The Impact of Digital Infrastructure on the Urban–Rural Income Gap: Empirical Evidence from 285 Cities in China

Abstract

Digitalization has reshaped economic systems worldwide, yet its distributional consequences remain uneven and raise new challenges for sustainable development. China, where digital infrastructure has expanded rapidly, provides a critical setting to examine these effects and their implications for sustainable and inclusive growth. Using a balanced panel of 285 prefecture-level cities from 2007 to 2023, this study constructs a text-based index of digital infrastructure from government work reports and applies two-way fixed effects, instrumental variables, nonlinear models, placebo tests, heterogeneity analysis, and spatial Durbin models. The results show that digital infrastructure significantly widens the urban–rural income gap, with the effect becoming increasingly convex as digital development deepens. Two mechanisms drive this pattern: the concentration of innovation resources in urban areas, which crowds out rural R&D, and a modest degree of wage-structure polarization. Spatial spillovers also matter; digital development in neighboring cities partially offsets local inequality by enhancing interregional connectivity and knowledge diffusion. These findings provide city-level causal evidence on the unequal distributional impacts of digitalization in large emerging economies and highlight the need for sustainability-oriented digital governance, inclusive innovation systems, and regionally coordinated strategies to prevent digital infrastructure from reinforcing structural disparities. Strengthening these policies is essential for achieving more sustainable urban–rural integration in the digital era.

1. Introduction

Over the past two decades, China and other large emerging economies such as India and Brazil have experienced an unprecedented wave of digital transformation. Massive investments in broadband networks, data centers, cloud platforms, and mobile communication systems have reshaped industrial structures, spatial linkages, and modes of service delivery [1,2,3]. Digital infrastructure is often expected to narrow information gaps and improve access to economic opportunities. Yet, growing empirical evidence suggests that digitalization may also amplify pre-existing inequalities by disproportionately benefiting urban, capital-intensive, and high-skilled sectors [4,5]. This ambiguity raises an important policy question for emerging economies: Does digital infrastructure function as an equalizer or a divider?

Despite rising scholarly interest, the existing literature remains inadequate in several respects. First, most studies rely on national or provincial data, which obscure the substantial city-level spatial heterogeneity through which digitalization affects income inequality [6,7]. Second, prior work typically focuses on a single channel, such as innovation upgrading or labor-market polarization, without integrating multiple mechanisms into a unified empirical framework [8]. Third, although scholars increasingly acknowledge spatial spillovers, few examine whether the digitalization of neighboring cities can offset local inequality, a possibility highly relevant for regionally interconnected economies [9]. Finally, conventional measures of digital infrastructure emphasize coverage or output indicators, but seldom capture the policy-driven, infrastructure-specific dynamics that shape the timing, intensity, and strategic priorities of digital deployment across cities [10].

To motivate our empirical investigation, Figure 1 illustrates the long-run evolution of China’s urban–rural income inequality and digital infrastructure development between 2007 and 2023. The Theil index steadily declined from 0.118 in 2007 to 0.045 in 2023, whereas the digital infrastructure index, constructed from more than 5000 official local government work reports, rose sharply, particularly following major national initiatives such as the Broadband China Strategy (2013) and the “Internet+” Action Plan (2015). This mirror-shaped pattern suggests that, despite an observable narrowing in headline income ratios, the rapid expansion of digital infrastructure may be reshaping more profound structural inequalities. Whether digitalization mitigates or reinforces these underlying disparities at the city level remains empirically unresolved.

To address these gaps, this study investigates the distributional consequences of digital infrastructure using a balanced panel of 285 prefecture-level cities from 2007 to 2023. We construct a novel text-based index of digital infrastructure from government work reports and employ fixed-effects models, nonlinear specifications, spatial dependence diagnostics, and Spatial Durbin Models (SDMs) to identify both direct and interregional effects. This empirical strategy enables a comprehensive examination of how digitalization affects inequality within cities and through spatial spillovers across cities. Based on these aims, the study addresses four core research questions:

RQ1: Does digital infrastructure significantly widen the Urban–rural income gap at the city level?

RQ2: Do agricultural-innovation crowding-out and wage-structure polarization jointly mediate this effect?

RQ3:Can neighboring cities’ digitalization generate spatial spillovers that offset local inequality?

RQ4: Under what institutional and market conditions does digitalization’s inequality effect weaken or reverse?

This study makes four key contributions. First, it provides the first long-horizon, city-level evidence on the inequality consequences of digital infrastructure in a large emerging economy. Second, it jointly assesses multiple mechanisms, innovation crowding-out and labor-structure polarization, within a unified identification framework. Third, it uncovers spatial offsetting effects. The paper proceeds as follows: Section 2 reviews related literature and develops hypotheses; Section 3 creates the theoretical framework and hypotheses; Section 4 details data and variables; Section 5 shows baseline, nonlinear, and mechanism results; Section 6 tests mechanisms; Section 7 covers spatial results; Section 8 and Section 9 conclude with policy implications for inclusive and regional digital development.

2. Literature Review

Urban–rural income inequality has long been a defining structural challenge in large emerging economies. Existing studies show that disparities in human capital, industrial structure, and access to public services jointly shape income divergence between cities and rural areas, and that these gaps tend to widen during periods of rapid transformation [11,12]. With the rapid expansion of digital technologies, scholars have increasingly examined whether digital infrastructure reduces or amplifies such persistent inequalities. A systematic review shows that the literature falls into three broad categories: (1) digitalization and distributional outcomes, (2) digitalization-driven mechanisms affecting inequality, and (3) spatial spillovers in digital development.

2.1. Digitalization and Distributional Outcomes

Digital infrastructure, including broadband networks, mobile communications, and data platforms, is broadly acknowledged as a catalyst for productivity, innovation, and equitable access to information [13,14]. However, growing empirical evidence indicates that the allocation of digital benefits is uneven across regions. Recent work highlights that technology adoption is shaped by organizational culture, institutional quality, and local economic development, implying that digitally advanced cities disproportionately capture the gains from digital upgrading [15]. These findings are consistent with studies emphasizing that digitalization can reinforce pre-existing socioeconomic disparities, particularly the urban–rural divide [16].

Theoretical research also offers contrasting insights. Endogenous growth theory posits that digital technologies enhance knowledge accumulation and innovation [17], while new economic geography suggests that such technologies may intensify spatial disparities by concentrating economic activity and high-skilled labor in urban centers [18,19]. Classic perspectives, such as Kuznets’ inverted-U hypothesis [20], imply nonlinearities in development processes [20]. In the digital era, these nonlinearities may be amplified: regions with limited digital access may fall further behind, while digitally advanced regions may experience accelerated returns [21,22]. Overall, the literature indicates that without supportive institutions and inclusive digital governance, digital infrastructure may widen spatial inequality.

2.2. Mechanisms Linking Digitalization to Income Inequality

Beyond documenting its aggregate distributional effects, a second strand of literature examines how digitalization reshapes economic structures and labor markets, thereby influencing income inequality. Two prominent mechanisms are frequently discussed.

2.2.1. Innovation Concentration and Crowding-Out Effects

Schumpeterian growth theory highlights the cumulative nature of technological progress: digitally advanced cities attract skilled labor, R&D investment, and high-tech industries, strengthening their innovation ecosystems [23]. At the same time, rural regions often lack the human capital and institutional capacity needed to integrate digital technologies effectively. Structural transformation theories suggest that digitalization accelerates the shift of economic resources from agriculture and traditional sectors toward urban and high-tech industries [24]. Empirical studies confirm that expansion of digital infrastructure increases urban patenting activity, while innovation in less developed rural regions stagnates or declines [25]. This spatial reallocation of innovation resources may weaken rural productivity growth, reinforcing the income gap.

2.2.2. Wage Structure Polarization

The second mechanism relates to labor-market dynamics driven by skill-biased technological change (SBTC). Digital technologies support high-skilled work but replace routine, low-skilled tasks [26]. As digital infrastructure advances, the returns to digital literacy and cognitive skills increase, particularly in urban sectors such as ICT, finance, and advanced manufacturing [27]. Low-skilled rural workers, by contrast, face displacement risks, limited digital training opportunities, and restricted access to urban labor markets due to institutional constraints [28]. The resulting wage polarization widens income differentials between skill groups and between urban and rural regions.

Together, these studies demonstrate that digitalization can generate multi-layered structural pressures that reinforce pre-existing inequalities.

