Threshold regression models provide a robust methodological framework for capturing nonlinear dependencies and identifying structural shifts in the impact of independent variables across endogenously determined thresholds. These models yield more interpretable characterizations of nonlinear relationships than conventional linear specifications (Liu et al., 2021). Given the multidimensional nature of digital financial inclusion’s effects on industrial structure upgrading and the heterogeneous characteristics of county-level samples, this study employs a threshold framework to examine how digital financial inclusion (DFI) modulates industrial structure upgrading (ISU) in counties.

The county-level DFI Index is co-developed by Peking University’s Digital Finance Research Center and Ant Financial Services Group, providing annual metrics on financial digitization. Explanatory and socioeconomic control variables are sourced from the China County Statistical Yearbook and the county-level module of the China Stock Market & Accounting Research (CSMAR) database. To ensure data consistency and integrity, this study employs a unified county/district administrative code as the primary key to cross-verify and supplement overlapping variables across the two datasets. The dataset with higher integrity and accuracy serves as the benchmark. Missing values in the other dataset are imputed using a “ID + year” matching approach, ultimately constructing a preliminary panel dataset.

The underlying county-level data encompass approximately 2,800 counties: the Digital Financial Inclusion Index extends from 2014 to 2021, while the socio-economic variables span 2000–2022. However, there are significant data missing of these two series data in 2021. Balancing data availability and annual data integrity, the final sample is an unbalanced panel of 1,772 Chinese counties observed from 2014 to 2020 after eliminating several county sample individuals whose data were seriously missing in the preliminary panel data. This paper finally constructs an unbalanced panel that extracts the maximal informational value from the available records.

When ISU is the explained variable, the levels of DFI_ia in 1772 county-level administrative units are estimated with no threshold, one threshold, and two thresholds, respectively. Drawing on Hansen’s “bootstrap” method, the p-value corresponding to the test statistic is obtained. Besides, repeated sampling 1,000 times to robustly test whether there is a threshold effect. The regression results are shown in Table 2.

There are two threshold values in this model, and Figure 3 shows the estimated threshold values which are 0.9079 and 1.0366, respectively, which can be considered credible and effective. This verifies the viewpoint proposed in the theoretical analysis that DFI has a threshold effect on the ISU.

To reinforce the economic meaning of estimated thresholds as markers of distinct DFI development stages, this study characterizes county-level economic and financial features across the three threshold intervals by drawing on core findings from the Peking University Digital Financial Inclusion Index Report (2021) and the China Rural Development Report (2020).

In the low DFI stage (DFIia ≤ 0.9079), counties align with the “underdeveloped rural areas” category defined in the China Rural Development Report (2020). These regions face binding constraints from inadequate digital infrastructure, with rural mobile payment terminal penetration below 35% and village-level digital service station coverage at only 40%, consistent with the traits of low-digitization regions documented in the Peking University Digital Financial Inclusion Index Report (2021). Financial penetration remains low, as the average per capita digital credit scale in these counties is less than 0.8 ten thousand RMB and rural household adoption of digital financial products falls below 25%. Economically, traditional agriculture dominates local output, the tertiary sector accounts for less than 38% of GDP, and per capita GDP is more than 40% below the national county average. Such weak economic fundamentals limit the ability of DFI to drive effective industrial restructuring.

In the moderate DFI stage (0.9079 < DFIia ≤ 1.0366), counties are classified as “transitional development areas” in the China Rural Development Report (2020). The Peking University Digital Financial Inclusion Index Report (2021) notes substantial improvements in digital infrastructure here, with rural mobile payment terminal penetration exceeding 60% and village-level digital service station coverage reaching 75%. Financial penetration rises markedly, with the average per capita digital credit scale ranging from 2.5 to 3.0 ten thousand RMB and rural household usage of digital financial products climbing to 55–60%. Economically, the tertiary sector’s share of GDP exceeds 42%, and emerging industries including local characteristic manufacturing and modern agriculture have started to agglomerate, forming a development pattern complementary to DFI progress. The coordinated upgrading of digital infrastructure, financial services and industrial structure allows DFI to fully exert its factor reallocation and innovation induction effects, delivering the maximum promotional impact on ISU.

In the high DFI stage (1.0366 < DFI_ia), counties fall into the “developed rural areas” category in the China Rural Development Report (2020). The Peking University Digital Financial Inclusion Index Report (2021) notes these regions have advanced digital infrastructure, with rural mobile payment terminal penetration above 85% and village-level digital service station coverage nearly saturated. Financial penetration is high, as the per capita digital credit scale exceeds 4.0 ten thousand RMB and rural household adoption of digital financial products surpasses 70%. Both reports, however, highlight persistent structural constraints in these regions: traditional industries still account for a large share of the local industrial structure, and the absorption capacity of high-value-added sectors is limited by path dependence.

