Agricultural Industrial Agglomeration and Agricultural Economic Resilience: Evidence from China

Abstract

Climate volatility and market uncertainty pose significant challenges to agricultural stability. We assess whether and how agricultural industrial agglomeration shapes China’s agricultural economic resilience using province-level panel data for 2003–2023 and a transparent, entropy-weighted index spanning resistance, recovery, and adaptability. Four results stand out. First, in a two-way fixed-effects model, agglomeration is associated with higher resilience on average, and this finding remains robust across multiple robustness tests and after addressing endogeneity concerns. Second, regional subgroup analyses reveal pronounced heterogeneity, providing evidence for geographically targeted policy design. Third, mechanism analysis reveals that the agricultural research intensity serves as a partial mediator between agglomeration and resilience. Fourth, the agglomeration-resilience relationship is nonlinear—N-shaped in the aggregate, while panel quantile regressions reveal an inverted-U among low-resilience provinces and an N-shaped pattern at the median and upper end of the distribution. In an extension, global Moran’s I statistics for three alternative resilience indices reveal significant positive spatial autocorrelation, indicating that agricultural economic resilience tends to cluster geographically and that spatial spillovers are likely to be present. In conclusion, agglomeration is a net enhancer of agricultural economic resilience, but its payoffs are agglomeration- and distribution-dependent: gains taper or reverse around the mid-range for low-resilience provinces, while the median and upper segments benefit again as specialization deepens, in a setting where resilience itself is spatially clustered. Reinforcing the research channel and tailoring actions to local resilience levels are therefore pivotal.

1. Introduction

1.1. Research Background and Conceptual Framework

Global warming, more frequent climate extremes, intensifying geopolitical shocks, and deepening supply chain interconnections have subjected agricultural systems to unparalleled external strain and systemic uncertainty regarding food security and eco-logical sustainability [1]. In response, enhancing the risk resistance and resilience of agricultural systems to protect food security and sustainable development has become a central concern of policymakers globally. Against this background, recent policy documents in China have shifted the emphasis from solely pursuing output increases to enhancing the stability of the agricultural system and promoting agricultural modernization. These documents emphasize the need to strengthen systemic adaptability and robustness.

In this context, the notion of agricultural economic resilience (AER) has gradually become one of the keys for promoting the high quality of agriculture. Its basic implication is that, from a “Pressure–State–Response” (PSR) perspective, the agricultural system as a structured socio-economic–ecological system is able to keep basic operation, adjust itself and go through structural transformation under environmental perturbations through the interaction of external pressures (P), internal structural states (S) and adaptive responses (R). Meanwhile, with the continual gathering of new agricultural operational entities such as agricultural industrial parks, leading enterprises, and co-operatives, the spatial structure of agriculture is evolving towards developing clear industrial agglomeration features. This development raises the question in academic circles of whether, and how, industrial agglomeration can improve the resilience of agricultural systems. Much of the literature examines the impact of agricultural agglomeration on performance variables such as production efficiency, resource allocation and farmers’ income. In recent years, other scholars have turned to a resilience perspective, arguing that there is an optimal level of agglomeration that can enhance system stability and adaptability via technology diffusion, organizational coordination, and scale economies.

1.2. Literature Review

Slijper et al. (2022) [2] utilized data from the European Farm Accountancy Data Network to construct farm-level resilience indicators and found that farm size, pro-duction diversification, and policy support significantly influence agricultural economic resilience. However, there is no consensus on how to measure resilience: some studies conceptualize it as a single elasticity or a set of volatility metrics, rather than adopting a systematic composite evaluation framework. Furthermore, in mechanism analyses, most studies rely on linear regression models without exploring potential nonlinear relationships, dynamic evolution patterns, or mediating mechanisms. In particular, the role of industrial agglomeration as a key driver of enhanced system resilience—through pathways such as technological innovation, organizational synergy, and factor mobility—remains underexplored and warrants both empirical validation and theoretical ex-tension.

The notion of “resilience” was originally proposed by Holling (1973) [3] in ecology to describe a system’s capacity to absorb external perturbations while maintaining its structural and functional organization. With global environmental change and economic uncertainty becoming more prominent, this concept has increasingly been applied to socio-economic systems as a lens for assessing the sustainability of regional and industrial development. By analogy with economic resilience in non-agricultural sectors, resilience in agriculture refers to an agricultural system’s ability to maintain production under resource constraints, environmental challenges, and income variability, that is, its resistance, adaptability, and ability to recover [4]. Methodologically, some studies adopt an elasticity-based approach that infers the vulnerability and resilience of agricultural economic systems from the elasticity of output with respect to shocks. For example, Zampieri et al. (2020) [5] construct a system-level resilience indicator from crop-yield data and characterize elasticity and recovery capacity via the mean–variance relationship. In recent years, researchers have turned to more systemic perspectives informed by the PSR framework, evaluating how agricultural systems perform and respond under stress. For instance, Chen et al. (2025) [6] and Yao et al. (2024) [7] develop three-dimensional indicator systems—capturing resistance, recovery, and adaptability—to provide a comprehensive assessment of the agricultural economy’s dynamic stability.

In recent years, research on the determinants of AER has become more systematic, spanning fiscal support, technological progress, institutional/financial safeguards, digitalization, and exogenous risk. On climate risk, studies show that extreme events significantly depress AER, while policy instruments can partially offset these shocks—e.g., extreme heat and drought reduce AER, whereas agricultural insurance mitigates the impact [8]. From the institutional and financial perspective, broader financial coverage helps cushion the adverse effect of population aging on resilience and thus plays a stabilizing role during structural shocks [9]. Along the digital and technological dimension, evidence highlights the role of data factors, digital technologies, and intelligent supply chains: using dynamic qualitative comparative analysis, provincial configurations that combine digital inclusive finance with agricultural industrial digitalization are found to enhance AER through heterogeneous pathways across regions [10]; consistent with this, spatial-panel estimates indicate that the diffusion of digital technologies strengthens disturbance resistance and adaptive adjustment capacity within the agricultural system [11]. Across these strands, effects are likely to be mutually reinforcing: broader financial coverage and insurance uptake can amplify the returns to digital adoption, while local exposure to climate extremes conditions the benefits of technological upgrading—underscoring the value of coordinated policy mixes rather than single levers. Finally, demographic pressure remains a headwind: higher levels of population aging exert a significant negative impact on AER [12].

Against this backdrop of the continuous popularization of new economic geography theory in agriculture and regional development policies and spatial layout adjustment, a growing literature begins to pay attention to the role that agricultural industrial agglomeration plays in driving regional coordination and rationalizing factor allocation. Some empirical works suggest that industry agglomeration can not only improve the efficiency of resource allocation within a region but also narrow the urban–rural and regional gaps and enhance the overall stability of the system via income distribution effect and spatial spillover effect. For instance, Feng et al. (2023) [13] argue that regional integration can promote regional economic resilience considerably through enhancing factor mobility and resource sharing. Ding (2023) [14] concludes that agricultural agglomeration promotes the development of agricultural enterprises and cooperatives, and it raises farmers’ income and exerts positive spatial spillovers over time on surrounding regions, which can improve system adaptability and coherence. These studies, viewed from the angles of income distribution and resource flows, reveal the possible indirect promoting effects of agricultural agglomeration on system resilience. Notably, in recent years scholars have begun to place agricultural industrial agglomeration within the analytical framework of AER. Chao & Li (2024) [15] construct an AER index system from the perspective of resistance, recovery, and adaptability, and empirically test the relationship between agglomeration level and resilience strength, offering new theoretical boundaries for resilience research. Xie et al. (2025) [16] further find that agricultural insurance enhances agricultural resilience primarily by promoting horizontal agricultural agglomeration rather than vertical industry integration. Overall, existing research has made certain advances in depicting system sensitivity but has not yet fully embodied the structural features of agricultural resilience under the PSR framework. At the same time, the construction of relevant indicator systems remains relatively narrow, and is insufficiently capable of fully reflecting the dynamic processes by which agricultural agglomeration may enhance the resilience of agricultural systems.

In summary, existing empirical research has identified evidence of linkages between agricultural economic resilience, industrial agglomeration, and related structural factors. However, when examined through the lens of the PSR framework and a structure–function perspective, this body of work reveals four key limitations. First, current measures of AER are often narrow—relying on isolated indicators such as elasticity or volatility—or only loosely anchored to a coherent PSR conceptual foundation. Second, agricultural industrial agglomeration is typically treated as a background control variable or a linear covariate, rather than being conceptualized as a dynamic structural force that simultaneously influences system pressures, states, and adaptive responses. Third, studies on regional heterogeneity and underlying mechanisms remain fragmented: while cross-regional and cross-production differences are frequently noted in qualitative terms, critical transmission channels—such as technological innovation—are rarely quantified within an integrated analytical framework. Fourth, potential nonlinearities and distributional dynamics in the agglomeration–resilience relationship are seldom investigated in a systematic manner. These gaps, interpreted through the PSR and structure–function paradigms, constitute the central motivation for the empirical analysis presented in this paper.