2.3. Spatial Spillovers in Digital Development

A third strand of literature emphasizes that digital development is intrinsically spatial: technological capabilities, labor mobility, and knowledge diffusion extend beyond administrative boundaries. Classical regional economics highlights the role of geographical proximity, industrial clustering, and shared supply chains in generating cross-regional externalities [29]. In the digital era, these spillover channels expand through online service platforms, interoperable digital systems, and regional data networks [30]. Empirical analyses increasingly confirm that digital infrastructure exhibits strong inter-city externalities, reducing regional disparities when neighboring areas undergo parallel digital upgrading [31].

In the context of urban–rural inequality, such spillovers can either mitigate or reinforce the inequality-enhancing effects of local digitalization. As surrounding cities improve digital accessibility, rural regions may benefit from shared digital services, integrated labor markets, and the diffusion of technology. Conversely, fragmented regional development can limit these benefits. Understanding the spatial dimension of digital infrastructure is therefore crucial for accurately assessing its distributional consequences.

Despite substantial progress, three significant gaps remain in the existing literature:

(1)

Limited city-level empirical evidence. Most studies rely on provincial or national data, leaving insufficient micro-spatial evidence on how digital infrastructure affects urban–rural inequality within cities.

(2)

Lack of an integrated multi-mechanism framework. Prior studies examine isolated channels (e.g., digital divide or SBTC) but do not integrate innovation dynamics, labor-market polarization, and spatial spillovers into a unified analytical framework.

(3)

Insufficient attention to spatial interdependencies. The literature increasingly acknowledges cross-city spillovers, yet few studies incorporate spatial econometric modeling to identify how digital development in neighboring regions influences local inequality.

3. Theoretical Analysis

Building on the gaps identified earlier, this section explores the theoretical mechanisms through which digital infrastructure influences the urban–rural income gap. While the literature offers fragmented insights, a unified theoretical framework remains lacking. Accordingly, this section synthesizes the core mechanisms into three pathways: direct productivity effects, indirect structural effects, and spatial spillovers, and derives the corresponding testable hypotheses.

3.1. Direct Impact: The Role of Digital Infrastructure in Urban–Rural Inequality

The direct impact of digital infrastructure on income inequality arises primarily from its differential effects on regional productivity. Digital infrastructure, including broadband internet, mobile networks, and data centers, enhances economic activities by facilitating access to information, markets, and services. However, its impact is far from uniform. Urban regions, which already possess better access to human capital, financial markets, and institutional capacity, tend to capture the majority of these productivity benefits.

Digital infrastructure acts as a complementary factor in urban economies where high-skilled labor and technological capabilities are abundant. This increases the return on digital investment, driving economic growth in urban centers [32]. Conversely, rural areas, characterized by lower levels of digital access and fewer resources, are less able to capitalize on these advancements fully [15]. Thus, the introduction of digital infrastructure may exacerbate spatial inequalities by disproportionately benefiting urban economies, leaving rural regions behind. Moreover, the productivity effects of digital infrastructure may be nonlinear. In the early stages, digital expansion may generate modest gains in both the urban and rural sectors. As digital development advances, however, its complementarities with skilled labor, innovation capacity, and institutional readiness grow stronger, indicating a potentially convex relationship between digitalization and inequality. This possibility justifies the nonlinear model tested in the empirical analysis.

Given the diverse economic structures across urban and rural regions, the direct influence of digital infrastructure is likely to widen the income gap between them. This mechanism is consistent with arguments in new economic geography, where technological advancements tend to concentrate economic activity in areas with stronger digital and human capital foundations, thereby amplifying pre-existing disparities. Accordingly, this research proposes:

H1. 

Digital infrastructure widens the urban–rural income gap by disproportionately benefiting urban areas, which have greater access to complementary resources, such as skilled labor and capital.

3.2. Indirect Mechanisms: Innovation Crowding-Out and Wage Polarization

In addition to the direct productivity effects, digital infrastructure influences the urban–rural income gap through indirect mechanisms that shape the allocation of innovation resources and the wage structure within labor markets.

3.2.1. Innovation Crowding-Out Effect

One important indirect mechanism is the crowding-out effect in innovation. Digital infrastructure enhances innovation by lowering the cost of information exchange, research collaboration, and technological development. However, these benefits are not equally distributed across regions. Urban areas, with better access to skilled labor, R&D facilities, and innovation networks, are better equipped to absorb and capitalize on these advancements [33]. As a result, digitalization may concentrate innovation activities in urban centers, leaving rural areas with limited capacity to generate or absorb new technologies.

This dynamic aligns with Schumpeterian growth theory, which argues that innovation is path-dependent and cumulative. As cities accumulate digital infrastructure, they attract additional innovation capital, skilled labor, and high-tech industries [34]. In contrast, rural areas, lacking these resources, fall behind, unable to fully integrate digital technologies into their local economies. This redistribution of innovation increases the urban–rural income gap by slowing technological progress in rural regions, particularly in agriculture and traditional industries. Furthermore, this research proposes the following:

H2. 

Digital infrastructure widens the urban–rural income gap by directing innovation efforts toward cities and pushing agricultural R&D and innovation out of rural areas.

3.2.2. Wage Structure Polarization Effect

Another significant indirect mechanism is wage polarization driven by skill-biased technological change (SBTC) [35]. As digital infrastructure expands, demand for high-skilled labor increases, especially in industries such as ICT, finance, and advanced manufacturing. At the same time, routine and low-skilled jobs in agriculture and traditional sectors face increasing risks of automation or outsourcing. This technological shift places upward pressure on wages for high-skilled urban workers while simultaneously suppressing wages for low-skilled rural workers [36].

The polarizing effect is particularly pronounced in urban areas, where high-skilled workers benefit from increased demand for their services. Rural areas, however, experience wage stagnation or decline as their economies remain heavily reliant on low-skill sectors with limited opportunities for digital adaptation. This structural transformation thus amplifies wage disparities between urban and rural regions. Furthermore, rural workers are constrained by institutional factors, including limited access to education, healthcare, and digital skills training. This worsens the wage gap because rural workers cannot acquire the digital skills necessary to meet the increasing demand for high-skilled labor in urban areas’ economies. So, this research proposes:

H3. 

Digital infrastructure widens the urban–rural income gap by increasing the wage premium for high-skilled labor in urban areas while depressing wages for low-skilled rural workers.

3.3. Spatial Spillover Effects: Regional Integration and Digital Externalities

A third critical mechanism through which digital infrastructure affects income inequality is spatial spillovers. Digital infrastructure, as an inherently networked technology, generates inter-regional externalities that can either mitigate or exacerbate the effects of local digital development [37]. In this context, neighboring cities or regions that simultaneously undergo digital upgrading can benefit from positive spillovers in several ways. First, digital infrastructure enhances the diffusion of knowledge, enabling rural or less developed regions to access digital services, data, and platforms developed in more advanced urban areas. These spillovers can expand access to education, financial services, healthcare, and other digital services, thereby reducing inequality by improving economic opportunities for rural populations [38]. Second, labor mobility is facilitated by digital connectivity, enabling rural workers to access job opportunities in urban centers [39]. This can help bridge the wage gap by giving rural workers access to higher-paying digital jobs in urban areas.

However, the magnitude of these positive spillovers depends on several factors, including geographic proximity, industrial complementarity, and regional coordination. In regions with uneven or poorly integrated digital infrastructure, the potential for positive spillovers is limited, and the benefits may remain concentrated in urban areas [40]. Thus, spatial spillovers can either attenuate or reinforce local inequalities depending on regional integration and policy coordination. When neighboring cities develop complementary digital ecosystems, the externalities generated can help reduce the income gap between urban and rural areas. Accordingly, this research proposes:

H4. 

Digital infrastructure development in neighboring cities generates spatial spillovers that reduce the income disparity by improving regional connectivity, labor mobility, and cross-regional knowledge diffusion.

The integrated framework in Figure 2 outlines three mechanisms, M1 (skill-based technological change), M2 (urban innovation concentration), and M3 (spatial digital spillovers), that explain how digital infrastructure affects the income gap.

M1 (Skill-based technological change) leads to routine job substitution and wage structure polarization, increasing inequality by raising wages for high-skilled urban workers while displacing low-skilled rural workers. M2 (Urban innovation concentration) exacerbates inequality by attracting innovation to urban areas, thereby reducing rural agricultural R&D and widening the income gap. M3 (Spatial digital spillovers) presents cross-regional benefits that can mitigate these effects, reducing the income gap when neighboring regions share digital resources and labor mobility. The balance of these mechanisms determines whether digital infrastructure increases or reduces urban–rural inequality.