The regression results of DFI_ia on industrial structure upgrading are shown in Table 3. Among them, column (1) uses the OLS method and only takes all control variables into account; column (2) adds individual-fixed effects and time-fixed effects based on column (1); column (3) clusters the standard errors to the county-level based on column (2). Column (1)–(3) shows the credibility of control variables and the importance of the DFI_ia level. The regression results all indicate that DFI significantly promotes ISU, in other words, the net impact of DFI on industrial structure upgrading ISU is significantly positive. This paper uses the regression results of column (4) as the benchmark regression results. When the threshold variable is DFI_ia, there are significant differences of the impact of different DFI_ia values on ISU. When the development level of DFI is relatively low (DFI_ia ≤ 0.9079), the correlation coefficient of regression is 2.08%; the development level of DFI is moderate (0.9079 < DFI_ia ≤ 1.0366), the correlation coefficient of regression is 3.40%; correspondingly, the development level of DFI is high (1.0366 < DFI_ia), the correlation coefficient of regression is 2.82%, above all results are significant at the 1% level. It can be seen that the positive impact of DFI on the ISU first increasing and then decreasing as the level of DFI increases.

DFI exerts a positive impact on ISU that rises and then falls with increases in DFI development, a nonlinear pattern rooted in the intrinsic dynamics of DFI’s operation in county-level economies rather than a simple case of diminishing marginal effects. First, DFI’s factor reallocation channel is constrained by the industrial foundations of counties, which are dominated by low-productivity traditional sectors with limited high-value-added industrial carriers. At high levels of DFI development, the scale of capital supply outstrips the absorption capacity of real industrial sectors, and excess capital cannot be efficiently channeled to technology-intensive or service-oriented industries, reducing the efficiency of resource reallocation. Second, DFI’s innovation-inducing effect is bounded by the digital adaptability of county-level market entities. Translating DFI into higher R&D investment and technological progress depends on the ability of microenterprises and market participants to utilize digital financial tools. At high DFI levels, the growing complexity of financial products and higher digital literacy requirements may exceed the adaptive capacity of these entities, hindering the conversion of financial support into tangible innovation outputs. Third, DFI’s industrial integration effect is limited by the capacity of county-level digital infrastructure to keep pace. The deep integration of financial services with real economic activity relies on robust digital infrastructure and tailored scenario-based application conditions. When DFI development outstrips the supporting capacity of such infrastructure, its ability to penetrate traditional industries and foster new integrated sectors is weakened. By contrast, at moderate DFI levels, coverage breadth and usage depth advance in coordination, and digital infrastructure and market adaptability fully match DFI development. This alignment allows DFI’s factor reallocation, innovation-inducing and industrial integration effects to operate synergistically, maximizing its promotional impact on ISU. This evidence confirms a nonlinear phase shift in DFI’s impact on ISU and validates Hypothesis 1.

China’s county economies operate under constraints of weak industrial foundations and sluggish factor mobility, making industrial structure upgrading a gradual process. A short-term marginal improvement of 2–3 percent thus represents substantive progress in county-level structural transformation. Moreover, DFI’s promotional effect on ISU will be continuously amplified through compound cumulative effects, policy synergies with digital infrastructure development and industrial support measures, and the tangible resource reallocation effects it generates. In further discussion, it is assumed that the proportion of the output value of the primary industry in a region remains unchanged, then a 1% increase in tertiary industry share triggers a symmetric 1% decrease in secondary industry share. This implies a structural shift in the advanced industrialization phase. The average industry composition of the sample as a whole (ratio of RegGDP_primary to RegGDP:18.80%; ratio of RegGDP_secondary to RegGDP:39.35%; ratio of RegGDP_tertiary to RegGDP:41.96%; ISU = 2.2291). Besides, this paper attempts to quantify the resource reallocation effect specific to the threshold of digital inclusion:

The baseline analysis is conducted using a model with both time and individual fixed effects, with standard errors clustered at the county level, to alleviate potential endogeneity issues. To specifically tackle reverse causality (Li et al., 2022), an instrumental variable approach (IV-2SLS) is implemented, utilizing the following two instruments:

