Heterogeneous human capital, spatial spillovers and regional innovation: evidence from the Yangtze River Economic Belt, China

Background: Human capital catalyzes technological innovation and urban innovation capacity. In China, skilled populations agglomerate in major urban regions (e.g., Beijing–Tianjin–Hebei, Yangtze River Delta), enabling knowledge and talent spillovers that raise regional innovation and can narrow interregional gaps. Thus, human capital can directly promote local innovation and, via spatial spillovers, raise neighboring regions’ innovation. Study area: The Yangtze River Economic Belt (YREB) spans 11 provincial units and 108 cities across China’s east, central, and west, accounting for 46.6% of national GDP in 2020 and hosting substantial professional and skilled talent pools. The YREB exhibits uneven development—highly innovative downstream coastal cities (e.g., Shanghai, Nanjing), relatively balanced midstream cities, and lagging upstream areas—making it a suitable context to study heterogeneous human capital and spatial spillovers. Objective: Quantify and analyze the interactions between heterogeneous human capital (low, intermediate, high education levels) and regional innovation, including spatial spillover effects, using a Spatial Durbin Model (SDM) with two-way fixed effects. The study also assesses spatial heterogeneity across upstream, midstream, and downstream YREB cities and proposes policy directions to enhance innovation and balanced development.

Endogenous growth theory posits human capital promotes growth via innovation and knowledge spillovers. Human capital—skills, knowledge, experience—exhibits strong spillovers that can yield increasing returns and local economic growth. At the city level, human capital influences the magnitude of knowledge spillovers and innovation output. Heterogeneous human capital and regional innovation: Empirical studies show positive relationships between human capital and innovation, reinforced by communication and collaboration. Improvements in human capital facilitate technology adoption and strengthen regional innovation capacity, explaining regional disparities in innovation and development. Given differing development stages, effects vary: in developing regions, raising low levels of human capital can notably improve innovation. This study tailors heterogeneity to China’s context by classifying human capital into low (primary), intermediate (secondary), and high (undergraduate and above) levels, reflecting China’s large share of secondary education. Human capital and spatial spillovers of innovation: Cross-regional human capital flows foster knowledge spillovers and technology sharing, enhancing local skills and innovation capacity, which in turn generate spatial spillovers. These spillovers differ by regional development level, with developed regions often more dependent on them. Prior work frequently focuses on macro (province/country) or micro (firm) scales; city-level heterogeneity remains underexplored. Methodologically, spatial econometric models (SLM, SEM, SDM) and spatial stochastic frontier approaches are common; SDM is typically the starting point, with degeneration tests to SLM/SEM where appropriate. To address heterogeneity of spillovers, geographically weighted regression (GWR) can reveal spatially varying relationships. This study contributes by (1) employing a three-tier human capital structure suited to Chinese cities and (2) focusing on the city scale across diverse YREB contexts.