1.3. Research Design, Main Findings, and Paper Structure

To address these gaps, this paper makes four integrated and methodologically coherent contributions: (i) the development of a multidimensional AER index grounded in the PSR framework, providing a transparent, theory-driven, and systematically justified measurement foundation; (ii) the mapping of context dependence through subsample regressions across diverse regional settings, revealing how the impact of industrial agglomeration varies with local socioeconomic and institutional conditions; (iii) the clarification of causal mechanisms via mediation analysis, identifying the pathways through which industrial agglomeration influences the evolution of resilience, thereby enhancing both pathway specificity and interpretability of underlying processes; and (iv) a rigorous examination of nonlinearity in the agglomeration–resilience nexus by incorporating squared and cubic terms of industrial agglomeration, enabling robust detection of potential N-shaped patterns, further complemented by quantile-based descriptive analyses and panel quantile regressions to uncover distribution-dependent heterogeneity. Beyond these core contributions, we conduct an extension analysis that evaluates and compares the spatial autocorrelation of AER indices derived from three distinct construction methods using Moran’s I statistics. The results indicate that while positive spatial clustering is consistently observed across all specifications, the magnitude and statistical significance of this clustering are sensitive to the choice of indicator composition and weighting scheme.

Building on prior research, this study establishes a PSR-based AER evaluation system encompassing multiple dimensions and constructs an entropy-weighted AER index for 31 Chinese provinces over the period 2003–2023, followed by estimation using two-way fixed-effects models. From a nonlinear perspective, the analysis reveals a clear N-shaped, three-phase evolutionary relationship between industrial agglomeration and resilience—comprising an initial phase of increasing returns, an intermediate congestion window, and a subsequent phase of specialization-driven gains. Applying panel quantile regressions, we demonstrate that this nonlinearity is regime-dependent: low-resilience provinces exhibit a positive–positive–negative N-shaped pattern with attenuation in the late stage, whereas provinces in the median and upper resilience quantiles display a more pronounced N-shaped trajectory with shifting turning points. We further conduct three subgroup analyses—by region (eastern vs. central vs. western), by grain production function (major producing vs. selling vs. balanced areas), and by investment level (high- vs. low-investment provinces)—to investigate regional heterogeneity. These analyses reveal systematic variations in effect magnitudes and distinct agglomeration–resilience profiles across subgroups. Finally, we examine a knowledge-sharing mechanism: agricultural agglomeration enhances agricultural research intensity (R&D), which in turn strengthens resilience. Mediation analysis confirms this pathway, and the result remains robust when employing an alternative AER index (AER2) that excludes innovation-related components, thus mitigating concerns about endogenous overlap with R&D inputs.

The paper is structured as follows: Section 2 outlines variable selection, indicator construction, and model specification; Section 3 presents baseline regression results, robustness checks, heterogeneity analysis, mediation mechanism tests, nonlinear dynamics, and an extended analysis of spatial spillover effects; Section 4 provides a comprehensive discussion of the findings and identifies promising directions for future research; and Section 5 summarizes the key conclusions and discusses their policy implications. Figure 1 illustrates the research roadmap for the core empirical modules.

2. Materials and Methods

2.1. Variable Definition

2.1.1. Agricultural Economic Resilience Index

Based on a balanced panel dataset of 31 Chinese provincial-level regions from 2003 to 2023, this study constructs a composite AER index as the dependent variable to capture the stability, recovery capacity, and adaptability of regional agricultural systems in response to external shocks. The analysis is conducted at the provincial level, with a focus on the post-2003 period, as this year marks the first time all provinces consistently report the full set of PSR-based indicators—covering production, technology, environment, and digital agriculture—in a sufficiently complete and harmonized manner. The provincial level is adopted because it is the lowest administrative tier at which this full PSR indicator set is consistently reported nationwide, and it also corresponds to the main level at which agricultural policies, fiscal support, and industrial and resilience strategies are formulated and implemented in China. Prior to 2003, several core variables suffer from substantial and uneven missing data, which would compromise the construction of a consistent and comparable longitudinal panel.

Given that there is currently no universally accepted definition of AER, we follow the approach of Lin et al. (2025) [17] and apply the entropy method to develop a comprehensive evaluation framework structured around the three dimensions of resistance, recovery, and adaptation. Building on the PSR framework introduced above, agricultural industrial agglomeration interacts with this configuration in a systematic way: on the pressure dimension, the spatial concentration of farms and agro-related enterprises can mitigate idiosyncratic risks through shared infrastructure and market access, but excessive concentration may also increase environmental stress and systemic exposure to extreme events; on the state dimension, agglomeration reshapes the structural characteristics of regional agriculture by raising average productivity, thickening local pools of labor and knowledge, and modifying the balance between specialization, diversity and redundancy; and on the response dimension, dense clusters facilitate collective actions such as joint R&D, contract farming, mutual insurance and coordinated recovery, which expand the set of feasible adaptive responses and speed up post-shock adjustment. Through these structure–function feedbacks, different degrees and forms of agglomeration translate into different combinations of resistance, recovery and adaptability, and hence into different levels of AER. The overall framework of indicator construction is illustrated in Figure 2.

As an objective method of weighting based on information entropy theory, the entropy method makes full use of the information features of each indicator. It also considers the degree of variation in indicator values to attribute weights without the need for human intervention, and the indicators with higher discriminatory power will be given higher weights enhancing the accuracy of the whole index system. Unlike traditional subjective weighting methods, the entropy method is free of human bias, and can be applied to complicated, multidimensional and diverse systemic variables such as AER [18,19].

Specifically, drawing on the different response stages of agricultural systems under external shocks, the index system is structured along three dimensions—resistance, recovery, and adaptation—each of which includes several secondary and tertiary indicators that comprehensively represent the stability, technological foundation, and adaptive capacity of agricultural systems [20]. To more effectively capture the technological adaptability of the system, the number of authorized agricultural invention patents is introduced as a representative indicator, reflecting regional agricultural technological innovation output.

The constructed AER index system is analogous to stability and the ability to recover and adapt to the agricultural system to external shocks including the basic elements of agricultural production, technology support, resource and environmental conditions, and financial assurance. Because of its relatively clear hierarchical structure and coverage, the system can be regarded as a rather good representation of the agricultural system’s performance at each stage, thereby providing good explanatory and operational power for both cross-regional resilience comparisons and the assessment of policies.

Table 1. Comprehensive Indicators of Agricultural Economic Resilience.

Primary Indicator	Secondary Indicator	Tertiary Indicator	Indicator Description	Attribute
Shock Resistance	Agricultural Production Capacity	Yield per Unit Area	Agricultural output value/sown area of crops	+
		Grain Yield per Unit Area	Total grain output/sown area of grain crops	+
	Natural Environmental Conditions	Forest Coverage Rate	Forest coverage rate in each region	+
	Agricultural Technology Level	Number of Digital Agriculture Enterprises	Number of digital agriculture enterprises in each region	+
Recovery Capacity	Agricultural Mechanization Level	Rural Per Capita Mechanical Power	Total agricultural machinery power/rural population	+
	Agricultural Disaster Management	Area of Soil and Water Conservation	Area of soil erosion control in each region	+
	Agricultural Fiscal Support	Agricultural Fiscal Expenditure	Expenditure on agriculture, forestry, and water affairs/general budget expenditure	+
	Innovation Capacity	Number of Granted Agricultural Invention Patents	Number of granted agricultural invention patents in each region	+
Adaptive Capacity	Agricultural Production Pressure	Disaster Severity	Affected area/disaster-hit area	−
	Resource and Environmental Load	Fertilizer Input	Fertilizer consumption (pure nutrients)/sown area of crops	−
		Pesticide Input	Pesticide consumption/sown area of crops	−
		Agricultural Film Input	Agricultural film consumption/sown area of crops	−

Note: In the AER indicator system, the symbol “+” denotes a positive indicator, where a higher value indicates greater resilience, while “−” denotes a reverse indicator, where a larger value reflects weaker resilience and requires reverse transformation prior to standardization. This directional coding ensures that the final synthesized AER index consistently follows the principle that higher values correspond to higher levels of agricultural economic resilience.

presents in detail the construction of the indicator system, variable selection, calculation methods and contributions.

2.1.2. Agricultural Industrial Agglomeration

Methods for measuring the degree of agricultural industrial agglomeration mainly include the location quotient, the Herfindahl–Hirschman Index, and the Gini coefficient. Among these, the location quotient has been widely applied in studies of agricultural agglomeration because of its computational simplicity, suitability for regional comparisons, and ability to stably reflect the relative concentration of a specific industry. Accordingly, this study adopts the location quotient as the measurement indicator of agricultural industrial agglomeration (LQ), calculated using data from 31 Chinese provincial-level regions and the national average over the period 2003–2023. The calculation formula is given as follows:

$${LQ}_{i,t =}{\displaystyle \frac{\frac{E_{i,t}}{E_{i,t}^{total}}}{\frac{N_t}{N_t^{total}}}}$$

(1)

where E_i,t denotes the agricultural GDP of region i at time t, E_i,t^total represents the total GDP of region i at time t, N_t is the national agricultural GDP at time t, and N_t^total is the national total GDP at time t. If LQ_i,t > 1, it indicates that the agricultural industry in a given region and year is more concentrated than the national average, reflecting the presence of agglomeration characteristics. In addition, we also construct an employment-based location quotient (LQ2), defined analogously to (1) by replacing output shares with employment shares to capture agglomeration from the labor perspective.