To further explore the nuanced relationship between DFI and ISU, this paper conducted threshold-based endogenous analyses. The full sample was divided into low, medium, and high development subgroups based on the DFI_ia thresholds identified earlier (0.9079 and 1.0366). IV-2SLS estimation was then performed within each subgroup, the regression results are shown in Table 4. The main conclusions of the endogeneity test are as follows. (1) The Kleibergen-Paap rk LM statistics are significant at the 1% level across all three subgroups, decisively rejecting the null of underidentification. (2) The Cragg-Donald Wald F statistics exceed the Stock-Yogo critical values (10% maximal IV bias level), confirming the absence of weak instruments. (3) Hansen J statistics reveal heterogeneity across subgroups: they remain insignificant for the Low and High DFI_ia groups (suggesting potential overidentification), but are significant for the Medium DFI_ia group at 90% confidence level. This indicates our instruments variables set satisfies the exogeneity requirements most reliably for counties with high DFI_ia development. In the first-stage regression, the F test of excluded instruments, Cragg-Donald Wald F statistic, and Kleibergen-Paap rk LM statistic are all significant, suggesting that the instrumental variables are highly correlated with the endogenous variable DFI_ia, and there is no under-identification issue, thus the instrumental variables are valid. On the whole, the Hansen J statistic and Stock-Wright LM S statistic show that the instrumental variables satisfy the exogeneity assumption, and there is no over-identification problem. In the second-stage regression, the coefficients of DFI_ia on the industrial structure upgrading index (ISU) are significantly positive in all threshold intervals, indicating that after controlling for endogeneity, the conclusion that DFI promotes county-level industrial structure upgrading still holds, and the baseline regression results are robust. In the second stage of regression, the coefficient of DFI_ia to ISU is significantly positive in all threshold intervals, indicating that after controlling endogenous, the conclusion that DFI nonlinearly promotes the upgrading of county industrial structure is still valid, and the baseline regression results are stable. The estimated coefficients are 0.0412, 0.2703, and 0.1518 for the low, medium, and high groups, respectively. This robustly affirms the core finding that digital financial inclusion promotes county-level industrial upgrading, with effects exhibiting heterogeneous persistence across DFI development levels.

The set of robustness tests provides substantial evidence for that there is no simple linear relationships between DFI and ISU, but threshold effect not only the trend but also the numerical magnitude. These findings align with the results obtained from the benchmark regression analyses, further validating their credibility.

The DFI_ia comprises three dimensions: DFI_cb (54%), DFI_ud (29.7%), and DFI_dl (16.3%). These dimensions collectively capture the development level of digital financial inclusion with greater specificity. Given their distinct focuses, a deeper examination of the relationship between these DFI components and ISU is essential to yield more practical and actionable insights. The developmental trends of these three DFI_sub are illustrated in Figure 4.

Similar to benchmark regression design, the threshold regression results of three sub dimensions are shown in Table 7. This indicates that all DFI_cb, DFI_ud, and DFI_dl have two threshold values on the promotion of industrial structure upgrading.

As shown in column DFI_cb of Table 8, when the development level of DFI_cb is relatively low (DFI_cb ≤ 0.7968), the regression coefficient is −0.0065 (statistically insignificant), indicating no significant promotional effect, even a weak inhibitory impact. At the moderate level (0.7968 ≤ DFI_cb ≤ 0.8765), the coefficient rises to 0.0185, exerting a significant positive effect on ISU. For the high-level regime (0.8765 < DFI_cb), the coefficient is 0.0068, remaining significant but with a smaller magnitude than the moderate level, reflecting diminishing marginal returns. This suggests that DFI_cb exhibits a threshold effect with diminishing marginal returns in the high-level stage.

As shown in column DFI_ud of Table 8, the impact of DFI_cb on ISU varies across levels. At the low level (DFI_ud ≤ 1.0044), the coefficient is 0.0006 (statistically insignificant), meaning no notable effect. For the moderate level (1.0044 < DFI_ud ≤ 1.3271), the coefficient becomes 0.0150, significantly promoting ISU. At the high level (1.3271 < DFI_ud), the coefficient is 0.0102, remaining significant but with a smaller magnitude, forming an inverse U-shaped pattern. This pattern implies that insufficient utilization fails to drive industrial upgrading, while excessive usage may lead to diminishing returns.

As shown in column DFI_dl of Table 8, DFI_dl affects ISU with distinct intensity across stages. At the low level (DFI_dl ≤ 0.9116), the coefficient is 0.0185, significantly positive. For the moderate level (0.9116 < DFI_dl ≤ 0.9691), the coefficient peaks at 0.0422, showing the strongest promotion. At the high level (0.9691 < DFI_dl), the coefficient declines to 0.0254, remaining significant but weaker than the moderate stage, presenting an inverse U-shaped pattern. This indicates that moderate digitization optimally drives industrial upgrading, while excessive digitization may trigger diminishing marginal gains.

