Will Digital Finance Reduce Agricultural Total Factor Productivity? Evidence from China

Abstract

Using a city-level panel for China (2011–2021), this paper estimates agricultural total factor productivity (TFP) with a stochastic-frontier approach and identifies the effect of digital finance through two-way fixed effects and instrumental-variable strategies. We document a statistically and economically significant negative association: a 1% increase in the digital finance index is linked to a decline of 1.5 in agricultural TFP. Evidence points to capital misallocation as the dominant channel, with the adverse effect most pronounced where agricultural capital markets are highly distorted. Heterogeneity analyses show stronger negative impacts in labor-intensive areas, non-major grain regions, and small-scale farming systems. Results are robust across alternative specifications and IV estimations. By moving from provincial aggregates to city-level variation, this study sharpens identification and uncovers within-province patterns that are invisible in coarser data. The findings highlight an important unintended consequence of digital financial expansion for agriculture and underscore a policy priority: improving the allocation and targeting of digital credit within rural economies to support productivity and sustainable development.

1. Introduction

Agricultural total factor productivity (TFP) is a key driver of economic development and a cornerstone of sustainable growth. Enhancing agricultural TFP not only improves food security by increasing output without proportional input growth but also releases labor and capital for industrial and service sectors, thereby facilitating structural transformation [1]. Sustained gains in agricultural productivity further contribute to poverty reduction and rural development, alleviate pressure on natural resources, and support more environmentally sustainable production systems [2,3]. As the global economy faces the twin challenges of feeding a growing population and mitigating climate change, raising agricultural TFP represents a crucial pathway to balancing economic growth with ecological sustainability.

Finance plays an increasingly important role in advancing agricultural productivity. Access to financial services enables farmers to adopt modern technologies, invest in higher-quality inputs, and manage production risks more effectively. Well-functioning credit and financial markets are especially critical in rural areas, where liquidity constraints often prevent smallholders from undertaking productivity-enhancing investments [4,5]. By easing these constraints, finance fosters innovation diffusion, mechanization, and more efficient resource allocation, thereby promoting long-term productivity growth [6]. Moreover, credit and insurance access enhance farmers’ resilience to shocks such as climate variability, helping stabilize productivity and support sustainable agricultural development [7]. Strengthening rural financial systems is thus essential not only for fostering agricultural TFP but also for promoting inclusive and sustainable economic growth.

The rapid advancement of digital technologies has reshaped financial systems worldwide, giving rise to digital finance services. By lowering transaction costs, broadening outreach, and improving the efficiency of credit allocation, digital finance has transformed access to capital in both urban and rural economies [8,9]. In agriculture, Internet-based services help overcome traditional barriers such as insufficient collateral, high monitoring costs, and geographic isolation. Through mobile banking, digital credit platforms, and fintech-enabled risk management tools, smallholders can more readily invest in modern technologies, stabilize consumption, and mitigate income volatility, thereby improving agricultural TFP [10,11]. However, these benefits are not guaranteed. Because agricultural borrowers often lack stable digital records and face higher income volatility, digital finance may disproportionately favor non-agricultural sectors where credit risks are easier to evaluate. This bias risks diverting capital away from agriculture, potentially reinforcing capital misallocation and undermining agricultural productivity. Thus, the diffusion of digital finance represents a double-edged channel through which digital transformation interacts with agricultural productivity and sustainable development.

Using city-level panel data for China from 2011 to 2021, we estimate agricultural total factor productivity (TFP) through a stochastic frontier analysis framework and examine its relationship with the expansion of digital finance. The baseline results consistently show that digital finance exerts a statistically significant negative effect on agricultural TFP: a 1% increase in the digital finance index is associated with an approximate decline of 1.5 in productivity.

Further analysis shows that the adverse effect operates mainly through capital misallocation and is amplified in areas with severe capital market distortions. The impact is particularly strong in labor-intensive, non-major grain-producing, and small-scale farming regions, where sensitivity to capital shortages and limited access to formal finance further intensify the negative consequences.