4. Research Design

4.1. Data Sources and Model Settings

This study uses a panel of 285 Chinese prefecture-level cities. Covering the years 2007–2023. All variables are harmonized under consistent administrative boundaries as defined by the National Bureau of Statistics of China (NBS), as illustrated in

Table 1. Data Sources and Variable Coverage.

Source	Time Span	Main Content	Used for Variables
China City Statistical Yearbook (CCSY)	2008–2024	GDP, population, industrial structure, and fiscal expenditure	Economic & demographic controls
China Statistical Yearbook	2007–2023	Urban and rural disposable income per capita	Calculation of the Theil index (DV)
China Regional Statistical Yearbook	2007–2023	Supplementary city-level indicators, missing data cross-check	Data validation
Wind Database/CEIC	2007–2023	FDI inflow, government revenue & expenditure	Macroeconomic control variables
Local Government Work Reports	2007–2023	Text frequency of digital infrastructure keywords	Construction of Digital Infrastructure Index (IV)

Note: All variables are adjusted to 2015 constant prices.

. Cities with incomplete income data are linearly interpolated in accordance with NBS conventions. Four directly governed municipalities (Beijing, Shanghai, Tianjin, and Chongqing) are retained as standard prefecture-level units.

4.2. Variable Specification

(1)

Dependent Variable (DV): Urban–rural Income Inequality

Urban–rural income inequality, measured by the Theil index (${\mathrm{T}\mathrm{h}\mathrm{i}\mathrm{e}\mathrm{l}}_{\mathrm{i}\mathrm{t}}$), which allows for decomposition into within- and between-group inequality. This measure captures both the intensity and structure of local income distribution. We use inflation-adjusted disposable incomes of urban and rural households, weighted by population shares. A higher Theil index value indicates a broader income gap. The formula is as follows:

$${\mathrm{T}\mathrm{h}\mathrm{i}\mathrm{e}\mathrm{l}}_{\mathrm{i}\mathrm{t}}={\sum }_{\mathrm{j}=1}^2\left({\displaystyle \frac{{\mathrm{I}}_{\mathrm{i}\mathrm{j}\mathrm{t}}}{{\mathrm{I}}_{\mathrm{i}\mathrm{t}}}}\right)\mathrm{l}\mathrm{n}({\displaystyle \frac{{\mathrm{I}}_{\mathrm{i}\mathrm{j}\mathrm{t}}}{{\mathrm{I}}_{\mathrm{i}\mathrm{t}}}}/{\displaystyle \frac{{\mathrm{N}}_{\mathrm{i}\mathrm{j}\mathrm{t}}}{{\mathrm{N}}_{\mathrm{i}\mathrm{t}}}})$$

(1)

where j = 1, 2 represent urban and rural areas, respectively. ${\mathrm{I}}_{\mathrm{i}\mathrm{t}}$ is the total disposable income of city i in year t, and ${\mathrm{I}}_{\mathrm{i}\mathrm{j}\mathrm{t}}$ denotes the income for the urban (j = 1) and rural (j = 2) population within that city. ${\mathrm{N}}_{\mathrm{i}\mathrm{t}}$ is the total population, and ${\mathrm{N}}_{\mathrm{i}\mathrm{j}\mathrm{t}}$ corresponds to the sizes of the urban and rural populations.

(2)

Independent Variable (IV): Digital Infrastructure

The key explanatory variable (${\mathrm{D}\mathrm{i}\mathrm{g}\mathrm{i}\mathrm{t}}_{\mathrm{i}\mathrm{t}}$) represents the degree of digital infrastructure development. It is derived from local government work reports via text mining and is normalized as the logarithm of the keyword frequency per 10,000 words. The measure covers 51 keywords related to broadband networks, data centers, 5G deployment, and cloud computing (see Appendix A). This index captures city-level policy attention and the intensity of digitalization implementation.

(3)

Mediating Variables

To uncover the mechanisms linking digital infrastructure to income inequality, we incorporate two mediating pathways consistent with the theoretical framework:

Innovation Crowding-Out Channel: measured by the ratio of agricultural research and development expenditure to total R&D (AgriR&D). A decline in agrarian R&D intensity indicates that digitalization may divert innovation resources toward urban, high-tech sectors.

Wage Structure Polarization Channel: captured by the standard deviation of average wages across industries (WageGap). This variable reflects how digitalization amplifies skill-biased returns and wage dispersion between high- and low-skill sectors.

(4)

Control Variables

To isolate the net impact of digital infrastructure, we include control variables covering economic development, industrial structure, fiscal capacity, openness, demographic density, and educational attainment. These variables are standard in regional inequality studies and ensure robustness against omitted-variable bias. All controls are lagged by 1 year to mitigate simultaneity further.

These variables operationalize the theoretical framework empirically, with detailed definitions, units, and signs in

Table 2. Variable Definitions, Measurement, and Expected Effects.

Type	Variable	Symbol	Definition/Construction	Source
Dependent Variable (DV)	Urban–rural Income Inequality	Theil	The Thiel index is based on per capita disposable income for urban and rural residents.	China Statistical Yearbook
Independent Variable (IV)	Digital Infrastructure	Digit	log(1 + frequency of digital-infrastructure keywords per 10,000 words in local government work reports); 51 keywords covering six dimensions (ICT, digital economy, big data, information infrastructure, etc.); cumulative (stock-based) index	Local Government Work Reports
Mediators	Agricultural Innovation (Agri R&D Intensity)	AgriR&D	Ratio of agricultural R&D expenditure/total R&D expenditure	(CCSY)
	Wage Polarization	WageGap	Standard deviation of industry-level wages; larger values = higher wage dispersion	Wind Database
Controls	GDP per Capita	lnGDPpc	GDP per capita (log)	(CCSY)
	Industrial Structure	Indus	Share of secondary-industry value-added in GDP	(CCSY)
	Fiscal Expenditure	Fiscal	Government fiscal expenditure per capita (log)	Wind Database
	FDI Inflow	FDI	Actual foreign direct investment	CEIC Database
	Population Density	PopDens	Total population/land area	(CCSY)
	Education Level	Edu	Higher-education enrollment per 10,000 persons	NBS
	Innovation Intensity	Innov	Ratio of science and technology expenditure/total public budget expenditure	(CCSY)
	Financial Development	Finance	Year-end loan balance of financial institutions (log)	Wind Database
	Urbanization Rate	Urban	Urban population/total population	(CCSY)
	Consumption Level	Consump	Total retail sales of consumer goods/GDP	(CCSY)
	Telecommunications Scale	Comm	Total telecommunications business revenue (log)	Wind Database

Notes: All continuous variables are log-transformed; missing observations (<1%) are interpolated according to NBS guidelines.

, complementing the data source overview in Table 1.

4.3. Model Specification

We estimate a series of econometric models to test the causal relationship between digital infrastructure and Urban–rural income inequality while accounting for spatial interdependence and endogeneity.

(a) Baseline Fixed-Effects Model

To investigate how digital infrastructure affects the urban–rural income gap, this study develops the following baseline model:

$${\mathrm{T}\mathrm{h}\mathrm{i}\mathrm{e}\mathrm{l}}_{\mathrm{i}\mathrm{t}}={\mathsf{\beta }}_0+{\mathsf{\beta }}_1{\mathrm{D}\mathrm{i}\mathrm{g}\mathrm{i}\mathrm{t}}_{\mathrm{i}\mathrm{t}}+\mathsf{\gamma }{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}\mathrm{l}}_{\mathrm{i}\mathrm{t}}+{{\mathsf{\mu }}_{\mathrm{i}}+{\mathsf{\lambda }}_{\mathrm{i}}+\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(2)

where ${\mathrm{T}\mathrm{h}\mathrm{i}\mathrm{e}\mathrm{l}}_{\mathrm{i}\mathrm{t}}$ denotes the Urban–rural income disparity of city $\mathrm{i}$ in year $\mathrm{t}$, measured by the Thei index (${\mathrm{D}\mathrm{i}\mathrm{g}\mathrm{i}\mathrm{t}}_{\mathrm{i}\mathrm{t}})$ is the level of digital infrastructure development. ${\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}\mathrm{l}}_{\mathrm{i}\mathrm{t}}$ represents a vector of city-level control variables. ${\mathsf{\mu }}_{\mathrm{i}}$ and ${\mathsf{\lambda }}_{\mathrm{i}}$ are city and year fixed effects; and ${\epsilon }_{it}$ is a random error term. If the regression results show that ${\mathsf{\beta }}_1>0$ and statistically significant, it indicates that greater digital infrastructure intensity widens the Urban–rural income gap (supporting H1).