2.1.3. Control Variables

The empirical model incorporates the dependent variable, AER, and the core explanatory variable, LQ, alongside a set of control variables to mitigate potential omitted variable bias. The selection and measurement of these controls are detailed as follows:

• Human Capital Level (Edu): Following the approach of Xiao (2021) [21], labor quality is proxied by the average years of schooling.
• Urbanization Rate (Urban): This is defined as the proportion of the urban population to the total population, serving as an indicator of urban development levels [22].
• Industrial Structure Upgrading Index (ISU): This index is measured by the proportion of the added value of the tertiary industry to the added value of GDP, measuring the extent of industrial transformation [23].
• Economic Development Level (E): Represented by the logarithm of per capita GDP, this variable controls for fundamental regional economic capacity and income levels, which constitute a foundational determinant of agricultural system resilience.
• Logarithm of total social fixed asset investment (I): Captures capital accumulation and investment capacity, which play an important role in improving agricultural production conditions and optimizing industrial structure.

2.1.4. Instrumental Variable

The number of landline phones per 100 people in each provincial administrative region in 1984 interacted with year (tele): To address endogeneity concerns in the fixed-effects model, we rely on a historical, time-varying instrument. In particular, we employ the number of landline telephones per 100 people in each provincial administrative region in 1984 interacted with a continuous year variable (2003–2023) as the instrumental variable (IV) for LQ. This strategy exploits long-lasting historical differences in regional telecommunication infrastructure, while allowing the IV to vary over time. It is conceptually similar to approaches in economic history that use historical infrastructure as a source of quasi-exogenous variation in contemporary spatial patterns [24]. The rationale is that pre-reform telecommunication capacity captures deep-seated regional infrastructure endowments which shaped subsequent industrial agglomeration yet largely predates the emergence of modern agricultural industrial parks, digital agriculture and resilience-oriented policies. At the same time, we recognize that historical infrastructure may also proxy for persistent provincial institutional capacity that could in principle affect AER directly. To mitigate this concern, we control for a rich set of time-varying socio-economic factors and include province and year fixed effects, so that any remaining impact of the IV on AER is expected to operate mainly through its effect on agglomeration. In this sense the instrument is relevant and plausibly exogenous, and the IV estimates are interpreted as a robustness check that complements, rather than replaces, the baseline fixed-effects results.

2.1.5. Mediating Variable

Agricultural research intensity: Technological progress in agriculture is widely recognized as a key driver for enhancing the adaptability and stability of agricultural systems [12]. Drawing on this line of research, this paper adopts the ratio of agricultural research expenditure to GDP as a mediating variable. To this end, we establish a mediation framework to uncover the underlying mechanism through which LQ enhances AER.

2.1.6. Alternative Agricultural Economic Resilience Index

To further assess the robustness of the main conclusion, we constructed two alternative indicators: one is an entropy-weighted index based on a reduced set of indicators, and the other is a principal component analysis index based on the complete indicator system. These two alternative indicators share the same conceptual framework and data sources as the baseline index, ensuring comparability in interpretation and magnitude.

AER2: Mitigating Mechanical Correlation with the Mediator

To mitigate potential mechanical overlap between the outcome index and the mediator, we construct an alternative resilience index, AER2. Specifically, we exclude two technological innovation-related indicators from the original AER—Number of Digital Agriculture Enterprises and Number of Granted Agricultural Invention Patents. All other indicators, data sources, processing steps, and the aggregation framework remain identical to those used for AER, ensuring conceptual and scale comparability. We use AER2 in the mediation analysis to verify that the identified mechanisms are not artifacts of outcome-mediator construct overlap.

Principal Component Analysis-Based Agricultural Economic Resilience Index

As an alternative approach, we construct an agricultural economic resilience index using principal component analysis (PCA), based on the same indicator system presented in Table 1. All original indicators are retained, and this PCA-based measure is employed as an additional robustness check for our baseline result. Variables for which higher values indicate weaker resilience (e.g., disaster losses, chemical inputs) are first reverse-coded by multiplying by −1 to ensure that higher values consistently reflect stronger resilience. All indicators are then standardized. PCA is applied to the correlation matrix to extract orthogonal common factors, and the first few components that account for the majority of the total variance are retained. Component scores are aggregated using variance-based weights proportional to the corresponding eigenvalues. The resulting composite scores are linearly rescaled to the [0, 1] interval to generate AER_PCA, where higher values denote greater agricultural economic resilience. The statistical description of all variables is shown in

Table 2. Descriptive Statistics of Variables.

Variables	Samples	Mean	Standard Deviation	Min	Max
AER	651	0.6242	0.1147	0.3544	0.9976
AER2	651	0.3110	0.0867	0.1207	0.6223
AER_PCA	651	0.4128	0.1493	0	1
LQ	651	1.2240	0.7023	0.04213	4.2090
LQ2	651	1.0295	0.4240	0.0647	2.0128
I	651	5.0665	1.1561	0.4253	6.8739
Urban	651	0.5443	0.1522	0.1389	0.8958
Edu	651	8.8070	1.2587	3.739	12.6810
E	651	4.5373	0.3355	3.5690	5.3020
ISU	651	0.8904	0.0593	0.6578	0.9980
tele	651	20,449.4000	16,846.41	1101.65	92,552.25
R&D	651	0.0019	0.0010	0.0002	0.0058

.

2.2. Model Construction

Given the pronounced regional heterogeneity and temporal dynamics of AER, we employ a two-way fixed effects model (FE). This model effectively accounts for unobservable time-invariant regional characteristics (e.g., natural endowments) and nationwide time-varying factors (e.g., policy adjustments or climate fluctuations), thereby mitigating omitted variable bias. Accordingly, the following two-way fixed effects model is specified:

$${AER}_{i,t}=a+\beta {LQ}_{i,t}+\gamma C_{i,t}+{\mu }_i+{\lambda }_t+{\epsilon }_{i,t}$$

(2)

where α is the intercept of the model, i denotes the region and t denotes the year. AER_i,t denotes the level of agricultural economic resilience in each region. LQ_i,t denotes the agricultural industrial agglomeration level in each region. C_i,t represents a set of control variables including human capital level, urbanization rate, industrial structure upgrading index, economic development level, and logarithm of total social fixed asset investment. μ_i denotes individual fixed effects, which control for unobservable province-level factors such as geography, climate, and policies; λ_t represents time fixed effects, which capture common temporal trends and nationwide shocks; and ε_i,t is the random error term. To avoid multicollinearity that could bias the regression results, this study conducts a correlation test on the main variables. The results show no evidence of severe multicollinearity, indicating that all variables can be jointly included in the model for regression analysis.

2.3. Data Sources

The data used in this study are primarily drawn from authoritative sources such as the China Statistical Yearbook, China Rural Statistical Yearbook, and China Science and Technology Statistical Yearbook. The research sample covers 31 provincial-level regions in China (excluding Hong Kong, Macao, and Taiwan) over the period 2003–2023. For a small number of missing observations, reasonable estimations were made using methods such as the geometric mean growth rate and linear interpolation, in order to ensure continuity of the sample and completeness of the indicators. For example, in constructing the AER index system, the values for agricultural plastic film usage and pesticide usage in 2022–2023 are forecasted, as are government expenditures on agriculture, forestry, and water affairs and general public budget expenditures for 2023, as well as the number of active enterprises related to digital agriculture in 2023. All variables were standardized before being entered into the entropy method model to calculate the AER index. Since the forecast values are limited to only a few observations at the end of the sample period, their proportion is negligible and does not affect the robustness and reliability of the overall analysis. In the empirical analysis, both RStudio (version 4.4.2) and Stata (version SE 19.5) software were used for data processing, index construction, and econometric estimation.

3. Results

3.1. Basic Regression

To comprehensively assess the impact of LQ on AER, this paper employs both a two-way fixed effects model and a random effects model that controls for individual and time effects (i.e., FE and RE, respectively) for regression estimation. The regression results of these models are reported in

Table 3. Basic Regression Analysis Results.

Variables	FE	RE
LQ	0.0241 ***	0.0232 ***
	(0.0051)	(0.0041)
I	−0.0028	0.0042 ·
	(0.0026)	(0.0023)
Urban	0.0175	0.0207
	(0.0173)	(0.0171)
Edu	−0.0053	−0.0090 **
	(0.0040)	(0.0032)
E	0.0623 **	0.0424 *
	(0.0189)	(0.0176)
ISU	0.1160 ·	0.0833
	(0.0661)	(0.0626)
Intercept	-	0.3755 ***
	-	(0.0722)
Time Fixed	Yes	Yes
Individual Fixed	Yes	Yes
p-value	6.4027 × 10−11	1.1682 × 10−13

Note: *** denotes significance at the 0.1% level; ** denotes significance at the 1% level; * denotes significance at the 5% level; and · denotes significance at the 10% level; the values in parentheses are standard errors, the same applies hereafter. In FE specifications, the common intercept α is absorbed by the individual and year fixed effects and is therefore not reported; the same applies hereafter.

.

Firstly, the Hausman test (p-value = 2.293 × 10⁻⁵) significantly rejects the null hypothesis that the random effects model is appropriate relative to the fixed effects model. Therefore, the fixed effects specification better fits the sample characteristics, and the two-way fixed effects model is adopted as the baseline regression model, serving as the foundation for the subsequent empirical analysis.

