and women (0.239***). Among women, married status dampens significance overall; married women with children (0.212***) and without children (0.328***) show differing magnitudes, indicating family responsibilities and training access shape responses.
• Factor allocation efficiency: Robot density significantly improves city-level technical efficiency (coefficients around 0.144–0.185), suggesting enhanced scale and productivity and more efficient capital allocation.
• Robustness: Results hold under multiple IV strategies (major robot-producing countries; historical telecom/treaty port instruments), alternative clustering, estimators (Probit/Logit), alternative variables (customs import values/quantities; job-type outcomes), and after accounting for the 2014 hukou reform.

Findings confirm that robot applications materially alter rural labor mobility by simultaneously exerting substitution and expansion effects. The substitution of routine, low-skill tasks raises displacement risks—manifested in increased re-migration intentions and reduced city move-ins—while expansion effects raise wages and hours, reflecting productivity-driven output growth. The stronger passive crowding-out among low-skilled, older workers, and those in low-skill occupations indicates skill mismatches as task content upgrades. Heterogeneous impacts by mobility, family accompaniment, and regional development underscore how migration costs, household constraints, and local technological trajectories mediate outcomes. The increased likelihood of returning to agriculture suggests sectoral reallocation when urban skill thresholds rise. At the macro level, improvements in factor allocation efficiency point to productivity gains from robot integration. Overall, the evidence supports the hypotheses that robots increase rural migration propensity, particularly among lower-skilled groups, and highlights the need for policies that ease mobility frictions and support skill upgrading to ensure inclusive adjustment.

This paper integrates a task-based framework with a spatial labor mobility model and microdata to quantify how robot applications affect rural labor migration in China. It shows that urban robot density significantly increases re-migration intentions, reduces urban migration rates, raises the likelihood of returning to agriculture, and differentially impacts groups by skill, age, occupation, mobility, family status, and regional development. Despite substitution risks, robot adoption currently coincides with higher wages, longer hours, and improved factor allocation efficiency, indicating strong expansion effects. Policy implications include: (1) reforming settlement systems to decouple hukou from welfare provision and to expand cross-regional access to public services, mitigating mobility frictions, especially for family-accompanied migrants; (2) investing in targeted vocational training and continuous skill development for low-skilled and vulnerable groups, alongside transition support (e.g., unemployment insurance, job placement) to cushion displacement; and (3) promoting strategic robot and AI deployment to harness productivity gains while coordinating labor market information and inter-regional matches to mitigate adverse distributional impacts. Future research could exploit longer panels and additional micro-mechanisms (e.g., firm-worker matching, training interventions) to assess dynamic adjustment paths and policy efficacy.

The study period and several analyses are restricted to 2014–2018 for comparability. Robot exposure is constructed via shift-share measures based on IFR industry stocks and baseline employment shares, which, despite extensive robustness and IV strategies, may still entail measurement error. The main focus is on the secondary sector and rural hukou migrants, limiting generalizability to other sectors or populations. Some auxiliary datasets (e.g., CMDS, CFPS) are subject to data-use agreements and are not publicly accessible; variable definitions (e.g., intentions-based mobility) may differ from realized moves. Potential reverse causality and omitted variables are addressed with multiple instruments and controls, but residual endogeneity cannot be fully ruled out.