Theoretical framework: Human capital affects local and neighboring cities’ innovation via information exchange, talent flows, and technology sharing. Local governments influence human capital through policy incentives, industrial agglomeration, and talent attraction, potentially creating positive or negative cycles depending on marginal effects and regional differences. Variables: Dependent variable is urban innovation capability (INNO), measured as per capita patent authorizations: lnINNO = ln(Pat/Pop). Patents include utility model and design patents to avoid long application lags; invention patents are fewer and excluded from timing considerations. Independent variables capture heterogeneous human capital as per capita counts of students by level: low (LEDU: primary), intermediate (MEDU: secondary), and high (HEDU: university). Controls include foreign direct investment intensity (FDI: FDI/GDP), science and education fiscal expenditure share (SFE), industrial structure (STRU: tertiary GDP/secondary GDP), and population agglomeration (POP: residents per area). All variables are log-transformed. Baseline models (Cobb–Douglas inspired): (1) lnINNO = α + α1 lnLEDU + α2 lnFDI + α3 lnSFE + α4 lnSTRU + α5 lnPOP (2) lnINNO = α + α1 lnMEDU + α2 lnFDI + α3 lnSFE + α4 lnSTRU + α5 lnPOP (3) lnINNO = α + α1 lnHEDU + α2 lnFDI + α3 lnSFE + α4 lnSTRU + α5 lnPOP Spatial modeling: Given expected spillovers, spatial dependence is tested and modeled. - Spatial autocorrelation: Global Moran’s I for lnINNO (2011–2020) using a geographic distance weight matrix shows significant positive spatial dependence each year (e.g., I=0.190 in 2011; I≈0.153–0.205 across years, all p=0.000), justifying spatial models. - Model selection: LM tests indicate both lag and error dependence are significant. Wald and LR tests support the Spatial Durbin Model (SDM) over SLM/SEM. A two-way fixed-effects SDM is adopted: lnINNO_it = α + ρ W lnINNO_it + Σ α_k X_kit + Σ θ_k W X_kit + ε_it, with X including each education level (in separate models 4–6) and controls. Effects are decomposed into direct (local) and indirect (spatial spillover) effects. - Spatial heterogeneity: GWR augments analysis by allowing coefficients to vary across space. An OLS specification with all three education levels (model 7) is compared to GWR (model 8) via ANOVA (log-likelihood, AICc, R²) to justify spatially varying relationships. Data: Panel of 108 YREB cities, 2011–2020, from China Urban Statistical Yearbook (2012–2020), China Statistical Yearbook (2012–2021), and 2021 provincial/city yearbooks. Missing values are imputed using growth rates and benchmarks. All variables are log-transformed to remove scale effects. Descriptive statistics show substantial spatial-temporal variation in innovation and human capital distributions. Robustness: The geographic distance matrix is replaced with an economic distance matrix in SDM estimations to test robustness of spatial spillovers and directionality of effects.

- Spatiotemporal pattern: Innovation capacity in the YREB exhibits clear spatial clustering and diffusion. From 2011 to 2020, patent authorizations increased broadly, radiating westward. Innovation intensity decreases from downstream (coastal) to upstream (inland), with downstream hubs (Shanghai, Ningbo, Nantong) as cores; midstream shows later balanced gains; upstream led by the Chengdu–Chongqing zone. - Spatial dependence: Moran’s I is significantly positive each year (e.g., 2011: I=0.190, p=0.000; 2012: 0.205; 2016: 0.170; 2020: 0.162; all p=0.000), indicating spatial autocorrelation in innovation. - SDM direct effects (geographic distance matrix): • Low-level human capital (lnLEDU): significantly negative direct effect on local innovation; a 1% increase in low-level human capital associates with a 5.44% decrease in patent authorizations (direct effect coefficient −0.544, p<0.01). • Intermediate human capital (lnMEDU): significantly positive direct effect; a 1% increase associates with a 3.56% increase in innovation (0.356, p<0.01). • High-level human capital (lnHEDU): significantly positive direct effect; a 1% increase associates with a 1.36% increase in innovation (0.136, p<0.05). - SDM indirect (spillover) effects (geographic distance matrix): • Low-level human capital: positive spillover (1.367, p<0.05), opposite to its local effect, indicating basic education expansion locally may benefit neighbors’ innovation. • Intermediate human capital: negative spillover (−5.625, p<0.01), suggesting cities upgrading secondary education attract talent from neighbors, suppressing their innovation. • High-level human capital: negative spillover (−0.473, p<0.05), consistent with siphoning of highly educated talent from nearby cities. - Controls (key patterns under SDM): • FDI: local direct effects often insignificant or mildly negative; spillovers tend to be negative, consistent with crowding-out of local R&D (e.g., indirect effects −0.517 to −0.525, p<0.01 in some models). • Science and education expenditure (SFE): direct effects weakly negative or insignificant; spillovers can be inhibitory, possibly due to low efficiency/high risk in fiscal input deployment. • Industrial structure (STRU): direct effects generally insignificant; negative spillovers indicate service-sector development may induce brain drain from neighbors. • Population agglomeration (POP): significant negative direct effects (e.g., −0.818 to −0.460) and mixed spillovers, implying congestion and resource constraints locally and talent loss in neighbors. - Spatial heterogeneity (GWR vs OLS): Across 2011–2020, GWR improves fit over OLS in every year (e.g., 2011 R²: OLS 0.733 vs GWR 0.827; AICc difference >3; similar improvements each year), supporting spatially varying relationships. Spatial evolution of coefficients: low-level human capital effects turn from negative to positive over time and spread from downstream to upstream; intermediate human capital effects decline over time and turn negative in upstream; high-level human capital shows polarization by 2020, with reduced downstream impact and some gains upstream. - Robustness (economic distance matrix): Core conclusions persist. Direct effects: low-level human capital negative (−0.696, p<0.01), intermediate positive (0.451, p<0.01), high-level positive (0.247, p<0.01). Indirect effects for intermediate and high human capital remain negative (−1.737 and −0.188, both significant). Some deviations (e.g., low-level human capital indirect effect becomes negative), but overall spillover patterns and heterogeneity are robust to alternative spatial weighting. - Regional nuance: In downstream cities (e.g., Shanghai, Nanjing), human capital’s marginal impact on innovation is not significant, reflecting saturation or siphon effects; medium and high human capital matter more for midstream and upstream cities.