Second, the overall F-test of the baseline regression is highly significant (p-value = 6.4027 × 10⁻¹¹). The coefficient of LQ in the FE model is 0.0241, significant at the 0.1% level, indicating that LQ exerts a significantly positive effect on AER. This finding is consistent with Chao & Li (2024) [15], who also document that agricultural production agglomeration enhances AER.

Concerning the control variables, E and ISU have positive and significant effects on AER, indicating that a stronger economic base and ongoing industrial upgrading support system stability. By contrast, Edu is insignificant in the FE model but negative and significant in the RE model, possibly reflecting regional educational structures or the out-migration of agricultural labor. Urban and I are not significant, which may relate to sample size or variable construction. Overall, the signs of most coefficients align with expectations, supporting the reliability of the estimates.

3.2. Robustness Tests

In this part, five sets of regression models are designed to conduct robustness tests, focusing on variable transformation, indicator substitution, dynamic effects, outlier treatment, and excluding observations with forecasted values.

First, in Model 1, we replace LQ with its logarithm, denoted LQ_log, to mitigate the influence of extreme values and to examine whether the elasticity interpretation delivers similar estimates. Second, Following the core-variable substitution approach in Ren (2025) [25], Model 2 replaces LQ with LQ2 to gauge the sensitivity of our estimates to how the core explanatory variable is constructed. Third, in Model 3, we replace LQ in (2) with its first-order lag term, denoted LQ_lag, to examine the dynamic effect of agricultural agglomeration on AER and to test the stability of the causal relationship over time. Model 4 applies a two-tailed 1% winsorization to the core explanatory variable, denoted LQ_w, to limit the influence of outliers and further assess the robustness of the estimates. Finally, to address potential bias from the interpolated values in 2023, Model 5 re-estimates the baseline regression after excluding all 2023 observations. Model 6 substitutes the dependent variable with the PCA-based resilience index, AER_PCA, while maintaining all other aspects of the specification unchanged. Results are reported in

Table 4. Robustness Test.

Variables	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6
LQ					0.0247 ***	0.02374 *
					(0.0046)	(0.0109)
LQlog	0.0215 ***
	(0.005)
LQ2		0.0111 ·
		(0.0062)
LQlag			0.0184 ***
			(0.0042)
LQw				0.0263 ***
				(0.0046)
Time Fixed	Yes	Yes	Yes	Yes	Yes	Yes
Individual Fixed	Yes	Yes	Yes	Yes	Yes	Yes
Control Variables	Yes	Yes	Yes	Yes	Yes	Yes
p-value	1.3084 × 10−8	1.5249 × 10−5	9.5348 × 10−8	4.3441 × 10−11	1.4136 × 10−10	6.0701 × 10−12

.

The robustness tests show that whether the core regressor is log-transformed, replaced by a re-estimated location quotient indicator, lagged by one period, subjected to two-sided 1% winsorization, estimated on the sample excluding 2023 observations, or paired with the alternative PCA-based resilience index AER_PCA as the dependent variable, the signs and significance levels remain highly consistent with those of the baseline model. Except for the specification using LQ2, which is significant at the 10% level, and the model using AER_PCA, which is significant at the 5% level, most estimates are statistically significant at the 0.1% level and exhibit magnitudes close to the baseline estimate. Overall, the consistent direction, statistical significance, and economic magnitude across all six specifications reinforce the robustness and reliability of the core finding that agricultural industrial agglomeration significantly enhances agricultural economic resilience.

3.3. Endogeneity Tests

To further address potential endogeneity arising from reverse causality and omitted variables, we implement an instrumental-variable (IV) strategy in which the 1984 number of landline telephones per 100 people in each provincial administrative region interacted with a continuous year variable (2003–2023) serves as an instrument for agricultural industrial agglomeration. This historical, time-varying IV exploits long-lasting differences in regional telecommunication infrastructure while allowing for temporal variation, in line with economic-history approaches that use historical infrastructure to isolate quasi-exogenous variation in contemporary spatial patterns. Because historical infrastructure may also proxy for persistent provincial institutional capacity that could in principle affect AER directly, the IV regressions include the same rich set of time-varying controls and province and year fixed effects as the baseline model, and the IV coefficients are interpreted as complementary robustness checks rather than as a replacement for the fixed-effects estimates.

Table 5. Endogeneity Test.

Variables	2SLS	LIML	GMM	GMM(CUE)	Placebo
LQ	0.0657 **	0.0657 **	0.0657 **	0.0657 **	−0.1321
	(0.0223)	(0.0223)	(0.0223)	(0.0223)	(0.2396)
Time Fixed	Yes	Yes	Yes	Yes	Yes
Individual Fixed	Yes	Yes	Yes	Yes	Yes
Control Variables	Yes	Yes	Yes	Yes	Yes
Cragg–Donald	24.5420	24.5420	24.5420	24.5420	0.5640
Stock–Yogo (10%)	16.3800	16.3800	16.3800	16.3800	16.3800
Anderson–Rubin	8.8500 **	8.8500 **	8.8500 **	8.8500 **	0.8400
Wu–Hausman	4.2070 **	4.2070 **	4.2070 **	4.2070 **	1.3580
N	651	651	651	651	651

reports five IV specifications that share the same controls as the baseline model together with region and year fixed effects; standard errors are computed as in the baseline. Column (2) reports two-stage least squares (2SLS) estimation using tele as the IV. Columns (3)–(5) re-estimate Column (2) with alternative estimators: limited-information maximum likelihood (LIML), two-step generalized method of moments (GMM), and continuously updated GMM (CUE), respectively. Because the model is just identified, point estimates are expected to be numerically similar across these estimators; the purpose is to demonstrate estimator insensitivity. Column (6) conducts a placebo test by shuffling the IVs order across observations and re-estimating 2SLS.

Across all IV estimators—2SLS, LIML, GMM, and GMM (CUE)—the coefficient on LQ remains positive and statistically significant under the same set of controls and two-way fixed effects. The first-stage results confirm strong instrument relevance, with Cragg–Donald F approximately equal to 24.54, far exceeding the Stock–Yogo 10% critical value (16.38), indicating that the IV is strong and the identification robust.

Furthermore, the placebo experiment—where the IV is randomly shuffled across provinces and years—yields an insignificant coefficient and a low first-stage F statistic approximately equal to 0.56, as expected. This falsification test confirms that the identification relies on the true historical-temporal structure of the IV rather than spurious correlation.

Taken together, these findings reinforce the robustness of the main conclusion: LQ significantly enhances AER [7,26]. Economically, the historical IV reflects pre-reform infrastructure endowments that shaped long-term regional connectivity and economic capacity. When scaled by the continuous year variable, it generates a gradual, exogenous source of variation in agglomeration intensity over time—satisfying both the relevance and exclusion restrictions of a valid instrument.

3.4. Heterogeneity Analysis

The considerable provincial disparities in China regarding natural resource bases, industrial composition, economic development, and policy frameworks posit that the influence of LQ on agricultural system stability and adaptability, together with the operative mechanisms, is not uniform but instead manifests distinct regional heterogeneities. Previous literature has indicated that technological progress, resource limitations, and environmental consequences vary dramatically by region. Zhao et al. (2022) [27] revealed that the disparities of regional green total factor productivity are mainly driven by the “trans-variation” part which is simultaneously influenced by structural and efficiency differences. The prospect of regional heterogeneity is further corroborated by Yang et al. (2025) [28], whose analysis reveals a starkly uneven spatial pattern of agricultural resilience at the provincial level in China. Their findings indicate not only a higher overall resilience in eastern coastal areas compared to the underdeveloped west but also significant spatial variations in the elasticity coefficients of the explanatory variables. More generally, regional economic resilience researchers have increasingly adopted multi-scale strategies to explore the diverse driving influences at various spatial scales, including the eastern-central-western pattern, urban agglomerations, and regional typologies hierarchies [29,30]. Furthermore, based on a system GMM model, Liang et al. (2025) [31] empirically tested urban agglomeration structures on different scales and discussed the variation in marginal effect in development stages and city-size groups.

Building on these insights, this paper conducts subgroup regressions across three dimensions: (i) eastern, central, and western regions; (ii) following the State Council’s designation, provinces are grouped into major grain-producing areas (Major), major grain-selling areas (Sell), and production-marketing balance areas (Balanced); and (iii) High vs. low investment, which is defined using the cross-sectional median of province-level full-period averages of total social fixed asset investment, yielding a time-invariant grouping. Such subgroup tests allow for the identification of potential heterogeneity in the effects of LQ. Figure 3 presents the map-based visualization of heterogeneity groups across Chinese provinces. The corresponding results are reported in

Table 6. Heterogeneity Analysis Results.