2. Related Literature

This study contributes to the growing literature on financial technology, which has largely focused on its modes of operation, associated risks, and economic impacts [12,13,14]. Enabled by big data, cloud computing, and artificial intelligence, fintech reduces information asymmetries in financial markets, broadens access to credit for small and medium-sized enterprises and households, and alleviates liquidity constraints. In doing so, it extends the reach of financial services and fosters inclusive finance [13,14,15]. Beyond financial inclusion, fintech as a form of technological change has been shown to enhance urban innovation, improve enterprise productivity, and stimulate household consumption and income growth through expanded credit availability and more efficient production processes [12,14,16,17,18].

However, the effects of fintech and digital technologies are not uniformly positive. Ref. [15] finds that the development of financial technology can suppress the total factor productivity of listed companies in regions with limited financial resources. Similarly, the diffusion of automation technologies, including digital tools and artificial intelligence, has been associated with declines in labor’s income share and overall employment [12]. The digital divide also risks exacerbating inequality: in the short run, it can widen educational disparities, while only in the longer run may fintech help ease credit constraints for rural households and narrow the urban–rural income gap [19]. Moreover, fintech institutions often channel credit primarily to borrowers with stronger reputations, higher education, and stable, well-paid jobs, making their clientele little different from that of traditional financial institutions [20]. Research further suggests that the digital economy may intensify sectoral income disparities. For example, the development of e-commerce may fail to generate sustained income growth for farmers [21], while digital inclusive finance can reduce overall agricultural output [22].

Recent agriculture-focused studies connect digital finance to productivity and investment on the farm side rather than to general fintech outcomes. At the aggregate level, Chinese evidence shows that digital inclusive finance is associated with higher green agricultural TFP [23] and with improvements in overall agricultural TFP [24]. At the technology margin, county-level analysis indicates that digital financial inclusion promotes mechanization and service uptake, pointing to concrete investment channels [25]. Complementing these correlational patterns, experimental evidence from smallholder settings demonstrates that mobile-money access can raise savings and farm investment, reinforcing the plausibility of finance-to-productivity linkages [26]. Together, this body of work highlights investment, technology adoption, and risk-management pathways through which digital finance may shape agricultural productivity—contextualizing and motivating our city-level TFP analysis.

Our work relates to research on resource allocation and agricultural development, which documents persistent inefficiencies in rural funding due to limited policies, urban-biased fiscal systems, and imperfect factor markets—factors that create distortions between agricultural and non-agricultural sectors in China [27,28,29]. This paper makes two key contributions. First, whereas prior studies have examined the Matthew effect of digital finance on human capital allocation, income inequality, and enterprise management [15,19,20,21], we focus on its impact on agricultural TFP. Second, in contrast to work based on provincial-level data, we estimate agricultural TFP using city-level data, allowing for more precise identification. Broadly speaking, our paper is also connected to the literature on financial friction and misallocation. Following the seminal contributions of [30,31], subsequent work has studied how information frictions and idiosyncratic uncertainty affect resource misallocation and aggregate efficiency [32,33,34,35,36,37,38].

3. Theoretical Analysis and Hypotheses

The digital economy is fundamentally transforming economic structures worldwide by integrating information technologies into traditional sectors and reshaping capital allocation and production processes. Financial technology, particularly Internet-based finance, has emerged as a key driver of this transformation. In agriculture, digital financial services can alleviate liquidity constraints, facilitate the adoption of modern inputs and technologies, and improve resource efficiency, thereby influencing total factor productivity (TFP) [16,21,39]. However, the uneven diffusion of digital tools across regions and sectors may generate heterogeneous outcomes, potentially constraining agricultural TFP. Understanding the impact of digital finance on agricultural productivity is therefore critical for both policymakers and researchers. This paper examines the role of digital finance in shaping agricultural TFP in China, assessing its effects in the context of a rapidly digitalizing economy.