Notably, the consistency in coefficient magnitude and robustness between the baseline regression results and those for DFI_dl underscores that digitization serves as the foundational driver of DFI’s overall impact on industrial structure upgrading. This alignment reflects that the technological infrastructure underlying digital finance, such as data transmission efficiency and smart service capabilities, exerts a more persistent and core influence compared to coverage breadth or usage depth. In contrast, the threshold effect of DFI_cb suggests that mere expansion of service networks yields limited returns until a critical scale is reached, likely due to underutilization of early-stage infrastructure. The inverse U-shape of DFI_ud indicates that moderate financial engagement optimizes resource allocation, while excessive usage may lead to misallocation or saturation in traditional sectors. Together, these findings highlight that DFI’s promotion of industrial upgrading relies on synergies across dimensions. Coverage breadth lays the foundation, usage depth amplifies resource flow, and digitization level sustains long-term efficiency gains, with digitization emerging as the most stable and impactful pillar.

Chose the sum of the number of patent invention publications and authorization represents the county innovation (Ino), the GDP of the secondary industry is used to represent the level of manufacturing development (MD) as the mediating variables. To mitigate the potential impact of excessive abnormal fluctuations, logarithmic transformations are applied to the Ino and the MD. This paper builds the mediation model shown below Equation (5): M it = α + β × DF I it + κ × C it + ε it (5)

Where, M_it presents mediator variables. To address endogeneity concerns including reverse causality and omitted variable bias that stem from the contemporaneous correlation between the mediating variables (Ino and MD) and both DFI and ISU, this study first employs a panel model with county and time fixed effects and clusters standard errors at the county level. This specification absorbs unobserved time-varying county heterogeneity and macro shocks that might induce co-movement between the core explanatory variable and mediating variables. Building on this framework, the IV-2SLS approach is further applied to the mechanism analysis, with one-period lagged DFI_ia selected as the instrumental variable for the core explanatory variable. l.DFI_ia acts as a predetermined variable that is exogenous to contemporaneous innovation, manufacturing development and ISU. This exogeneity effectively breaks the reverse causal association between current mediating variables and DFI, allowing the study to isolate the exogenous impact of DFI on the mediating variables.

The IV-2SLS regression results of Ino and MD are shown in Table 9. The correlation coefficients of innovation and manufacturing industry development are 175.08 and 5.38%, respectively, which are significant at the 1% confidence level. This shows that DFI can effectively promote the vigorous development of the manufacturing industry and innovation, which means the intermediary effect exists. The hypothesis H2a and H2b are proved. To quantify the mechanism’s economic significance, this study calculates the impact of one standard deviation (sd) increase in DFI_ia on mediating variables. With sd of DFI_ia = 0.2383 (Table 1), a one sd increase in DFI_ia promotes Ino by 1.7509 × 0.2383 ≈ 0.4172 (41.72%) and MD by 0.052 × 0.2383 ≈ 0.0124 (1.24%). These results confirm that DFI exerts a more pronounced impact on technological innovation than on manufacturing output expansion, which aligns with the theoretical expectation that digital finance prioritizes high-return innovation activities.

Mechanistic analysis of manufacturing development and regional innovation offsets the limitation of the ISU index in failing to fully capture within-sector technological upgrading and quality improvement, and bridges structural reallocation with technological progress. These findings collectively demonstrate that DFI drives industrial upgrading through two distinct channels: structural reallocation that elevates ISU and technological upgrading fueled by manufacturing expansion and regional innovation. This establishes a comprehensive understanding of the relationship between DFI and industrial upgrading. Though a dual strategy is adopted to address endogeneity in the analysis, residual endogeneity may still exist in the mechanism tests due to the inherent complexity of economic linkages. This residual endogeneity, however, has a negligible impact on the core conclusions of the mechanism analysis. The instrumental variable employed is one-period lagged DFI_ia, which is exogenous to all contemporaneous economic variables and highly correlated with current DFI_ia. This identification strategy ensures that the estimates capture the primary exogenous variation in DFI’s impact on the mediating variables. Furthermore, the causal transmission mechanism linking DFI, the mediating variables and ISU is consistent with the fundamental theoretical framework of financial development and industrial structural upgrading. DFI relieves microeconomic financing constraints through innovations in digital financial services, which in turn expands manufacturing production scale and raises regional innovation investment. It is noteworthy that the core conclusion confirming the existence of the mediating effect relies on the direction and statistical significance of the estimated coefficients, not their exact numerical values.