Variables	Eastern	Central	Western	Major (RE)	Sell	Balanced	High-Investment	Low-Investment
LQ	0.0448 **	0.0432	0.0221 ***	0.0071	0.0657 ***	0.0316 ***	0.0103	0.0296 ***
	(0.0160)	(0.0364)	(0.0008)	(0.0074)	(0.0157)	(0.0120)	(0.0073)	(0.0069)
Control Variables	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Intercept	-	-	-	0.3788 ***	-	-	-	-
				(0.1318)
Time Fixed	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Individual Fixed	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
p-value	5.0120 × 10−5	0.19135	4.8721 × 10−6	5.8354 × 10−7	2.8633 × 10−6	0.0005	1.1188 × 10−5	0.0002

Note: In the model specification tests, the Hausman test for the sample of major grain-producing areas yields a p-value of 0.7369, failing to reject the null hypothesis that the random effects model is consistent relative to the fixed effects model. Therefore, a two-way random effects model is adopted for this group. For the other subsamples, the Hausman tests significantly reject the random effects specification, and the two-way fixed effects model is ultimately selected as the estimation method.

.

The results of regression demonstrate the positive and significant effect of LQ on AER in the eastern and western regions. This means that the agglomeration effect contributed positively to the stability and adaptability of the system, in either the relatively developed eastern region or the underdeveloped western region. By comparison, the estimated coefficient is positive in the central region but statistically insignificant (the model’s overall F-test is also not significant), suggesting that the gains to agglomeration are not fully realized at the central region. This may be due to lagging industrial adjustment, inefficient distribution of resources, or differences in policy enforcement.

The effect of agricultural agglomeration on resilience is not statistically significant in major grain producing regions. This may be because agricultural production in these areas is more reliant on policy support and institutional arrangements, which could weaken the independent role of agglomeration. In contrast, for grain-selling and producing-selling balanced regions, the agglomeration effect is positive and significant at the 0.1% level, implying that in the case of no or offsetting policy bias, the market externalities and agglomeration-produced economies of scale can be more adequately exploited to enhance the region’s ability to resist external shocks [28].

Moreover, after considering a split by the level of total social fixed investment, the result further reveals that the positive effect of agglomeration on agricultural resilience is more significant in the “low investment” group. This suggests that in the context of financial constraints, the local economies and agricultural actors were more dependent on the agglomeration externalities (e.g., factor sharing, technology diffusion, and collaborative development) which significantly enhanced the marginal effect of agglomeration on systemic stability. By contrast, capital concentration is not significant in the high total investment sample, suggesting that an abundance of capital and supportive policies may undermine the importance of capital concentration mechanisms.

3.5. Mechanism Analysis

Advances in agricultural technology are generally considered to be important sources for the adaptability and stability of agricultural systems, while the uneven development of such technology is an important factor that determines the effect of industrial agglomeration on the resilience of agricultural economy. Existing studies provide strong evidence for this mechanism. For example, Wan et al. (2024) [32] revealed that technological innovation in agriculture has a significant positive influence on resilience, but with possible nonlinear and heterogeneous impacts among regions with disparate fiscal support levels. Ren et al. (2025) [25] also argues that economic and technological development and industrial structure upgrading are pivotal means to enhance AER. Based on these findings, the present study incorporates R&D as the mediating factor to directly capture how research investment impacts resilience. Investment in research and development can contribute to increase flexibility and stability in agricultural systems through the development of innovation in technology and management of production. Therefore, a fixed effects based three-stage regression model of “LQ → R&D→ AER” is developed to empirically test if LQ enhances AER indirectly through R&D.

Specifically, Model 1 presents the baseline regression, testing whether LQ has a significant overall effect on AER, consistent with (2). Model 2 examines whether LQ significantly affects the mediating variable R&D (see (3)). Model 3, with R&D included, investigates its effect on AER and whether the impact of LQ on AER is transmitted via the research investment channel, thereby identifying partial or full mediation of (4). Model 4 replaces R&D with its first-order lag term R&D_lag, which can reduce the synchronicity bias between R&D and contemporaneous measures of AER and to alleviate concerns about reverse causality. Model 5 replaces the core explanatory variable with LQ2, to test the robustness of the mechanism analysis to alternative measurement specifications. Finally, Model 6 replaces AER with AER2—which excludes mechanically related indicators—to mitigate mechanical overlap with R&D. Results are reported in

Table 7. Mechanism Analysis Results.

Variables	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6
LQ	0.0241 ***	0.0010 ***	0.0184 ***	0.0195 ***		0.0556 ***
	(0.0043)	(0.0001)	(0.0046)	(0.0048)		(0.0046)
LQ2					0.0144 *
					(0.0061)
R&D			5.8720 **		8.8957 ***	4.5406 **
			(1.7789)		(1.6734)	(1.7544)
R&Dlag				5.2400 **
				(1.8694)
Control Variables	Yes	Yes	Yes	Yes	Yes	Yes
Time Fixed	Yes	Yes	Yes	Yes	Yes	Yes
Individual Fixed	Yes	Yes	Yes	Yes	Yes	Yes
Sobel			0.0057 **	0.0050 **		0.0044 **
Aroian			0.0057 **	0.0050 **		0.0044 **
Goodman			0.0057 **	0.0050 **		0.0044 **
p-value	6.4027 × 10−11	2.22 × 10−16	1.4514 × 10−12	1.168 × 10−10	1.6189 × 10−10	2.22 × 10−16

.

$${RD}_{i,t}=a+\beta {LQ}_{i,t}+\gamma C_{i,t}+{\mu }_i+{\lambda }_t+{\epsilon }_{i,t}$$

(3)

$${AER}_{i,t}=a+\beta {LQ}_{i,t}+\gamma C_{i,t}+\delta {RD}_{i,t}+{\mu }_i+{\lambda }_t+v_{i,t}$$

(4)

The results of mediation analysis show that the effect of LQ on AER is not only significantly and positively direct, but the system adaptability and system stability can also be indirectly improved by the R&D channel. In particular, LQ exerts a strong positive effect on the mediator variable—R&D—which indicates that agglomeration leads to more funds devoted to research. With R&D added, it is still significantly positive but the coefficient of LQ decreases, so there is a partial mediating effect. The Sobel and Goodman tests confirm the significance of this indirect path at the 1% significance level, which is also suggested for interpretations of partial mediation. When R&D is replaced with its first-order lag term, the results remain consistent, which further confirms the robustness. Models 4 and 5 represent robustness checks using different specifications, and the results are consistent with the main findings. With AER2 (Model 6), the positive and significant associations of LQ and R&D with resilience persist, indicating that the mediation result is not driven by mechanical overlap and that our conclusions are robust. In short, the study reveals that investment in agricultural science and technology has acted as a key mediator in the way that industrial agglomeration affects the AER. Several studies, such as Wan et al. (2024) [32], have shown that system resistance and recovery enhancement of agricultural technological innovation. It is also backed up by other research in agglomeration and innovation: Lee (2009) [33] demonstrates that firms within industrial agglomeration have higher R&D intensity than non-agglomeration firms indicating agglomeration promotes greater research investment; Guo et al. (2023) [34] also reveal a significantly increasing trend of time pressures on firms in a cluster setting to continually invest in R&D to sustain their competitive advantage over time, thereby securing a continuous supply of technology advancements and inventive capabilities. Agricultural agglomeration fosters R&D investment through several interconnected channels. First, intensified competitive pressure within the cluster compels firms to innovate to maintain their market position. Second, knowledge spillovers and technological diffusion lower both the costs and risks associated with R&D, enhancing its expected returns. Finally, agglomeration facilitates collaborative innovation—easing access to pooled talent and resources—and attracts targeted policy support, such as R&D subsidies and tax incentives, which collectively underpin and encourage research activities.

To conclude, agricultural industrial agglomeration not only directly increases AER but also indirectly enhances system adaptability and recovery by spurring R&D investment, thereby providing solid theoretical and empirical foundations for cluster-based development strategies.

3.6. Nonlinear Relationship Analysis

Evolutionary economics views industrial agglomeration as a process in which the various stages of development are characterized by different forms. Thus, its effect on economic resilience is likely to be non-monotonic and to reflect the attributes of evolution in stages and multiple mechanisms through coevolution. At different phases of development, different degrees of agglomeration may trigger contrasting effects through resource pooling, technological spillovers, or factor misallocation. Such a double-edged sword impact may enhance the system’s resilience or weaken its adaptability. Therefore, empirical relevance should accomplish stage-based identification as well as refined characterization. As a subject of study, nonlinear mechanisms relating agglomeration to regional economic resilience are gaining growing interest among academics. Previous studies generally focused on the interaction between agglomeration and economic growth and revealed that its impacts fluctuated across stages. For instance, Zhang et al. (2023) [35] argues that agglomeration is diluted when over-concentrated, and the marginal effect of agglomeration on technology advancement is negatively related to the level of agglomeration—showing the nonlinearity of declining returns. Du et al. (2024) [36] discloses that the effect of manufacturing agglomeration on green total factor productivity demonstrates nonlinear heterogeneity across urban, geographical, and scale features. Wang (2024) [37] also shows that the relationship between secondary industry agglomeration and ecological/economic resilience: an inverted U-shaped turning point. Chao & Li [15] also find that agricultural production agglomeration has a nonlinear effect on AER. In general, studies in agricultural economics are still relatively sparse, and there are hardly any systematic examinations of potential multi-stage effect structures. The elements of agricultural systems, such as land and labor, are more foundational and exhibit distinct agglomeration patterns in both organizational and spatial structure compared to manufacturing. To systematically investigate these potential nonlinear effects on AER, this paper incorporates the second- and third-order terms of the agglomeration indices (LQ and LQ2) into the empirical model. This methodological approach seeks to expand the theoretical and empirical understanding of the agglomeration-resilience relationship, as captured by (5) and (6):

$${AER}_{i,t}=a+\beta {LQ}_{i,t}+\phi {{(LQ}_{i,t})}^2+\eta ({{LQ}_{i,t})}^3+\gamma C_{i,t}+{\mu }_i+{\lambda }_t+{\epsilon }_{i,t}$$