The impact of digital technologies on agricultural TFP can be conceptualized through two opposing mechanisms. On one hand, digital innovations may increase TFP by enhancing resource allocation efficiency and lowering transaction costs. Digital financial platforms can ease credit constraints for rural households, facilitating timely investment in machinery, high-quality seeds, and advanced agronomic practices. Moreover, digital tools improve access to information, reduce production uncertainties, and strengthen risk management through enhanced weather forecasting and insurance services, collectively fostering technological progress and productivity growth in the agricultural sector.

Conversely, as digital technologies advance, Internet financial intermediaries are more effective at reducing information asymmetries in non-agricultural sectors, where borrowers generate abundant digital footprints through e-commerce, online payments, and formal employment records. By contrast, agricultural borrowers often lack reliable digital records and face volatile, seasonal incomes, making credit risk assessment more difficult. As a result, digital finance disproportionately favors non-agricultural sectors, diverting attention and resources away from agriculture. Smallholder farmers—who make up a large share of the rural population in China—are particularly disadvantaged, as limited scale, technical skills, and infrastructure hinder their adoption of digital tools, further widening productivity gaps. Overall, the expansion of digital technologies in rural areas risks reinforcing capital misallocation, thereby weakening agricultural investment and reducing TFP.

In summary, the rise of digital finance is transforming capital allocation and production, with key implications for agriculture. It can ease liquidity constraints, improve access to technology, and enhance risk management, boosting agricultural total factor productivity. However, uneven adoption and challenges in assessing creditworthiness may divert attention and finance to non-agricultural sectors, potentially limiting TFP gains. Based on these arguments, we propose two hypotheses:

Hypothesis 1a.

Digital finance improves agricultural TFP.

Hypothesis 1b.

Digital finance reduces agricultural TFP.

4. Empirical Analysis

This section presents the empirical analysis in four parts. We assemble a balanced city–year panel covering Chinese prefecture-level cities over 2011–2021 (exceeding 280 cities), merging agricultural production accounts, input endowments, and meteorological controls with the Peking University Digital Financial Inclusion Index at the city level. Agricultural total factor productivity is estimated via a stochastic frontier approach, producing time-varying measures of technical efficiency and TFP for each city. Our empirical strategy exploits both cross-city and within-city temporal variation, estimating two-way fixed-effects models with city and year fixed effects and the full set of standard controls.

Studying cities rather than provinces alleviates aggregation bias: provincial aggregates pool heterogeneous agro-ecological zones, production structures, and input quality, potentially attenuating or even reversing the sign of covariate–TFP relationships. City-level data preserve meaningful within-province variation in digital finance penetration and agricultural technology, enabling identification under more comparable institutional, regulatory, and climatic environments. This finer granularity also reduces omitted-variable concerns tied to unobserved provincial composition effects and enhances the credibility of our instrumental-variable strategy, which relies on the net of differential exposure to national digital finance expansion and province-level trends on the focal city.

Section 4.1 describes the measurement of agricultural TFP, including the data sources and the stochastic frontier approach used for estimation. Section 4.2 specifies the baseline regression model, introduces the control variables, and reports the main results on the relationship between digital finance and agricultural productivity. Section 4.3 explores the underlying mechanism, focusing on the role of capital misallocation and testing its mediating effect. Finally, Section 4.4 conducts heterogeneity analysis across different regional and structural conditions to assess how the impact of digital finance varies with local characteristics.

4.1. Measurement of Agricultural TFP

To evaluate the impact of digital finance on agricultural TFP, we first estimate TFP using stochastic frontier analysis (SFA) [40,41,42].

A variety of methods exist for productivity measurement. In the industrial organization literature, approaches such as Olley–Pakes (OP) and Levinsohn–Petrin (LP) are widely used to address input endogeneity and selection bias at the firm level. However, these methods assume firm-level optimization and require micro-level data on investment or intermediate inputs, conditions that do not apply in our context where data are aggregated at the city level and agricultural input allocation is shaped by institutional and natural constraints.