(b) Mediation Models

To analyze how digital infrastructure influences urban–rural income inequality, we adopt a unified two-stage mediation framework in which all mediators are denoted by ${\mathrm{M}}_{\mathrm{i}\mathrm{t}}$. In this study, ${\mathrm{M}}_{\mathrm{i}\mathrm{t}}$ represents either agricultural R&D intensity (the innovation-crowding-out channel) or wage polarization (the labor-market restructuring channel).

Step 1: From Digitalization to Mediators

$${\mathrm{M}}_{\mathrm{i}\mathrm{t}}={\mathsf{\alpha }}_0+{\mathsf{\alpha }}_1{\mathrm{d}\mathrm{i}\mathrm{g}\mathrm{i}\mathrm{t}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\alpha }}^{\prime }{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}\mathrm{l}}_{\mathrm{i}\mathrm{t}}+{{\mathsf{\mu }}_{\mathrm{i}}+{\mathsf{\lambda }}_{\mathrm{i}}+\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(3)

Step 2: From Mediators to Income

$${\mathrm{T}\mathrm{h}\mathrm{i}\mathrm{e}\mathrm{l}}_{\mathrm{i}\mathrm{t}}={\mathsf{\beta }}_0+{\mathsf{\beta }}_1{\mathrm{d}\mathrm{i}\mathrm{g}\mathrm{i}\mathrm{t}}_{\mathrm{i}\mathrm{t}}+\mathsf{\theta }{\mathrm{M}}_{\mathrm{i}\mathrm{t}}+\mathsf{\gamma }{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}\mathrm{l}}_{\mathrm{i}\mathrm{t}}+{{\mathsf{\mu }}_{\mathrm{i}}+{\mathsf{\lambda }}_{\mathrm{i}}+\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(4)

A significant ${\mathsf{\alpha }}_1$ and $\mathsf{\theta }$ provides evidence of the corresponding mediating pathway.

(c) Spatial Durbin Model (SDM)

To capture both local and inter-regional interactions, we extend the baseline specification to a Spatial Durbin Model (SDM) that includes spatial lags of both dependent and independent variables:

$${\mathrm{T}\mathrm{h}\mathrm{i}\mathrm{e}\mathrm{l}}_{\mathrm{i}\mathrm{t}}=\mathsf{\rho }+{\mathsf{\beta }}_1{\mathrm{d}\mathrm{i}\mathrm{g}\mathrm{i}\mathrm{t}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\theta }}_1\mathrm{W}{\mathrm{d}\mathrm{i}\mathrm{g}\mathrm{i}\mathrm{t}}_{\mathrm{i}\mathrm{t}}+\mathsf{\gamma }{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}\mathrm{l}}_{\mathrm{i}\mathrm{t}}+\mathsf{\eta }\mathrm{W}{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}\mathrm{l}}_{\mathrm{i}\mathrm{t}}+{{\mathsf{\mu }}_{\mathrm{i}}+{\mathsf{\lambda }}_{\mathrm{i}}+\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(5)

Here $\mathrm{W}$ is the spatial-weight matrix (constructed using contiguity or inverse-distance criteria). The parameter $\mathsf{\rho }$ measures the spatial autocorrelation of inequality, while ${\mathsf{\theta }}_1$ captures the indirect spatial effect of neighboring cities’ digitalization. A negative and significant ${\mathsf{\theta }}_1$ (${\mathsf{\theta }}_1<0$) indicates an offsetting spatial impact, consistent with H4. The inclusion of $\mathrm{W}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}{\mathrm{l}}_{\mathrm{i}\mathrm{t}}$ accounts for potential environmental spillovers associated with control variables.

5. Empirical Results

5.1. Descriptive Statistics and Multicollinearity Diagnostics

Table 3. Descriptive statistics.

Variable	N	Mean	SD	Min	Median	Max
Theil	4614	0.0805	0.0484	0.3181	0.0698	0.318
Digit	4614	0.0065	0.00707	0.0633	0.00370	0.0633
Indus	4614	43.14	15.54	85.08	45.34	85.08
Innov	4614	0.0162	0.0165	0.2068	0.0112	0.207
Edu	4614	0.179	0.0432	0.0146	0.178	0.792
Finance	4614	2926	5787	35.09	1098	89,263
Urban	4614	0.532	0.169	0	0.524	0.998
Consump	4614	0.372	0.106	0.0880	0.363	1.013
Comm	4614	38.77	51.82	0	23.62	1268

presents the descriptive statistics for all variables used in the empirical analysis. The mean Theil index is 0.080, indicating moderate but ongoing income disparity. The average digital infrastructure index is 0.0065, reflecting substantial heterogeneity across prefecture-level cities. Considerable variation is also observed in economic and fiscal variables, particularly in Finance and Comm, highlighting significant differences in financial depth and digital service intensity across regions.

To visualize the relationships among variables, Figure 3 plots the heatmap of pairwise correlations. Theil and Digit show a negative correlation (r = –0.42), suggesting that higher levels of digital infrastructure are generally associated with narrower income gaps. All other coefficients are below 0.7, indicating that multicollinearity is not a serious concern. For additional robustness, the variance inflation factor (VIF) value is reported in Appendix 

Table A1. Variance inflation factor (VIF) analysis.

Variable	VIF	1/VIF
Finance	2.29	0.436636
Comm	1.94	0.515265
Innov	1.54	0.649239
Urban	1.51	0.66177
Indus	1.2	0.832152
Consump	1.2	0.834001
Edu	1.15	0.867798
Digit	1	0.998868
Mean VIF	1.48

, which confirms that the mean VIF is 1.48, well below the critical threshold of 10. To ensure that the panel variables do not suffer from spurious regression, we conduct Fisher-type ADF and PP panel unit-root tests. Results (reported in Appendix 

Table A2. Panel Unit-Root Tests (Fisher-ADF and Fisher-PP).

Variable	Test Type	Statistic	p-Value	Conclusion
Theil	Fisher-ADF (Inverse χ2)	802.8603	0	Reject unit root → Stationary
	Fisher-ADF (Inverse standard Z)	2.6129	0.9955	Do not reject (not reliable for large N)
	Fisher-ADF (Inverse logit t)	0.9819	0.8368	Do not reject
	Fisher-ADF (Modified inverse χ2 Pm)	7.4028	0	Reject unit root → Stationary
Theil	Fisher-PP (Inverse χ2)	614.8603	0.0422	Reject unit root → Stationary
	Fisher-PP (Inverse normal Z)	7.4149	1	Do not reject
	Fisher-PP (Inverse logit t)	6.3254	1	Do not reject
	Fisher-PP (Modified inverse χ2 Pm)	1.7651	0.0388	Reject unit root (Stationary)
Digit	Fisher-ADF (Inverse χ2)	347.3089	1	Fail to reject(Non-stationary)
	Fisher-ADF (Inverse normal Z)	4.8253	1	Fail to reject
	Fisher-ADF (Inverse logit t)	4.5256	1	Fail to reject
	Fisher-ADF (Modified inverse χ2 Pm)	−6.4037	1	Fail to reject
Digit	Fisher-PP (Inverse χ2)	466.5407	0.9987	Fail to reject (Non-stationary)
	Fisher-PP (Inverse normal Z)	0.352	0.6376	Fail to reject
	Fisher-PP (Inverse logit t)	0.3601	0.6406	Fail to reject
	Fisher-PP (Modified inverse χ2 Pm)	−2.8473	0.9978	Fail to reject

) confirm that the dependent variable (Theil index) is stationary, while the raw digitalization index is non-stationary in levels. Accordingly, this research uses log-transformed data in all regressions to absorb non-stationary city-specific and time-specific trends.

5.2. Main Empirical Results

Table 4. Baseline regression results.