(5)

$${AER}_{i,t}=a+\beta {LQ2}_{i,t}+\phi {{(LQ2}_{i,t})}^2+\eta ({{LQ2}_{i,t})}^3+\gamma C_{i,t}+{\mu }_i+{\lambda }_t+{\epsilon }_{i,t}$$

(6)

As shown in (5), we estimate a cubic polynomial in LQ to allow for potential N-shaped nonlinearity. (6) keeps the same polynomial form and substitutes LQ with LQ2 to probe dimension-specific nonlinearities associated with employment-intensive agglomeration. Throughout, C_it denotes the set of control variables. We estimate two nonlinear fixed-effects specifications. In (5), the output-based LQ shows the configuration characteristic of an N-shaped relationship, but the quadratic and cubic terms are not statistically significant, so the evidence for nonlinearity is weak. In (6), replacing LQ with the employment-based LQ2 yields the same N-shaped configuration with significant higher-order terms, providing clear evidence of nonlinearity. Following the quantile-regression approach of Ren et al. (2022) [38], we examine how the nonlinear relationship varies across quantiles. We estimate panel quantile regressions using the Powell (2022) [39] estimator for panel quantiles with nonadditive fixed effects, keeping the same set of controls and including year fixed effects; results are reported in

Table 8. Nonlinearity Analysis.

Variables	Model 1	Model 2	0.25	0.5	0.75
LQ	0.02333 ***
	(0.0045)
LQ2	0.0027
	(0.0138)
LQ3	−0.0015
	(0.0064)
LQ2		0.1462 *	0.070 ***	0.3450 ***	0.3350 ***
		(0.0592)	(0.0090)	(0.0030)	(0.0023)
LQ22		−0.1046 *	0.061 ***	−0.2570 ***	−0.1310 ***
		(0.0517)	(0.0100)	(0.0260)	(0.0190)
LQ23		0.0246 ·	−0.0290 ***	0.0730 ***	0.0080 ·
		(0.0144)	(0.0030)	(0.0060)	(0.0040)
Control Variables	Yes	Yes	Yes	Yes	Yes
Time Fixed	Yes	Yes	Yes	Yes	Yes
Individual Fixed	Yes	Yes	Yes	Yes	Yes
p-value	6.9547 × 10−10	4.7171 × 10−6	2.22 × 10−16	2.22 × 10−16	2.22 × 10−16

.

Model 2 employs a nonlinear fixed-effects specification based on the employment-based agglomeration index LQ2, revealing a statistically significant N-shaped relationship: the coefficients of the linear, quadratic, and cubic terms are positive, negative, and positive, respectively, with all terms significant at least at the 5% level. The thresholds—derived from the first derivative of the cubic function with respect to LQ2—indicate that the marginal effect of agglomeration on AER turns negative when LQ2 lies between approximately 1.2439 and 1.5938, while remaining positive outside this range. This finding suggests that agricultural industrial agglomeration initially enhances resilience at low levels, exerts a dampening effect in the intermediate range—consistent with factor misallocation, congestion, or structural rigidities—and subsequently restores its positive influence at higher levels through mechanisms such as specialized division of labor, collaborative networks, and knowledge diffusion. Thus, the cubic specification captures a theoretically grounded PSR dynamic, rather than reflecting a purely data-driven empirical fit.

The panel quantile regression for (6) further uncovers substantial distributional heterogeneity around this N-shaped pattern. At the 0.50 and 0.75 quantiles, the coefficients of LQ2, LQ2², and LQ2³ consistently follow a positive–negative–positive sign pattern, fully aligning with the N-shaped profile identified in Model 2. In contrast, at the lower tail of the distribution (q = 0.25), the estimates for LQ2, LQ2², and LQ2³ are positive, positive, and negative, respectively—indicating that agglomeration yields net gains in resilience at low-to-moderate levels of industrial concentration, but these benefits diminish and eventually reverse as density increases. This behavior is consistent with congestion effects or capacity constraints commonly observed in low-resilience regions. Provinces at lower resilience quantiles are typically less developed or inland, characterized by limited market access, constrained fiscal and R&D capacity, and sluggish reallocation of labor and capital. While moderate agglomeration can generate efficiency improvements through input sharing and matching, rising industrial density amplifies structural frictions, leading to increased congestion, heightened production pressures, and greater vulnerability to external shocks. By contrast, provinces at the median and upper quantiles generally possess more advanced infrastructure, deeper financial markets, and more integrated agro-industrial value chains. These institutional advantages enhance cross-sectoral flexibility and value-chain adaptability, enabling firms and workers to reconfigure resources efficiently. After a mid-range slowdown in marginal returns, these regions can reactivate specialization economies, collaborative innovation, and knowledge spillovers at higher levels of LQ2, thereby restoring the positive impact on resilience. This differentiated response is supported by a growing body of empirical evidence on agglomeration and economic resilience. A meta-analysis of agglomeration economies in developing countries finds that while the average returns to agglomeration are positive, they exhibit substantial heterogeneity across regions and are highly sensitive to local institutional environments and structural conditions—collectively pointing to inherent nonlinearities and threshold effects in the returns to density [40]. At the urban level, Jiang et al. (2022) show that population agglomeration in China enhances urban economic resilience and generates spatial spillovers to neighboring cities, with the magnitude of these effects varying significantly according to levels of human capital and labor force composition [41]. At the regional scale, Zheng et al. (2023) demonstrate that the coupling between industrial agglomeration and regional economic resilience displays pronounced spatiotemporal heterogeneity and is shaped by disparities in development stages and underlying structural foundations [42]. Taken together, these findings provide a coherent explanation for why provinces at lower resilience quantiles are more prone to entering a “late-stage congestion zone” as density increases, whereas those with stronger socioeconomic and institutional capacities are better equipped to overcome this phase and realize the “third leg” of the N-shaped relationship identified in the median and upper-quantile estimates.

Figure 4 displays the quantile-specific average marginal effects (AME) of LQ2 on AER at q = 0.25 and q = 0.50 (the 0.75 quantile curve is omitted as it closely tracks the median within the observed range, to avoid visual clutter). The 0.25 quantile curve rises initially when LQ2 is in the low-to-moderate range and then declines markedly at higher agglomeration levels, exhibiting a hump-shaped pattern consistent with congestion or input mismatch in low-resilience provinces during high-density phases. In contrast, the 0.50 quantile curve follows a U-shaped trajectory: marginal gains weaken in the middle range of LQ2 but rebound significantly at higher levels of agglomeration, indicating that regions with stronger underlying capacities are better able to leverage mechanisms such as specialized division of labor, collaborative innovation, and knowledge diffusion under advanced agglomeration conditions. The shaded bands represent 95% confidence intervals for each estimated curve. Overall, agricultural industrial agglomeration tends to enhance resilience, yet its incremental returns are strongly state-dependent—smallest in the medium agglomeration range and most fragile among provinces with lower resilience.

The comparison across different agglomeration measures shows that, in the linear two-way fixed-effects specifications, the baseline results are broadly consistent whether we use the traditional output-based location quotient or the employment-based. However, once we introduce quadratic and cubic terms, the shape of the nonlinear relationship becomes more sensitive to how agglomeration is measured. When higher-order terms in LQ are added, their coefficients are small and statistically unstable, so that no clear N-shaped profile can be identified. By contrast, LQ2 more accurately captures the intra-industry allocation of agricultural labor and the breadth of the agro-industrial base, and in the nonlinear specifications it delivers a more stable and precisely estimated N-shaped pattern. We therefore rely on LQ2 as the main indicator for tracing nonlinear and distributional effects, while treating the LQ-based estimates as a robustness check on the average impact of agglomeration on resilience. Figure 5 illustrates the N-shaped effect of LQ2: gray dots represent province-level observations after controlling for other covariates, the black solid line shows the fitted curve from the cubic fixed-effects model, and two vertical dashed lines mark the estimated turning points (LQ2 ≈ 1.2439 and LQ2 ≈ 1.5938). The marginal effect is positive at low levels of agglomeration, turns negative in the intermediate range, and becomes positive again at high levels—closely aligning with the “rise–squeeze–reconfiguration” three-stage mechanism proposed earlier. Overall, this N-shaped pattern complements the inverted U-shaped relationship identified by Wang (2024) [37] in a study of Yangtze River Delta cities using a dynamic spatial Durbin model, collectively demonstrating that the effect of agglomeration on resilience exhibits pronounced nonlinearity and stage-specific dynamics across varying spatial scales and measurement approaches.