By contrast, SFA is well suited to agriculture, where output is highly sensitive to exogenous shocks such as weather variability, pests, and diseases. SFA decomposes deviations from the production frontier into random noise and inefficiency, thereby separating uncontrollable shocks from differences in productive efficiency. It is also widely applied in agricultural economics at the sectoral and regional levels. For these reasons, we adopt the SFA framework to estimate agricultural TFP, specifying a stochastic frontier production function with a time-varying inefficiency component:

$$y_{it}=f(x_{it};\beta )\times \exp (v_{it}-{\mu }_{it})$$

(1)

$${\mu }_{it}={\mu }_i\times \exp [-\eta (t-T)]$$

(2)

$$\gamma ={\displaystyle \frac{{\delta }_{\mu }^2}{{\delta }_{\mu }^2+{\delta }_v^2}}$$

(3)

Here, $y_{it}$ denotes the real agricultural output of city i in period t, $f(·)$ represents the deterministic frontier production function, and $x_{it}$ is the vector of agricultural inputs. The parameter vector $\beta$ is to be estimated. The error term $v_{it}$ is assumed to be independently and identically distributed, while ${\mu }_{it}$ captures technical inefficiency and follows a non-negative truncated normal distribution.

Equation (2) specifies a time-varying inefficiency model, where ${\mu }_{it}$ is expressed as the product of a city-specific inefficiency level (${\mu }_i$) and an exponential function of time. The parameter $\eta$ reflects the rate of change in efficiency: when $\eta <0$, technical efficiency declines over time, with the speed of decline depending on the magnitude of $\eta$. The level of technical efficiency is given by $\exp (-{\mu }_{it})$, which lies within the interval $(0,1]$.

Equation (3) defines $\gamma$ as the share of the composite error attributable to inefficiency. A $\gamma$ close to 1 indicates that deviations from potential output are largely due to inefficiency, justifying the use of SFA. Conversely, a $\gamma$ approaching 0 suggests that random shocks dominate, in which case ordinary least squares would be more appropriate.

Following [42], we specify a translog production function as the functional form of the stochastic frontier model (Model “a0”):

$$\begin{array}{cc}\hfill \ln y_{it}& ={\beta }_0+{\beta }_n\ln n_{it}+{\beta }_k\ln k_{it}+{\beta }_m\ln m_{it}+{\beta }_{kn}\left(\ln k_{it}\right)\left(\ln n_{it}\right)\hfill \\ \hfill & +{\beta }_{km}\left(\ln k_{it}\right)\left(\ln m_{it}\right)+{\beta }_{nm}\left(\ln n_{it}\right)\left(\ln m_{it}\right)+{\beta }_{nn}{\left(\ln n_{it}\right)}^2\hfill \\ \hfill & +{\beta }_{kk}{\left(\ln k_{it}\right)}^2+{\beta }_{mm}{\left(\ln m_{it}\right)}^2+{\lambda }_t+v_{it}-{\mu }_{it}\hfill \end{array}$$

(4)

Here, $y_{it}$ denotes agricultural output per unit of cultivated land, and $n_{it}$, $k_{it}$, and $m_{it}$ represent labor, capital, and fertilizer inputs per unit of land for city i in year t, respectively. The coefficients $\beta$ are unknown parameters to be estimated, ${\lambda }_t$ captures time fixed effects, $v_{it}$ is the random error term, and ${\mu }_{it}$ is the inefficiency term. Based on this specification, agricultural TFP is obtained from the estimated frontier and can be expressed as $\exp ({\beta }_0+{\lambda }_t-{\mu }_{it})$, conditional on normalized input levels.

Following [42,43], we measure agricultural TFP using city-level data on aggregate agricultural output (farming, forestry, animal husbandry, and fishery), agricultural employment, total power of agricultural machinery, and net fertilizer application, all adjusted by relevant price indices for the period 2011–2021. Agricultural TFP is estimated using the likelihood-ratio (LR) test within the stochastic frontier framework. The data are drawn from statistical yearbooks and annual city-level statistical bulletins. Because of severe data gaps, the Xinjiang Uygur Autonomous Region is excluded from the sample, while linear interpolation is applied to handle missing observations in other regions.