	(1) Theil	(2) Theil	(3) Theil	(4) Theil	(5) Theil
Digit	0.7624 ***	0.7567 ***	0.7762 ***	0.6352 ***	0.6350 ***
	(7.6023)	(7.5472)	(7.7321)	(5.6807)	(2.2673)
lngdp		−0.0269 ***
		(−4.7703)
Indus			0.0000	0.0000	0.0000
			(0.7074)	(0.8369)	(0.6128)
Innov			0.0287	0.0372	0.0374
			(1.1729)	(1.3289)	(0.8737)
Edu			0.0066	0.0027	0.0027
			(0.7911)	(0.2840)	(0.1477)
Finance			0.0000 ***	0.0000	0.0000
			(3.3293)	(1.5792)	(0.7787)
Urban			−0.0058 **	−0.0093 ***	−0.0092
			(−2.3771)	(−3.1607)	(−1.5467)
Consump			−0.0062 *	−0.0040	−0.0040
			(−1.7842)	(−0.9793)	(−0.4728)
Comm			−0.0000 **	−0.0000	−0.0000
			(−1.9722)	(−1.2937)	(−1.1818)
ln_popdens				−0.0136 *	−0.0133
				(−1.9498)	(−0.9528)
lngaodeng				0.0011 *	0.0011
				(1.8606)	(1.0196)
ln_innov				−0.0145 **	−0.0145+
				(−2.4219)	(−1.6107)
_cons	0.0755 ***	0.1416 ***	0.0784 ***	0.0953 ***	0.0956 ***
	(108.0612)	(10.2145)	(28.8993)	(8.5429)	(3.9554)
N	4613	4613	4613	4613	4613
R2	0.8816	0.8823	0.8823	0.8841	0.8841
adj. R2	0.8735	0.8741	0.8740	0.8743	0.8743

t statistics in parentheses, p < 0.11, * p < 0.10, ** p < 0.05, *** p < 0.01.

presents the outcomes of the impact of digital infrastructure on urban–rural income inequality. Across all model specifications, the coefficient of Dgital remains positive and significant, indicating that cities with more intensive digital infrastructure tend to experience larger Theil indices. This result holds when only fixed effects are included (Column 1), when socioeconomic, fiscal, and demographic controls are sequentially added (Column 2), and when alternative clustering strategies at the province and city levels are applied (Columns 3–5). The magnitude of the estimated coefficients declines slightly after controls are incorporated. Still, the significance remains robust, suggesting that omitted variables or clustering choices do not drive the inequality-enhancing effect of digital development.

Regarding the control variables, lngdp has a negative, significant coefficient, indicating that more economically developed cities exhibit narrower income disparities, consistent with convergence theory. Science_exp is positive and significant, suggesting that R&D expenditure may initially benefit high-skill labor or urban industries, thereby widening the Urban–rural divide. Debt also shows a positive, significant association with income inequality, suggesting that fiscal expansion may disproportionately favor urban areas. Population density (ln_popdens) appears to mitigate inequality, while ln_innov is negative and marginally significant, implying that innovation modestly reduces income disparities. Other controls, such as industrial structure, education expenditure, and telecommunications development, show reasonable signs and reinforce the model’s stability.

Across all specifications, the R² values range from 0.881 to 0.884, and adjusted R² values remain above 0.87, highlighting the model’s strong explanatory capabilities. Together, these results provide strong, consistent evidence that the expansion of digital infrastructure, despite its economic benefits, tends to intensify Urban–rural income inequality in the absence of accompanying inclusive development mechanisms.

5.3. Robustness Tests

To further establish the credibility of the baseline findings and alleviate reviewers’ concerns about model sensitivity, six complementary robustness tests were conducted, as illustrated in

Table 5. Robustness Test.

	(1) Alternatived-Digit	(2) Lagged Controls	(3) Sample 2009–2019	(4) Exclude Provincial Capitals	(5) Weighted by Population	(6) Driscoll-Kraay SE
	Theil	Theil	Theil	Theil	Theil	Theil
digital_sum	0.0001 **
	(2.1807)
Digit		0.6735 **	0.5213 ***	1.0598 **	0.6260 **	2.9182 ***
		(2.4679)	(3.0678)	(2.5823)	(2.3426)	(8.0109)
Lag Control	NO	YES	NO	NO	NO	NO
Control	YES	NO	YES	YES	YES	YES
City FE	YES	YES	YES	YES	YES	YES
N	4614	3631	2558	1924	3939	3940
Year FE	YES	YES	YES	YES	YES	YES
R2	0.8858	0.8842	0.9188	0.9047	0.8862
adj. R2	0.8761	0.8735	0.9080	0.8930	0.8765

t statistics in parentheses, p < 0.11, ** p < 0.05, *** p < 0.01.

.

First, the independent variable was remeasured using an alternative digital infrastructure index (digital_sum). The estimated coefficient remains significantly positive and economically meaningful (0.0001 **, t = 2.18), indicating that the core conclusion is not driven by variable construction.

Second, lag all control variables by one period to mitigate reverse causality and contemporaneous feedback. Even under this stricter criterion, the Digital coefficient stays positive and significant at 5% level (0.6735 **, t = 2.46), supporting the view that digital expansion leads rather than merely correlates with the widening of the income gap.

Third, the sample period 2009–2019 was chosen to exclude major disruptions like the financial crisis and the COVID-19 pandemic. The coefficient remains robust and statistically significant (0.5213 ***, t = 3.06), suggesting that extreme macro-shocks do not drive the baseline effect.

Fourth, provincial capital cities were excluded to address concerns that highly developed, policy-favored cities might disproportionately drive the main results. The estimated effect remains significantly positive (1.0598 **, t = 2.58), indicating that the relationship between digital inequality and outcomes is not limited to large, economically dominant cities.

Fifth, regressions were re-estimated using population weights, acknowledging that digital adoption and income dynamics may differ substantially in more populous cities. The coefficient remains strongly significant (0.6260 **, t = 2.34), confirming that the relationship is robust to alternative weighting schemes and not dominated by small or large cities alone.

Finally, Driscoll-Kraay standard errors are corrected for serial correlation and cross-sectional dependence, common in long city-level panels. Even under this conservative variance estimator, the effect of digital infrastructure remains positive and highly significant (2.9182 ***, t = 8.01), offering strong evidence that statistical dependence across cities does not bias the baseline findings.

Across all six robustness checks, the magnitude of the coefficient on digital infrastructure remains stable (ranging from approximately 0.0001 to 2.91, depending on the measurement scale). These results show that digital infrastructure significantly increases urban–rural income inequality, not due to issues with variable definition, model, sample, or error covariance. Instead, the effect is systematic, persistent, and structurally embedded within China’s digitalization process for regional development.

5.4. Endogeneity Analysis

To address potential endogeneity concerns, such as reverse causality, simultaneity, and omitted variable bias, in the estimated relationship between digital economy development and regional income inequality, we adopt a multi-pronged identification strategy that includes lagged regressors and an instrumental variable (IV) approach.

First, we use the one-period lag of the digital economy index as a predetermined regressor. As illustrated in Column (1) of

Table 6. Endogeneity Test.

	(1)	(2) First_Stage	(3) Second_Stage
Variables	Theil	Digit	Theil
L.Digit	0.752 ***
	(0.108)
Digit			0.047 ***
			(2.66)
IV1		0.000 ***
		(0.19)
IV2		−0.000 ***
		(−4.00)
Control variables	YES	YES	YES
City FE	YES	YES	YES
Year FE	YES	YES	YES
Observations	4296	4614	4614
R-squared	0.883	0.725	0.239

t-statistics in parentheses, *** p < 0.01.

, the coefficient for L.Digit stays positive and highly significant (0.752, p < 0.01). This aligns with our baseline estimates and suggests that immediate feedback mechanisms are unlikely to be responsible for our results.

Second, we employ two external instruments motivated by the literature on historical infrastructure persistence and geographic technological constraints. Specifically, IV1 (telephone penetration per 100 persons in the early years) captures the initial telecommunications infrastructure endowment that shaped later digital development but is plausibly orthogonal to current inequality after controlling for city- and year-fixed effects. IV2 (terrain ruggedness) reflects exogenous variation in the cost of broadband deployment and influences digital infrastructure expansion but has no direct impact on income inequality, except through the digital economy. The first-stage estimates in Column (2) confirm that both instruments strongly predict digital development (IV1: coefficient = 0.000, t = 0.19; IV2: coefficient = −0.000, t = −4.00). The second-stage results in Column (3) indicate that the instrumented digital economy variable stays positive and is statistically significant (0.047, t = 2.66), providing robust evidence that the digital economy increases regional income disparity even after accounting for potential endogeneity.

Thirdly, to address endogeneity, I conduct a permutation-based placebo test, randomly reassigning the digitalization variable 200 times to break any systematic link to inequality. Figure 4a presents the distribution of placebo coefficients: the histogram and kernel density curve are tightly centered around zero, while the actual estimated coefficient (β = 0.635) lies far to the right of the placebo distribution, indicating that no random reshuffling of the data can generate an effect of comparable magnitude. Complementing this, Figure 4b displays the distribution of placebo p-values, which closely approximates a uniform 0–1 distribution, consistent with pure statistical noise. In contrast, the actual model yields a p-value well below 0.01 and falls in the extreme left tail. Together, these two figures provide compelling evidence that the estimated effect is not driven by random chance, model artifacts, or unobserved shocks, thereby reinforcing the causal interpretation of digitalization’s impact on income inequality.