In conclusion, this study demonstrates that the impact of agricultural industrial agglomeration on agricultural economic resilience is characterized by pronounced nonlinearity and is sensitive to how agglomeration is measured. In the linear two-way fixed-effects specifications, both the traditional output-based location quotient and the employment-based location quotient yield broadly consistent evidence that moderate agglomeration enhances resilience. However, the employment-based location quotient most clearly captures the N-shaped relationship in the nonlinear and distributional analyses, aligning more closely with theoretical expectations regarding factor allocation dynamics and structural transformation, and exhibiting greater empirical robustness when identifying turning points. Accordingly, we use the employment-based index as the main indicator for tracing stage-specific dynamics, while the output-based location quotient serves both as an important benchmark that ensures comparability with the existing literature and as a supplementary check on the consistency of our results. These findings underscore the importance of deliberate and theoretically informed selection of agglomeration measures in empirical analysis and provide a solid foundation for advancing research on the “phased and conditional” mechanisms underlying the agglomeration–resilience nexus.

3.7. Extended Analysis

Geographically, China’s agricultural economic activities are characterized by pronounced regional interconnections and patterns of factor mobility. The level of agricultural economic resilience does not evolve in isolation within individual provinces but is shaped through spatial interactions driven by technology diffusion, policy emulation, and industrial relocation. A growing body of recent literature has adopted a spatial analytical framework to systematically examine this phenomenon. For instance, Luo et al. (2024), after constructing a comprehensive AER index and applying Moran’s I statistic along with the spatial Durbin model, identified significant positive spatial autocorrelation and notable spillover effects across Chinese provinces [43]. Chen et al. (2025), using provincial panel data, further demonstrated that determinants such as market size, agricultural financial investment, and urbanization exert both significant direct impacts on local AER and substantial indirect effects on neighboring regions [4]. Likewise, Li et al. (2024), based on county-level data and employing kernel density estimation and spatial conditional distribution analysis, confirmed the presence of persistent positive spatial dependence in agricultural economic resilience at the sub-provincial level [44]. Together, these findings underscore that agricultural resilience and its associated indicators are not independently distributed but are embedded within a tightly interconnected regional network. In contrast, much of the existing literature remains concentrated on single-dimensional measures of resilience or green productivity, with limited comparative assessment of how different methodological specifications influence the observed spatial agglomeration and potential spillover patterns of AER. To address this gap, this study extends its analysis by constructing a spatial weight matrix based on provincial contiguity to compute the global Moran’s I for three alternative AER measures—AER, AER2, and AER_PCA—thereby mapping the spatial clustering patterns and evolutionary trajectories of agricultural economic resilience across Chinese provinces from 2003 to 2023. This approach provides more systematic empirical evidence for understanding the mechanisms of cross-regional linkage and spatial diffusion in agricultural resilience.

3.7.1. Spatial Weight Matrix

The spatial weight matrix W is a core element of spatial econometric models, whose role is to depict the strength of spatial connections among different regions. In line with the characteristics of the research question, this paper constructs an adjacency matrix based on the bordering relationship of provincial administrative regions, reflecting the direct geographical adjacency relationship among regions. This matrix is in 0-1 form, that is, 1 if the regions border each other and 0 if they do not. The form of the spatial weight matrix is shown in (7):

$$W=\left[\begin{array}{ccc}\begin{array}{c}{w}_{11}\\ w_{21}\end{array}& \begin{array}{c}{w}_{12}\\ w_{22}\end{array}\dots & \begin{array}{c}{w}_{1n}\\ w_{2n}\end{array}\\ ⋮& \ddots & ⋮\\ w_{n1}& \dots & w_{nn}\end{array}\right]$$

(7)

where W is an n × n square matrix. The elements w_ij of the matrix represent the correlation between region i and region j, while the diagonal elements are usually set to zero (i.e., w_ii = 0).

3.7.2. Spatial Autocorrelation Analysis

The Global Moran’s I is the most widely used statistic for testing spatial autocorrelation, employed to assess whether a given variable exhibits a clustered, dispersed, or random spatial distribution across the entire study region. Its formal definition is shown in (8):

$$I={\displaystyle \frac{n}{{\sum }_{i=1}^n{\sum }_{j=1}^{n}wij}} \times {\displaystyle \frac{{\sum }_{i=1}^n{\sum }_{j=1}^{n}wij(x_i-\overline{x})(x_j-\overline{x})}{{{\sum }_i\left(x_i-\overline{x}\right)}^2}}$$

(8)

where n denotes the number of observations, w_ij represents an element of the spatial weight matrix, and x_i and x_j denote the observed values for regions i and j, respectively, while $\overline{x}$ stands for the overall sample mean. Global Moran’s I is a classical measure for assessing global spatial autocorrelation, with its value typically falling within the interval [−1, 1]. A statistically significant positive I value indicates spatial clustering of similar high or low values, reflecting positive spatial dependence. Conversely, a significant negative I value suggests a spatial pattern in which high values are surrounded by low values and vice versa—commonly referred to as spatial dispersion or “heterogeneous adjacency.” When I is close to zero and statistically insignificant, the spatial distribution of the variable is considered approximately random, implying no discernible spatial pattern. In this paper, we compute Global Moran’s I for three alternative measures of agricultural economic resilience (AER, AER2, and AER_PCA). The annual Moran’s I statistics are reported in

Table 9. Spatial Autocorrelation Test.

Year	Moran’s I
	AER	AER2	AER_PCA
2003	0.0482	0.2260 *	0.3678 ***
2004	0.0849	0.2211 *	0.4302 ***
2005	0.0910	0.2314 *	0.4059 ***
2006	0.0833	0.2422 **	0.4184 ***
2007	−0.0330	0.1825 *	0.4590 ***
2008	0.0014	0.2016 *	0.4264 ***
2009	0.0619	0.1980 *	0.4595 ***
2010	0.0918	0.2336 *	0.4519 ***
2011	0.0771	0.2215 *	0.4381 ***
2012	0.0660	0.2417 **	0.4330 ***
2013	0.1565	0.2046 *	0.4028 ***
2014	0.1403	0.1955 *	0.3711 ***
2015	0.1603 *	0.2106 *	0.3604 ***
2016	0.2240 *	0.2492 **	0.3830 ***
2017	0.2419 **	0.2211 *	0.2061 ***
2018	0.1563	0.1939 *	0.4054 ***
2019	0.0500	0.2089 *	0.4105 ***
2020	0.1172	0.2281 *	0.4297 ***
2021	0.1756 *	0.2219 *	0.4442 ***
2022	0.1756 *	0.2320 *	0.4409 ***
2023	0.1348	0.2202 *	0.4009 ***

Note: Due to space limitations, Table 9 does not present standard errors but only provides significance symbols; otherwise, the table would be overly long.

, and their temporal evolution is illustrated in Figure 6.

A synthesis of Table 9 and Figure 6 reveals that the three alternative measures of agricultural economic resilience generally exhibit a spatial pattern dominated by positive spatial autocorrelation at the provincial level. However, notable differences exist in the strength and statistical significance of spatial agglomeration across these indicators. Specifically, while the Moran’s I for the original AER index remained mostly positive from 2003 to 2023, it fluctuated near the 10% significance threshold in most years and failed to achieve statistical significance in several early periods. In contrast, both AER2 and AER_PCA yielded consistently positive and statistically significant Moran’s I value throughout the entire period. AER2 remained stable within the range of 0.19–0.24, whereas AER_PCA persistently exhibited higher values between 0.37 and 0.45. A key factor underlying this divergence in spatial significance lies in the reconfiguration of indicator composition: AER2 was recalculated using the original entropy weighting framework after excluding two variables—“Number of Digital Agriculture Enterprises” and “Number of Granted Agricultural Invention Patents.” The former displays pronounced “head clustering,” with extreme concentration in a few economically advanced provinces, while most central, western, and traditional agricultural provinces register near-zero values. The latter exhibits high temporal volatility and sharp spikes in specific provinces, making it sensitive to changes in data reporting practices and short-term policy incentives. Although these two indicators possess high temporal dispersion—leading the entropy method to assign them substantial weights—their spatial distributions are highly skewed, resembling a pattern of “a few extreme outliers surrounded by extensive low-value regions.” Rather than aligning with the broader east-central-west developmental gradient, they introduce spatial noise that obscures the dominant regional structure. Retaining them in the AER index amplifies idiosyncratic fluctuations in certain provinces during specific years, diluting the shared component of regionally coordinated resilience dynamics. This weakens the spatial covariance term in Moran’s I, reducing its magnitude and preventing statistical significance in earlier years. By removing these indicators in AER2, the entropy-derived weights are reallocated to more spatially coherent variables reflecting output capacity, ecological conditions, mechanization levels, disaster resilience, fiscal support, and resource-environmental pressures. As a result, systematic inter-provincial disparities and spatial clusters—such as “high-high” and “low-low” groupings—are more clearly captured, yielding a purer signal of spatial dependence. Consequently, AER2 demonstrates a robust, uniformly significant positive spatial autocorrelation over time, with a markedly higher overall Moran’s I compared to the original AER.