Based on Equation (4), the baseline specification is denoted as Model “a0”. To assess the model’s validity, we test two hypotheses. The first assumes that the coefficients of the interaction terms are jointly zero, corresponding to the restricted specification Model “a1”. The second assumes that the coefficients of statistically insignificant variables are jointly zero, corresponding to Model “a2”. The LR test results are reported in

Table 1. Hypothesis test results of stochastic frontier model.

Model Test	Assumption	DF	LR	Results
a0 vs. a1	${\beta }_{kn}={\beta }_{km}={\beta }_{nm}={\beta }_{nn}={\beta }_{kk}={\beta }_{mm}=0$	6	75.13 ***	reject
a0 vs. a2	${\beta }_n={\beta }_{km}={\beta }_{kk}=0$	3	3.84	accept

Note: *** $p<0.01$.

.

As shown in Table 1, the interaction terms significantly affect agricultural output, supporting the adoption of the translog functional form. Given these results, Model “a2” is selected for subsequent analysis. We add a correlation matrix in

Table A1. Correlation matrix for the variables used in TFP estimation.

	Variables
	${\mathit{lny}}_{\mathit{it}}$	${\mathit{lnk}}_{\mathit{it}}$	${\mathit{lnm}}_{\mathit{it}}$	$\left({\mathit{lnk}}_{\mathit{it}}\right)\left({\mathit{lnn}}_{\mathit{it}}\right)$	$\left({\mathit{lnn}}_{\mathit{it}}\right)\left({\mathit{lnm}}_{\mathit{it}}\right)$	${\left({\mathit{lnn}}_{\mathit{it}}\right)}^{\mathit{2}}$	${\left({\mathit{lnm}}_{\mathit{it}}\right)}^{\mathit{2}}$
${\mathit{lny}}_{\mathit{it}}$	1.0000
${\mathit{lnk}}_{\mathit{it}}$	0.4398	1.0000
	(***)
${\mathit{lnm}}_{\mathit{it}}$	0.3795	0.3559	1.0000
	(***)	(***)
$\left({\mathit{lnk}}_{\mathit{it}}\right)\left({\mathit{lnn}}_{\mathit{it}}\right)$	0.5122	0.6357	0.2047	1.0000
	(***)	(***)	(***)
$\left({\mathit{lnn}}_{\mathit{it}}\right)\left({\mathit{lnm}}_{\mathit{it}}\right)$	−0.0812	−0.1698	0.3130	−0.5191	1.0000
	(***)	(***)	(***)	(***)
${\left({\mathit{lnn}}_{\mathit{it}}\right)}^{\mathit{2}}$	0.4554	0.3656	0.0401	0.8196	−0.3441	1.0000
	(***)	(***)	(**)	(***)	(***)
${\left({\mathit{lnm}}_{\mathit{it}}\right)}^{\mathit{2}}$	−0.2575	−0.2082	−0.8683	−0.0507	−0.2491	0.1555	1.0000
	(***)	(***)	(***)	(***)	(***)	(***)

Note: ** $p<0.05$, *** $p<0.01$.

for variables used in the TFP estimation in the Appendix A.

Table 2. Estimation results of agricultural TFP.

	Model a0	Model a1	Model a2
$\ln n_{it}$	−0.0697	0.1397 ***
	(0.0446)	(0.0153)
$\ln k_{it}$	0.1022 **	0.1600 ***	0.1181 ***
	(0.0511)	(0.0127)	(0.0143)
$\ln m_{it}$	0.3402 ***	0.1306 ***	0.3390 ***
	(0.0456)	(0.0116)	(0.0332)
$\left(\ln k_{it}\right)\left(\ln n_{it}\right)$	0.1513 ***		0.1051 ***
	(0.0420)		(0.0225)
$\left(\ln k_{it}\right)\left(\ln m_{it}\right)$	0.0184
	(0.0316)
$\left(\ln n_{it}\right)\left(\ln m_{it}\right)$	−0.1698 ***		−0.1380 ***
	(0.0276)		(0.0217)
${\left(\ln n_{it}\right)}^2$	−0.0891 ***		−0.0859 ***
	(0.0274)		(0.0272)
${\left(\ln k_{it}\right)}^2$	0.0076
	(0.0233)
${\left(\ln m_{it}\right)}^2$	0.1249 ***		0.1192 ***
	(0.0210)		(0.0208)
$\mu$	1.9044	1.8891	1.9118
$\eta$	−0.0024	−0.0026	−0.0022
$\gamma$	0.9563	0.9551	0.9561
Observations	2937	2937	2937