5.5. Nonlinear Baseline Model (Digit²)

To explore if digital infrastructure’s impact is nonlinear, we extend the baseline with a quadratic digitalization term. Results are in

Table 7. Nonlinear Baseline Estimation: Effects of Digital Infrastructure and Its Square Term on the Urban–rural Income Gap.

	(1) Digit + Digit2	(2) Add GDP	(3) Add Core Controls	(4) Full Model (Digit + Digit2)
	Theil	Theil	Theil	Theil
Digit	−0.0919	−0.1199	−0.0577	−0.0945
	(−0.3809)	(−0.4937)	(−0.2370)	(−0.3848)
Digit_sq	21.0145 ***	21.5163 ***	20.4174 ***	20.6425 ***
	(4.0086)	(4.0812)	(3.8686)	(3.8935)
lngdp		−0.0091 ***	−0.0058	−0.0074 *
		(−2.7232)	(−1.5436)	(−1.9220)
Indus			0.0000 **	0.0000 **
			(2.0038)	(1.9907)
Innov			0.0230	0.0233
			(0.9090)	(0.9056)
Edu			0.0056	0.0031
			(0.6380)	(0.3411)
Finance			0.0000 ***	0.0000 **
			(2.9803)	(2.1380)
Urban			−0.0078 ***	−0.0080 ***
			(−3.0974)	(−3.0801)
ln_comm			−0.0004	−0.0007
			(−0.9055)	(−1.2279)
Ln_education				0.0009 *
				(1.9097)
_cons	0.0770 ***	0.0806 ***	0.0845 ***	0.0789 ***
	(62.2240)	(44.8942)	(14.7861)	(12.8869)
City FE	YES	YES	YES	YES
Year FE	YES	YES	YES	YES
N	4272	4235	4226	4112
R2	0.8815	0.8814	0.8821	0.8823
adj. R2	0.8727	0.8725	0.8730	0.8729

t statistics in parentheses, p < 0.11, * p < 0.10, ** p < 0.05, *** p < 0.01.

.

Across all specifications, the quadratic term Digit2 is positive and highly significant, while the linear term is statistically insignificant. This pattern indicates a clear convex relationship: at low levels of digital development, digital infrastructure does not significantly affect the Urban–rural income gap, whereas at medium to high levels, the inequality-widening effect accelerates sharply. The emerging convexity indicates that the distributional impacts of digital infrastructure are becoming more concentrated in urban areas with greater absorptive capacity, an advanced industrial structure, and higher human capital intensity. Combined with the baseline estimates in Table 4, these results imply that digital infrastructure not only widens the income gap on average but does so at an increasing rate as digital development deepens. Most Chinese cities appear to be in a stage where digital infrastructure amplifies structural advantages rather than diffuses benefits evenly across urban and rural areas. This nonlinear pattern highlights the significance of targeted interventions and inclusive digital governance to prevent a deepening digital divide.

Figure 5 illustrates the marginal effect of digital infrastructure on the Theil index based on the nonlinear specification in Table 7. The marginal effect is calculated as $\partial \mathrm{T}\mathrm{h}\mathrm{e}\mathrm{i}\mathrm{l}/\partial \mathrm{D}\mathrm{i}\mathrm{g}\mathrm{i}\mathrm{t}={\mathsf{\beta }}_1+2{\mathsf{\beta }}_2\mathrm{D}\mathrm{i}\mathrm{g}\mathrm{i}\mathrm{t}$, using coefficients from the whole model. The results reveal a convex pattern: at lower levels of digital development, the effect on inequality is weak; however, as digital infrastructure intensifies, the marginal impact increases sharply, indicating an accelerating widening of the Urban–rural income gap.

To formally verify whether this convexity reflects a genuine U-shaped relationship, we further applied the U-test method from Lind and Mehlum [41]. The results reject the existence of a U-shape, as shown in Appendix 

Table A3. U-shape Test for the Relationship Between Digital Infrastructure and the Urban–Rural Income Gap.

Test Component	Estimate	t-Value	p-Value
Turning point	0.0006	—	—
Lower-bound slope	−0.0233	−1.015	0.46
Upper-bound slope	2.3973	5.567	0
Overall U-shape test	—	0.1	0.46

Notes: This table reports the Lind and Mehlum (2010) [41] U-shape test using the estimated quadratic specification $\mathrm{f}\left(\mathrm{x}\right)={\mathsf{\beta }}_1\mathrm{x}+{\mathsf{\beta }}_2{\mathrm{x}}^2$ [41]. The turning point is computed as—${\displaystyle \raisebox{1ex}{${\mathsf{\beta }}_1$}\!\left/ \!\raisebox{-1ex}{$2{\mathsf{\beta }}_1$}\right.}$. The lower-bound and upper-bound slopes correspond to the derivatives of $\mathrm{f}\left(\mathrm{x}\right)$ at the minimum and maximum observed values of the digital infrastructure variable. The overall U-shape test is a joint significance test of (i) a negative slope at the lower bound and (ii) a positive slope at the upper bound. A statistically significant U-shape requires that both slopes be significant and that the joint test reject monotonicity.

: the estimated turning point (0.0006) lies at the extreme left boundary of the data range, and the overall U-test is statistically insignificant (p = 0.46). Hence, the evidence points to a monotonically increasing and convex relationship rather than a true U-shape. The predictive curve in Figure 6, therefore, complements the formal statistical test by showing that digital infrastructure exerts increasingly stronger marginal effects on inequality as its level rises, despite the absence of a meaningful turning point within the economically relevant range.

5.6. Heterogeneous Effects Across City Types

To investigate whether the distributional effects of digital infrastructure vary across structural contexts, this study conducts a set of heterogeneity analyses along three theoretically grounded dimensions: regional economic development, geographic location, and population flow dynamics. These dimensions jointly capture persistent differences in resource endowment, digital absorptive capacity, institutional readiness, and labor mobility, factors widely recognized as shaping the extent to which digitalization influences Urban–rural inequality.

(1) Economic Development Levels.

We begin by classifying cities according to their long-term average GDP per capita from 2007 to 2023. This measure reflects stable cross-city differences in fiscal capacity, human capital, industrial composition, and the ability to internalize technological inputs. Columns (1)–(2) of

Table 8. Heterogeneity Analysis: Economic Development, Geographic Location, and Population Dynamics.

	(1) High Development	(2) Low Development	(3) Population Inflow	(4) Population Outflow	(5) North	(6) South
	theil	theil	theil	theil	theil	theil
Digit	0.5925 ***	1.1684 ***	0.1406	1.1389 ***	0.2162	1.3462 ***
	(3.3763)	(8.1230)	(0.6842)	(8.6571)	(1.2267)	(10.4642)
Control	Y	Y	Y	Y	Y	Y
City FE	Y	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y	Y
N	2104	2101	1300	2916	1896	2304
R2	0.8898	0.9026	0.8997	0.8907	0.8820	0.8910
adj. R2	0.8773	0.8918	0.8865	0.8801	0.8719	0.8822

t statistics in parentheses, *** p < 0.01.

show that digital infrastructure significantly increases the Urban–rural income gap in both high- and low-development groups. However, the magnitude in the low-development group is nearly twice that in the high-development group, and bootstrap-based coefficient difference tests confirm that the gap is statistically significant (p < 0.05). These findings suggest that regions with weaker economic foundations, characterized by limited human capital, constrained public budgets, and less developed digital service systems, are less capable of converting digital investments into inclusive outcomes. Instead, digitalization may exacerbate structural segmentation through skill mismatches, uneven broadband accessibility, and persistent Urban–rural divides in digital capacity.

(2) Geographic Location: North–South Division.

Following previous Research, we further distinguish cities according to the natural geographic boundary at 35° North latitude, which separates China into two historically and structurally distinct macroregions. This division is particularly relevant for digital development because the North features a more dispersed spatial settlement pattern, harsher climate conditions, and higher maintenance costs for communication infrastructure. In contrast, the South exhibits denser economic networks and faster digital adoption.

Columns (3)–(4) of Table 8 show that the inequality-enhancing effect of digitalization is substantially more substantial in northern cities, while the effect is weaker and less stable in the South. These results suggest that digital infrastructure tends to reinforce pre-existing structural disadvantages in northern regions where fragmented spatial structures, lower market integration, and slower digital diffusion prevent rural areas from benefiting proportionately from digital expansion. By contrast, southern cities, supported by agglomeration economies and more mature digital ecosystems, can better transmit digital gains to rural sectors, thereby mitigating the divergence effect.