Overall, our findings align with recent empirical evidence on agricultural economic resilience in China, which consistently reports significant positive spatial autocorrelation and well-defined “high–high” and “low–low” spatial clusters at both provincial and county levels (Chen et al., 2025; Luo et al., 2024; Li et al., 2024) [4,43,44]. In contrast to these prior studies—most of which rely on a single entropy-weighted AER index—this paper systematically evaluates three alternative measurement specifications: the full-system AER, the reconstructed AER2, and the AER_PCA. The results demonstrate that while spatial autocorrelation remains predominantly positive from 2003 to 2023 across all measures, the magnitude and statistical significance of spatial clustering vary considerably depending on the index construction approach. Specifically, the spatial pattern is weakest and least statistically significant under the original AER specification, whereas it becomes markedly stronger and uniformly significant for both AER2 and AER_PCA. This not only reinforces the understanding that agricultural economic resilience is characterized by stable regional interdependence and spatial spillover effects but also underscores the critical role of index design in spatial analysis. Methodological choices about indicator selection and weighting can affect the identified degree of spatial dependence and the estimated spillover effects. Future spatial econometric models and policy simulations should take these measurement decisions into account.

4. Discussion

This study investigates the relationship between LQ and AER from an evolutionary perspective, with a particular emphasis on understanding “why this is so.” The overall evidence suggests a three-stage mechanism chain: at low levels of agglomeration, processes such as input sharing, labor and knowledge exchange, matching, and learning are predominant; resilience increases alongside agglomeration. As agglomeration moves into the medium range, coordination frictions, mismatches, and resource or environmental constraints become more pronounced, leading to diminished marginal returns and, if left unmanaged, to localized environmental stress and widening disparities in resilience across regions. When agglomeration deepens further, specialized divisions of labor together with denser upstream–downstream collaboration and knowledge platforms emerge. This reactivates the system’s capacity for adaptation and restructuring and strengthens the positive effects. Such dynamics are consistent with functional evolution within the pressure–state–response framework.

Regarding distributional heterogeneity, panel quantile regression analysis reveals that regions situated in the lower segment of the resilience distribution are more likely to experience congestion and mismatch in the middle section. By contrast, the median and upper segments exhibit a pattern of “initial gains—pressures in the middle—recovery at higher sections,” indicating that once foundational capabilities and collaborative networks reach a certain maturity, high agglomeration can again enable regions to benefit from knowledge sharing and collaboration. Average effects therefore cannot be interpreted in a one-size-fits-all manner; state dependence calls for a hierarchical reading of the results.

In terms of region and type, the positive effects in the east and west are relatively stable, whereas those in the central region are insignificant, which may reflect industrial adjustment, factor frictions, or mismatches in policy implementation stages. In major grain-producing areas, the effect is also insignificant, plausibly because policy goals and institutional arrangements partially offset or mask market externalities. Major grain-selling areas and production–marketing balance areas are more likely to benefit from economies of scale and network synergies and thus exhibit stronger resistance and recovery capacity. Overall, the regional economic context and stage of development shape the returns to an additional unit of industrial agglomeration. At the same time, the global Moran’s I result shows that provincial AER is predominantly characterized by positive spatial autocorrelation, with high- and low-resilience provinces forming distinct spatial clusters. This pattern is robust across different resilience indices but is markedly stronger and more uniformly significant for the reconstructed AER2 and the PCA-based AER_PCA, underscoring that the spatial manifestation of resilience is sensitive to how the index is constructed.

Mechanism identification further shows that research and development (R&D) acts as a key mediator: agglomeration raises research intensity, and including lagged R&D attenuates—but does not eliminate—the direct effect of agglomeration on resilience, consistent with partial mediation. Cross-validating the mediation channel with AER2, which is specifically recalculated after excluding two innovation-related indicators to mitigate mechanical correlation with the R&D index, yields very similar estimates. This alleviates concerns that the identified R&D channel is driven by mechanical overlap in indicator construction rather than by genuine behavioral responses.

In terms of limitations and future research, cross-provincial panels are inherently limited in their ability to capture the adjustment paths of micro agents. It is therefore important to link firm- and household-level data, enabling event studies and distributed-lag identification around policy pilots or market shocks. Given the presence of positive spatial autocorrelation documented in this paper, future work should move beyond descriptive Moran’s I and explicitly employ spatial econometric models—such as the spatial Durbin model—to decompose direct, indirect, and total effects of agglomeration on resilience. Nonlinearities could also be revisited using splines, generalized additive models, or semi-parametric threshold methods. On the measurement side, future research may further extend the “pressure” dimension to incorporate climate extremes and supply chain disruptions and triangulate the results with alternative agglomeration indicators such as the Ellison–Glaeser index, Herfindahl–Hirschman index, spatial Gini coefficient, and input–output network centrality. These extensions would deepen our understanding of the causal chain linking “agglomeration–capacity–resilience” and clarify how spatial interactions and measurement choices jointly shape the observed patterns.

5. Conclusions

5.1. Research Conclusions

Using panel data for 31 Chinese provinces over 2003–2023, we find that agricultural industrial agglomeration—measured by the location quotient—significantly raises agricultural economic resilience. The result is robust to extensive checks and strategies addressing potential endogeneity. Five features qualify this average effect: (i) pronounced regional heterogeneity linked to geography, industrial base, and development stage; (ii) an N-shaped relationship captured by the employment-based LQ2, with two empirically relevant thresholds (LQ2_i,t ∈ [1.2439, 1.5938]); (iii) state dependence across the conditional distribution of resilience in panel quantile regressions; (iv) partial mediation through agricultural research intensity, indicating that agglomeration enhances resilience in part by deepening innovation capacity; and (v) A consistently positive pattern of spatial autocorrelation is observed across all three resilience indices, with AER2 and AER_PCA exhibiting statistically significant spatial clustering in every year, while the original AER index displays weaker and intermittently significant spatial dependence—collectively indicating non-negligible interprovincial spillovers in agricultural economic resilience.

5.2. Policy Conclusions

Based on the above findings, we propose several targeted policy implications:

(1) Strategically promote agricultural agglomeration to harness agglomeration economies. It is essential to leverage the advantages of agglomeration in resource integration, large-scale production, and collaboration. By building specialized division-of-labor systems, extending industrial chains, and optimizing market layouts, the overall efficiency and risk resistance of agricultural production can be improved. Particularly under the frequent occurrence of external shocks, properly guiding the development of agricultural agglomeration can help reinforce the long-term stability and resilience of agricultural systems.

(2) Implement region-specific and differentiated development strategies. Due to the regional differences in resource endowment, industrial structure, market scale, and development stage, agglomeration paths as well as agglomeration effects vary. Therefore, policy should be local-specific. For instance, resource-rich regions can emphasize the agglomeration of high-end factors and technology-led development while under-developed areas may be based on characteristic industries and comparative advantage to gradually cultivate agglomeration effects. This differentiated approach can avoid one-size-fits-all policies and make agglomeration more effective in enhancing regional resilience.

(3) Adopt stage-contingent governance strategies that are closely aligned with the nonlinear agglomeration–resilience relationship. Building on the estimated thresholds at LQ2 ≈ 1.24 and 1.59 and the quantile-specific responses, governance should follow a concrete, stage-contingent strategy. When LQ2 is below approximately 1.24—particularly in lower-resilience inland provinces with limited market access and constrained factor mobility—policy should prioritize backbone infrastructure and entry facilitation by expanding transport and logistics corridors, easing access to working-capital credit for small and medium-sized agro-processing enterprises, and scaling up extension services and basic digital infrastructure to foster upstream–downstream linkages without triggering premature congestion. In the intermediate range of 1.24–1.59, where our estimates indicate a negative marginal effect, the focus should shift toward mitigating congestion and resource misallocation: reducing interprovincial barriers to labor and capital mobility (e.g., via coordinated labor-matching platforms and mutual recognition of qualifications), redirecting subsidies and fiscal transfers toward R&D and shared service platforms, and tightening scrutiny of land-intensive, low-value-added expansions and associated environmental and safety risks. Once LQ2 exceeds approximately 1.59—typically in provinces near or above the median and upper resilience quantiles—policy should deepen specialization and manage externalities by supporting collaborative R&D consortia, building digital knowledge-sharing and supply chain coordination platforms, and guiding firms toward higher value-added upgrading to capture network externalities. At the same time, regular monitoring of LQ2 and the AER sub-indices (resistance, recovery, adaptability) should be built into performance evaluation and early-warning systems to keep agglomeration within sustainable limits and to address emerging environmental and congestion pressures in a timely manner.

(4) Amplify the mediating role of research investment. Research investment has been confirmed as a key channel through which agricultural agglomeration enhances resilience. The design of policies should focus on building innovation systems and distributing research factors, and promote interaction and knowledge diffusion among research organizations, leading firms and farmers. Promoting the transfer and application of the fruits of research and development could strengthen the long-term innovation capability of agriculture and thus contribute to the enhancement of the ability to recover when agricultural systems face external shocks.

(5) Coordinate policy at the regional scale to internalize spatial spillovers. Evidence of positive spatial autocorrelation and non-negligible interprovincial spillovers in agricultural economic resilience demonstrates that policy coordination should operate at the regional level rather than be confined to individual provinces. High-resilience, high-agglomeration provinces can serve as “growth and resilience hubs” through cross-provincial joint planning of agricultural industrial parks, shared R&D and extension platforms, and coordinated disaster preparedness and recovery mechanisms. This approach facilitates the more effective diffusion of agglomeration benefits to neighboring low-resilience regions while simultaneously preventing the displacement of congestion and environmental pressures across provincial borders.