Note: ** $p<0.05$, *** $p<0.01$.

reports the estimation results of agricultural TFP. The parameter $\eta$ suggests that China’s agricultural technical efficiency has been gradually declining over time, with an average annual decrease of approximately 0.25%. Across all model specifications, the values of $\gamma$ range from 0.95 to 0.98, indicating that more than 95% of the deviation from the frontier is attributable to inefficiency rather than random shocks. This confirms the appropriateness of applying the stochastic frontier production function to estimate agricultural productivity.

We must admit that our estimates under the SFA framework do not recover farm- or plot-level production functions, within-activity technical change, or micro-adoption decisions. Because the data are city aggregates, composition shifts across agricultural subsectors may contribute to measured TFP even when within-subsector efficiency is unchanged; while we control for subsector shares and report robustness to alternative normalizations, residual composition effects and measurement noise at aggregation may remain. Consequently, our findings should be read as city-aggregate effects rather than micro-causal estimates, with finer decomposition left for future work when linked farm-level data become available.

4.2. Baseline Regression Results

The core explanatory variable is the digital finance index, obtained from China’s Digital Inclusive Financial Index, which has been widely used in studies of financial technology in China [13,16]. Following prior research [44,45], we adopt the “Credit Usage Index” to capture digital finance at the city level. Additional data are sourced from statistical yearbooks and annual city-level statistical bulletins covering the period 2011–2021.

After estimating agricultural TFP, we examine the impact of digital finance using the following empirical specification:

$$TF{P}_{it}={\alpha }_0+{\alpha }_1\ln \left(\mathrm{Digital}{\mathrm{finance}}_{it}\right)+{\alpha }_c{\mathrm{Control}}_{it}+{\mu }_i+{\mu }_t+{\epsilon }_{it}$$

(5)

Here, $TF{P}_{it}$ denotes the agricultural TFP of city i in year t. The key explanatory variable, $\mathrm{Digital}{\mathrm{finance}}_{it}$, represents the digital finance index at the city level. ${\mu }_i$ captures unobserved, time-invariant city-specific factors, while ${\mu }_t$ represents year fixed effects, controlling for common shocks across all cities. ${\epsilon }_{it}$ is the idiosyncratic error term. We use the Huber–White (and clustered) covariance estimator, which is robust to heteroskedasticity and within-panel autocorrelation.

The vector of control variables includes additional factors that may influence agricultural productivity. Specifically, $\ln \left(\mathrm{Rural}-\mathrm{power}\right)$ measures rural electricity consumption, $\ln \left(\mathrm{Fin}\right)$ denotes government expenditure on farming, forestry, and water affairs, and $\ln \left(\mathrm{Infra}\right)$ reflects highway mileage. $\ln \left(\mathrm{Income}\right)$ captures the level of regional economic development, proxied by per capita regional GDP. We also account for structural composition of agricultural output: Forestry is the share of forestry output in total agricultural output, Animal is the share of animal husbandry output, Fishery is the share of fishery output, and Agriculture ratio denotes the share of total agricultural output in gross regional output.

Table 3. Descriptive statistics of main variables.

Variable	Obs	Mean	S.D.	p5	p50	p95
TFP	2937	68.5679	40.7734	27.8930	58.3041	148.8810
$\ln \left(\mathrm{Digital}\mathrm{finance}\right)$	2911	4.7216	0.5655	3.9154	4.9387	5.2232
$\ln \left(\mathrm{Rural}-\mathrm{power}\right)$	2714	11.3884	1.6321	8.1259	11.4457	14.0147
$\ln \left(\mathrm{Fin}\right)$	2649	3.4886	0.7802	2.3219	3.5349	4.4355
$\ln \left(\mathrm{Infra}\right)$	2862	9.3322	0.6361	8.2918	9.4186	10.1310
$\ln \left(\mathrm{Income}\right)$	2897	10.5988	0.5487	0.0057	0.0337	0.1590
Forestry	2937	0.0526	0.0619	0.0012	0.0320	0.3240
Animal	2937	0.3183	0.1363	0.0179	0.1154	0.7093
Fishery	2889	0.0821	0.1205	0.0012	0.0320	0.3240
Agriculture ratio	2935	0.1567	0.0179	3.9154	0.1154	0.7093

reports the descriptive statistics of the main variables.