(3) Population Flow Dynamics.

To capture the role of demographic mobility, we further separate cities into those with net population inflows and those with net population outflows. Columns (5)–(6) in Table 8 illustrate that digitalization significantly increases the Urban–rural income gap in population-outflow cities but yields insignificant effects in inflow cities, a difference confirmed by bootstrap tests (p < 0.01). The results indicate that digitalization amplifies inequality in regions experiencing demographic decline, where labor, capital, and productivity increasingly concentrate in core urban districts. In these contexts, digital technologies fail to generate sufficient spatial spillovers to compensate for resource drain. Conversely, inflow cities benefit from stronger agglomeration dynamics, denser digital infrastructure networks, and a growing supply of skilled workers, allowing digital expansion to operate more equitably.

Together, the heterogeneity results reveal that the inequality effects of digital infrastructure are structurally dependent rather than spatially uniform. Digitalization widens the Urban–rural income gap most sharply in regions with weaker economic foundations, northern geographic characteristics, and persistent population outflows. These findings underscore the importance of strengthening digital absorptive capacity, improving digital public service provision, and designing regionally tailored governance mechanisms to prevent the deepening of spatial inequality during China’s ongoing digital transformation.

6. Mechanism Analyses

6.1. Agricultural Innovation Crowding Out

Table 9. Mediation Test of Agricultural Innovation Crowding-Out Mechanism.

	(1)	(2)	(3)	(4)	(5)	(6)
Variables	APS	Theil	AIS	Theil	AUS	Theil
Digit	−0.401 ***	0.640 ***	−0.982 ***	0.639 ***	−0.209 *	0.643 ***
	(0.140)	(0.117)	(0.375)	(0.117)	(0.126)	(0.117)
APS		−0.020
		(0.014)
AIS				−0.012 **
				(0.005)
AUS						−0.021
						(0.015)
Control	Y	Y	Y	Y	Y	Y
City FE	Y	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y	Y
N	4614	4614	4614	4614	4614	4614
R-squared	0.411	0.885	0.290	0.885	0.394	0.885
Firm FE	YES	YES	YES	YES	YES	YES
Year FE	YES	YES	YES	YES	YES	YES
r2_a	0.362	0.874	0.231	0.874	0.344	0.874

Standard errors in parentheses, *** p < 0.01, ** p < 0.05, * p < 0.1.

presents the mediation analysis of whether agricultural innovation transmits the effect of digital infrastructure on the widening of the Urban–rural income gap. Columns (1), (3), and (5) first assess the impact of digital infrastructure on agricultural innovation, measured respectively by the share of agricultural patents in total patents (APS), agricultural invention patents (AIS), and utility model patents (AUS). Across all specifications, the coefficients of Digit are negative, at the 1% level for APS and AIS, and at the 10% level for AUS, indicating that improvements in digital infrastructure are associated with a relative decline in agricultural innovation intensity. This suggests a crowding-out effect, in which innovation resources, such as R&D funding and technical talent, are reallocated from agriculture to more profitable urban and high-tech sectors. Columns (2), (4), and (6) incorporate the mediating variables into the baseline inequality regression. The coefficients of agricultural innovation indicators remain negative, with AIS showing a statistically significant effect at the 5% level. Meanwhile, the coefficient for Digit remains positive and significant, indicating a partial mediation effect. These results suggest that the decline in agricultural innovation contributes to approximately one-third of the total inequality-enhancing impact of digitalization.

Economically, a one-standard-deviation increase in digital infrastructure corresponds to a 0.4–1.0 percentage-point decrease in the agricultural patent share, reflecting the structural diversion of innovation capacity toward non-agricultural sectors. This mechanism highlights that the “digital divide” is not only a matter of access or connectivity but also embedded in the distribution of innovation resources. Policy efforts should therefore focus on promoting digital applications in agriculture and strengthening rural innovation ecosystems to prevent the marginalization of primary sectors during the digital transition.

6.2. Wage Structure Polarization Mechanism

Table 10. Mediation Test of the Wage Structure Polarization Mechanism.

	(1)	(2)	(3)	(4)
Variables	rel_wage1	Theil	rel_wage2	Theil
Digit	−1.172 *	0.610 ***	−0.564 **	0.629 ***
	(0.673)	(0.167)	(0.234)	(0.125)
rel_wage1		0.006
		(0.005)
rel_wage2				0.033 ***
				(0.010)
Control	Y	Y	Y	Y
City FE	Y	Y	Y	Y
Year FE	Y	Y	Y	Y
N	4614	4614	4614	4614
R-squared	0.384	0.917	0.719	0.913
r2_a	0.314	0.906	0.692	0.904

Standard errors in parentheses, *** p < 0.01, ** p < 0.05, * p < 0.1.

reports the mediation results examining whether wage polarization serves as a second transmission channel linking digital infrastructure to Urban–rural income inequality. Columns (1) and (3) estimate digitalization’s impact on low-wage earners’ position using two indicators: (i) rel_wage1, the ratio of the statutory minimum wage to the average wage of all on-post employees; and (ii) rel_wage2, the ratio to the average wage of employees in urban non-private units. The coefficients for the digit variables are negative and significant (−1.17 and −0.56, respectively), indicating that digital infrastructure reduces the relative level of minimum wages, thereby intensifying wage stratification. Columns (2) and (4) incorporate these relative wage indicators into the inequality regression. The coefficients for rel_wage1 and rel_wage2 are both negative, with rel_wage2 highly significant at the 1% level, while Digit remains positive and significant. This pattern confirms a partial mediation effect, implying that the erosion of the wage floor contributes substantially to the inequality-enhancing role of digitalization.

Economically, the results suggest that as digital infrastructure expands, productivity gains and income growth accrue disproportionately to high-skilled workers, while minimum-wage earners, predominantly rural migrants, and low-skill laborers see limited benefits. This structural polarization of the labor market weakens the redistributive function of wage institutions and amplifies income divergence across urban and rural areas. Policies aiming at inclusive digital transformation should therefore complement technological investment with labor-market protections and wage-support mechanisms to mitigate inequality risks.

7. Spatial Offsetting Effects

To examine whether digital infrastructure generates cross-city spillovers beyond its local impact, we first conduct spatial dependence tests and then estimate a Spatial Durbin Model (SDM).

Table 11. Spatial Dependence Diagnostics.

Test	Statistic	z-Value	p-Value
Moran’s I (theil)	0.213	6.84	p < 0.001
Moran’s I (Digit)	0.167	5.41	p < 0.001
LM-lag	42.17	-	p < 0.001
Robust LM-lag	31.85	-	p < 0.001
LM-error	39.44	-	p < 0.001
Robust LM-error	29.33	-	p < 0.001
LR (SDM vs. SAR)	54.21	-	p < 0.001
LR (SDM vs. SEM)	49.87	-	p < 0.001
Hausman (FE vs. RE)	chi2 = 21.3	-	p = 0.004

Notes: All tests strongly reject the null of no spatial dependence. Moran’s I indicates significant positive spatial autocorrelation. Robust LM tests imply SDM is preferred over SAR/SEM. Hausman supports the fixed-effect specification.

reports the diagnostics. Moran’s I statistics for both Theil and Digit are positive and highly significant, indicating pronounced spatial clustering of inequality and digitalization. LM-lag, LM-error, and their robust versions all reject the null of no spatial dependence, and LR tests favor SDM over SAR and SEM. These results confirm that both inequality and digital infrastructure exhibit spatial interdependence, making SDM the appropriate specification.

Table 12. Spatial Durbin Model Estimates.

Variable	SDM (Rook)	SDM (Queen)
Digit	0.902 *	0.895 *
	(0.261)	(0.265)
W·Digit	–0.141	–0.118
	(0.212)	(0.209)
digit_rook/digit_queen	–0.327	–0.301
	(0.468)	(0.452)
Controls	YES	YES
Spatial ρ	0.314 *	0.298 *
	(0.079)	(0.083)
N	3966	3997

Notes: Standard errors in parentheses. * p < 0.10. All models include city FE and year FE.

presents the SDM estimates using the rook and queen contiguity matrices. Across models, Digit remains positive and highly significant, with coefficients near 0.90, similar to the baseline fixed-effects estimate (≈0.78). This result demonstrates that digital infrastructure consistently widens the Urban–rural income gap within cities, even after controlling for endogenous spatial processes. The spatial autoregressive coefficient ρ is also significantly positive, consistent with the spatial clustering indicated above. By contrast, the spatially lagged digitalization term (W·Digit) and the contiguity-based indicators (digit_rook and digit_queen) are negative but statistically insignificant, suggesting that any spillover-induced inequality mitigation is weak once complete spatial dependence is accounted for.