The study addresses how heterogeneous human capital affects urban innovation locally and through spatial spillovers within the YREB. Findings demonstrate that human capital’s impact depends on education level and location. Locally, intermediate and high-level human capital promote innovation, while low-level human capital suppresses it—consistent with the greater innovation conversion capacity of more educated workforces. Spillovers invert these patterns: expansion of secondary and tertiary education in one city can siphon talent from neighbors, reducing their innovation; basic education expansion locally can benefit neighbors’ innovation via broader skill diffusion. These results highlight the importance of aligning human capital development with regional industrial structures and coordinating across neighboring jurisdictions to mitigate negative spillovers. The observed downstream-to-upstream diffusion pattern underscores the role of leading hubs in seeding innovation, while congestion and crowding-out dynamics (FDI, population agglomeration) can hinder both local and neighboring innovation if unmanaged. The GWR results confirm significant spatial heterogeneity, implying that one-size-fits-all human capital policies are suboptimal; instead, cities should tailor strategies to their position (downstream, midstream, upstream), industrial base, and capacity to absorb talent. Policy relevance includes: fostering a human-capital-friendly ecosystem, leveraging downstream hubs’ radiation effects to uplift midstream/upstream regions, and reforming higher education downstream while strengthening basic education upstream and midstream. Together, these measures can enhance regional innovation while balancing development.

Using a two-way fixed-effects SDM on 108 YREB cities (2011–2020) and complementary GWR analysis, the study shows: (1) innovation exhibits significant spatial heterogeneity and diffusion from downstream to upstream; (2) locally, low-level human capital inhibits innovation, whereas intermediate and high-level human capital promote it; spillovers display opposite signs, with intermediate and high-level human capital exerting negative spillovers on neighbors; (3) human capital’s innovation impact is insignificant in downstream cities but more pronounced in midstream/upstream. The study emphasizes targeted human capital strategies by region and level, coordination to mitigate negative spillovers, and differentiated investments in basic versus higher education. Future research should adopt more comprehensive human capital metrics (e.g., average years of schooling), expand to other economic belts, and strengthen causal identification.

Human capital is proxied by per capita numbers of students at different education levels, capturing structural distribution but not the full stock/quality of human capital (e.g., average years of schooling, on-the-job skills). Data compiled from multiple public sources may have inconsistent standards and introduce uncertainties. The design is observational with potential endogeneity, limiting causal inference. Future work should employ richer human capital measures, apply the methodology to other regions for comparison, and improve identification strategies.