Table 4. Baseline regressions.

	(1)	(2)	(3)
$\ln \left(\mathrm{Digital}\mathrm{finance}\right)$	−1.6185 ***	−1.6700 ***	−1.5036 **
	(0.5866)	(0.6036)	(0.6017)
$\ln \left(\mathrm{Rural}-\mathrm{power}\right)$			−1.4222
			(1.0751)
$\ln \left(\mathrm{Fin}\right)$			−4.7761 ***
			(1.264)
$\ln \left(\mathrm{Infra}\right)$			−4.4512 ***
			(1.5634)
$\ln \left(\mathrm{Income}\right)$			0.5513
			(1.2733)
Forestry			5.2880
			(14.9415)
Animal			−7.1579
			(5.5942)
Fishery			46.4908
			(32.2697)
Agriculture ratio			41.3134 ***
			(9.0485)
Year FE	Yes	Yes	Yes
City FE	No	Yes	Yes
Observations	2911	2911	2415
$R^2$	0.7338	0.7338	0.7628

Note: ** $p<0.05$, *** $p<0.01$.

presents the baseline regression results for empirical model (5). Across all specifications, digital finance has a statistically significant negative effect on agricultural TFP. In column (1), without city fixed effects, the coefficient of digital finance is −0.162 and significant at the 1% level. Column (2), which adds city fixed effects, produces a nearly identical estimate (−0.167), underscoring the robustness of the result. In the fully specified model (column 3), after controlling for rural electrification, fiscal expenditure on agriculture, infrastructure, rural income, agricultural structure, and subsectoral shares, the coefficient remains negative and highly significant (−0.150). Substantively, a 1% increase in the digital finance index corresponds to an average decline of about 0.15 in agricultural TFP.

These results suggest that the expansion of digital financial services, despite their recognized benefits in other sectors, may exacerbate inefficiencies in agriculture. By diverting capital away from farming activities and amplifying resource misallocation, digital finance can inadvertently hinder productivity growth in the agricultural sector.

4.3. Mechanism: Capital Misallocation

As emphasized in seminal studies such as [30,46], a key mechanism underlying the negative effect of digital finance on agricultural TFP is capital misallocation. Unlike standardized industries such as manufacturing or services, agriculture is highly heterogeneous, and reliable information on production opportunities and risks is more difficult to obtain. As a result, agricultural capital markets are particularly prone to distortions, with resources often failing to flow to their most productive uses. The expansion of digital finance exacerbates this problem: financial institutions and investors, attracted by more transparent and profitable opportunities, tend to redirect attention and capital toward non-agricultural sectors. This reallocation of resources reduces the effective use of agricultural capital and ultimately suppresses agricultural TFP.

The strength of this adverse effect is expected to vary across regions. In areas where agricultural capital misallocation is already severe, financial institutions may find it more profitable to divert resources toward non-agricultural sectors. As a result, they are more likely to withdraw capital from agriculture, leading to a sharper decline in TFP as digital finance expands. By contrast, regions with relatively efficient capital allocation are less vulnerable to such adverse effects.