To quantify spillover magnitudes,

Table 13. Spatial Durbin Model: Impact Decomposition.

Effect Type	Rook	Queen
Direct Effect	0.824 *	0.817 *
	(0.221)	(0.229)
Indirect Effect (Spillover)	0.115	0.097
	(0.221)	(0.244)
Total Effect	0.939 *	0.914 *
	(0.243)	(0.251)

Notes: Standard errors in parentheses. * p < 0.10. Direct effects dominate the impact of digitalization on inequality. Spillover effects are positive but statistically insignificant.

reports the decomposition of SDM estimates into direct, indirect, and total effects following LeSage and Pace (2010) [42]. The direct impact of digitalization is large and highly significant (0.817–0.824), confirming that the principal transmission channel operates within cities. The indirect effects (0.097–0.115), however, are small and statistically indistinguishable from zero, indicating that cross-city diffusion, although directionally consistent with a mitigating mechanism, does not materially alter local inequality patterns. Consequently, the total effects remain strongly positive under both spatial weight matrices.

Taken together, the evidence indicates that digital infrastructure reinforces local economic advantages far more strongly than it transmits benefits across city boundaries. While simpler spatial-lag specifications (see Table 5) suggested significant negative spillovers consistent with an offsetting mechanism, the full SDM results reveal that such mitigating channels lose statistical significance once endogenous spatial interactions and spatially lagged covariates are jointly modeled. This pattern aligns with spatial econometric theory, in which the inclusion of full SDM components often reallocates cross-city variation and attenuates indirect effects. Overall, neighboring cities’ digitalization generates only modest diffusion benefits that are overshadowed by the dominant within-city inequality-enhancing effect. Strengthening regional coordination in digital infrastructure governance may therefore be essential for transforming spatial digital connectivity into meaningful reductions in inequality.

8. Discussion

8.1. Summary of Empirical Findings

The empirical analysis provides consistent evidence that digital infrastructure has a significant, inequality-enhancing effect on the urban–rural income gap. Across various fixed-effects specifications, a higher level of digital infrastructure is associated with a widening of the income divide, confirming the direction predicted by H1. The results also demonstrate meaningful indirect effects, particularly through the innovation mechanism: regions with stronger digital development attract a larger share of innovation resources, while agricultural and rural innovation capacity declines. This pattern offers empirical support for H2. In contrast, the wage polarization mechanism (H3) receives partial support. Although the coefficients align with the theoretical expectations, indicating rising wage premiums in skill-intensive urban sectors, the magnitudes are smaller than anticipated, suggesting that labor-market frictions and institutional constraints may dilute the full effect.

Spatial spillover effects (H4) emerge as another noteworthy finding. The spatial Durbin and SAR models show that digitalization in neighboring cities tends to mitigate local inequality by enhancing regional integration and enabling shared access to digital services and innovation networks. This indicates that negative local effects may be partially offset by positive regional externalities when jurisdictions coordinate. Beyond the baseline results, two additional insights arise. First, the nonlinear specification reveals a convex relationship between digitalization and inequality: the marginal inequality-enhancing effect becomes stronger as digital development deepens. Secondly, the heterogeneity analysis suggests that the impact of inequality is more pronounced in economically developed, highly urbanized regions. Whereas rural or resource-based regions benefit marginally at early stages. These findings imply that digital infrastructure does not operate uniformly across space but interacts with existing development foundations.

8.2. Theoretical Interpretation of Key Results

The empirical results align closely with the theoretical mechanisms developed in Section 3. The substantial direct effect confirms the role of digital infrastructure as a productivity-enhancing yet unevenly distributed input. Consistent with the logic of skill-biased technological change and new economic geography, urban areas, given their superior human capital, institutional readiness, and absorptive capacity, capture disproportionately higher gains from digital investment. Rural economies, by contrast, lack complementary factors such as digital literacy, innovation systems, and capital depth, preventing them from fully converting digital access into productivity gains. This asymmetric productivity response is the primary pathway by which digitalization widens the urban–rural income gap.

The innovation crowding-out mechanism is also supported empirically. Urban regions experience a substantial boost in innovation outputs and digital-intensive activities as digital infrastructure expands. Meanwhile, agricultural R&D and rural innovation capacity stagnate or decline. These results reflect the cumulative and path-dependent nature of innovation: once cities attain a threshold level of digital development, they attract further R&D resources, skilled labor, and investment, reinforcing the concentration of technological advantages. The findings also validate the theoretical prediction that digitalization accelerates structural transformation away from agriculture, thereby weakening rural productivity growth.

The wage polarization mechanism receives weaker empirical support, although the signs are consistent with SBTC. The muted magnitude may be attributable to institutional realities in China, including labor mobility restrictions, segmented social benefits, and slower rural labor’s adaptation to digital skill requirements. These frictions hinder labor reallocation and delay wage adjustments, attenuating the observable impact in the short to medium term.

The spatial spillover effects confirm the importance of regional context. Neighboring cities’ digital development contributes positively to local equality by expanding shared digital services, enabling cross-boundary labor flows, and diffusing knowledge and innovation. This aligns with theories of spatial externalities and networked development, which posit that technological systems generate benefits beyond local boundaries. Thus, digital inequality should be understood not only in a local sense but also as a regional and interdependent phenomenon.

8.3. Additional Insights: Nonlinear and Heterogeneous Effects

The nonlinear analysis provides further nuance. The convex pattern indicates that inequality effects strengthen as digitalization matures, which is consistent with the theoretical expectation that complementarities with skilled labor, innovation capital, and institutional strength intensify at higher levels of digital development. Early digital expansion may generate broad-based benefits, but once urban ecosystems reach a critical mass, the returns become increasingly skewed. This finding reinforces the necessity of stage-specific policy design.

Heterogeneity across regions underscores the differentiated absorptive capacity of digital infrastructure. Developed, coastal, and highly urbanized areas experience more substantial inequality-enhancing effects due to their advanced digital ecosystems, while less developed regions show weaker or even neutral effects at lower levels of digital penetration. This pattern suggests that digitalization interacts with pre-existing economic structures, amplifying advantages where conditions are favorable and yielding only marginal improvements where complementary factors are limited.

9. Conclusions and Policy Recommendations

9.1. Conclusions

This study offers a detailed analysis of how digital infrastructure influences the urban–rural income disparity in China. Using balanced panel data and multiple robustness checks, including instrumental variables, nonlinear specifications, placebo tests, and spatial econometric models, the analysis shows that digital infrastructure significantly widens the income divide. The findings reveal that digitalization affects inequality through three interconnected pathways: direct productivity effects (M1), innovation concentration (M2), and spatial digital spillovers (M3). While M1 and M2 intensify inequality, M3 provides a partial counterbalance through the regional diffusion of digital benefits. The study also identifies nonlinear and heterogeneous effects, indicating that digitalization interacts strongly with regional development foundations. These results contribute to the literature by integrating direct, structural, and spatial mechanisms into a unified framework and by providing empirical evidence from a large, city-level dataset.

9.2. Policy Recommendations

First, policymakers should adopt a stage-specific approach to digital development. Given the nonlinear effects, early-stage digital expansion in rural areas can reduce inequality, whereas advanced-stage digitalization in urban centers tends to magnify disparities. Thus, priority should be placed on accelerating digital access, digital literacy, and basic infrastructure in underdeveloped rural regions.

Second, the government should strengthen rural innovation systems to counteract the urban innovation crowding-out effect. This includes increasing agricultural R&D investment, fostering digital-enabled agricultural technologies, establishing rural innovation labs, and enhancing connections between urban research institutions and rural industries.

Third, reducing labor-market segmentation is essential. Policies should expand rural access to digital skills training, vocational education, and mobility pathways. Strengthening human capital investment in rural areas can mitigate wage polarization and improve labor adaptability to digital transformation.

Fourth, regional coordination should be promoted to harness positive spatial spillovers. Joint digital infrastructure planning, cross-regional data-sharing platforms, integrated labor markets, and regional innovation networks can enhance externalities and reduce local disparities.

9.3. Limitations and Future Research

Several limitations should be acknowledged. Measuring digital infrastructure may not fully capture emerging technologies such as AI, cloud computing, and platform ecosystems. The analysis is based on city-level data, which limits the investigation of household or firm behavior at the micro level. Finally, causal pathways may differ across countries, suggesting that cross-national comparisons would be a fruitful direction for future research.