As a consequence of the heterogeneous nature of China’s economic development across different regions, notable disparities arise in terms of the economic development level, resource allocation, and industrial structure within county-level administrative units. The different development levels of DFI in different regions may have heterogeneity in their impact on upgrading industrial structure.

A subsample analysis is conducted by classifying counties according to their status as national-level poor counties. The results, shown in columns (1)–(2) of Table 10, reveal that digital financial inclusion significantly promotes industrial upgrading in poor counties, with consistently positive and significant coefficients. In contrast, the coefficient estimates are generally greater in magnitude for non-poor counties, implying a stronger promotional effect in these regions.

DFI’s promotional effect on industrial upgrading varies significantly by regional poverty status, with a more pronounced impact in non-poor counties. This stronger effect in non-national-level poor counties derives from three interrelated economic and institutional characteristics. First, these counties operate under relatively relaxed fiscal constraints. Their more stable fiscal revenues support complementary public investments in digital infrastructure and financial literacy initiatives, reducing frictions in financial resource allocation and enhancing DFI’s role in driving industrial upgrading. Second, non-national-level poor counties have more market-oriented financial environments. With a more complete layout of traditional financial institutions, credit markets exhibit greater competition and transparency. DFI thus complements the existing financial system, extending services to medium and high-value-added industrial sectors that were previously underserved. Third, these counties possess greater policy flexibility in industrial development. Less constrained by targeted poverty alleviation priorities, they can allocate financial resources including DFI to sectors with higher technological intensity and growth potential, such as advanced manufacturing and modern services, thereby amplifying DFI’s structural upgrading effects. In contrast, national-level poor counties face binding fiscal constraints, underdeveloped financial markets and policy orientations focused on basic livelihood security, all of which limit DFI’s capacity to foster high-value-added industrial growth.

The sample is stratified by whether counties belong to China’s contiguous areas of dire poverty. The regression results for these subsamples, presented in columns (3) and (4) of Table 10, are numerically and directionally consistent with the baseline findings. Within contiguous areas of dire poverty, the coefficients for DFI are positive, albeit only partially statistically significant, suggesting a positive yet relatively weak promotional effect on industrial upgrading. Conversely, for counties outside these areas, the coefficients are positive, statistically significant at the 1% level, and larger in magnitude, indicating a more pronounced effect of digital financial inclusion in promoting industrial structural upgrading.

The stronger promotional effect of DFI on ISU in non-contiguous areas of dire poverty is closely linked to their more favorable regional economic ecosystems and institutional environments. First, these areas have more concentrated industrial layouts. Proximity to urban agglomerations or transportation hubs fosters industrial agglomeration, enabling DFI to efficiently satisfy the diversified financial needs of clustered enterprises such as supply chain financing and R&D investment, and in turn accelerating industrial cluster upgrading. Second, non-contiguous areas of dire poverty suffer from lower degrees of financial repression. Less geographically isolated and with more open market conditions, these regions have deeper financial development that allows DFI to penetrate the real economy without hindrance from excessive administrative intervention or severe information asymmetry. Third, policy implementation is more effective in these areas. Better coordination among government departments ensures efficient execution of policies supporting industrial upgrading and digital finance, allowing DFI to align with industrial policies and generate synergistic effects. In contrast, contiguous areas of dire poverty face fragmented industrial layouts, high financial repression and challenges in policy implementation due to geographic and institutional constraints, leading to a relatively weaker impact of DFI on industrial structure upgrading.

In conclusion, DFI’s promotional effect on ISU exhibits significant poverty and regional heterogeneity, with its intensity closely tied to regional economic foundations, the sophistication of financial infrastructure and the level of industrial agglomeration. The relatively weaker effect in poor counties and contiguous areas of dire poverty reflects inherent structural constraints in regional economic development rather than failure of DFI to promote industrial upgrading in these regions. Poor counties face multiple structural barriers including weak industrial foundations, incomplete digital infrastructure and low digital financial literacy. These barriers directly reduce the efficiency with which DFI’s financial support translates into industrial upgrading momentum, resulting in a lower relative effect size compared to non-poor regions. This does not negate DFI’s foundational role in easing financing constraints and facilitating initial industrial development accumulation in poor areas. On the contrary, the fragility of industrial foundations and prevalence of financial exclusion in poor areas make DFI a more critical and irreplaceable financial tool for their industrial structure upgrading. This observation forms the core rationale for prioritizing poor areas in DFI development policies.