To empirically test this mechanism, we first construct a measure of agricultural capital misallocation. Following [30,46], we define ${\tau }_{ki}$ as the degree of capital misallocation in region i, given by

$${\displaystyle \frac{1}{1+{\tau }_{ki}}}={\displaystyle \frac{K_i/K}{\left(s_i{\beta }_{ki}\right)/{\beta }_k}}$$

(6)

Here, $K_i/K$ denotes the share of capital in region i relative to total capital, and $s_i$ is the share of agricultural output in region i relative to total agricultural output. ${\beta }_{ki}$ represents the output elasticity of capital in region i, while the weighted average elasticity is ${\beta }_k={\sum }_{i=1}^N{s}_i{\beta }_{ki}$. Equation (6) thus defines ${\tau }_{ki}$. When ${\tau }_{ki}>0$, the capital share in region i falls short of the level implied by its output elasticity, suggesting that the marginal product of capital is relatively high and the region is undercapitalized. Larger values of ${\tau }_{ki}$ therefore indicate more severe capital misallocation.

To examine heterogeneity, we divide the sample into two groups based on the degree of agricultural capital misallocation: low (below the median) and high (above the median). Columns (1) and (2) of

Table 5. The role of capital misallocation.

Dependent Variable: Agricultural TFP	Capital Misallocation (Type 1)		Capital Misallocation (Type 2)
	(1) Low	(2) High	(3) Low	(4) High
$\ln \left(\mathrm{Digital}\mathrm{finance}\right)$	−0.6450	−1.9530 **	−0.8532	−1.9108 **
	(0.5288)	(0.7988)	(2.0681)	(0.8038)
Controls	Yes	Yes	Yes	Yes
City FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
Observations	1252	1163	1121	1068
$R^2$	0.8991	0.6548	0.9029	0.6984

Note: We compute two types of capital misallocation using alternative measures of capital input. For capital misallocation (type 1), capital input is measured by the total power of agricultural machinery; for type 2, it is measured by agricultural fixed asset investment. ** $p<0.05$.

measure capital input using the total power of agricultural machinery, while columns (3) and (4) use agricultural fixed asset investment. The results show that digital finance exerts a statistically significant negative effect on agricultural TFP in regions with high capital misallocation (columns 2 and 4), whereas the effect is smaller and statistically insignificant in regions with low misallocation. These findings provide supporting evidence that the adverse impact of digital finance on agricultural productivity operates primarily through the channel of capital misallocation.

4.4. Heterogeneity Analysis

The preceding analysis suggests that digital finance affects agricultural TFP primarily through capital misallocation. To further investigate heterogeneity, we examine three factors: labor output elasticity, natural resource endowment, and cultivation scale. Specifically, we assess how these factors condition the interaction between digital finance and capital misallocation.

First, labor output elasticity may amplify the adverse impact of capital misallocation. In regions with high labor elasticity, agricultural output is more sensitive to input allocation. Since capital misallocation also influences labor input, the negative effect of digital finance is expected to be stronger in such areas.

Table 6. Heterogeneity Test I: Labor output elasticity in agriculture.

	(1) Low	(2) High
$\ln \left(\mathrm{Digital}\mathrm{finance}\right)\times \mathrm{Capital}\mathrm{misallocation}$	−5.4774	−12.1478 **
	(4.3556)	(5.2863)
$\ln \left(\mathrm{Digital}\mathrm{finance}\right)$	−6.1506 *	−9.4297 ***
	(3.3545)	(2.0386)
Capital misallocation	5.0360 *	6.1673 ***
	(2.5946)	(2.0930)
Controls	Yes	Yes
City FE	Yes	Yes
Year FE	Yes	Yes
Observations	1170	996
$R^2$	0.8059	0.7652

Note: * $p<0.1$, ** $p<0.05$, *** $p<0.01$.

reports the heterogeneity analysis by labor output elasticity. Cities are divided into two groups: low (below the median) and high (above the median). The results show that the interaction between digital finance and capital misallocation is significantly negative only in regions with high labor output elasticity, indicating that the adverse effect of digital finance on agricultural TFP is concentrated where labor contributes more strongly to agricultural production. Figure 1 shows the above results.

A second dimension of heterogeneity arises from grain production status. Non-major grain-producing areas typically receive weaker financial support and less policy protection, leaving them more vulnerable to capital crowding-out.

Table 7. Heterogeneity Test II: Grain production status.

	(1) Non-Major Grain Areas	(2) Major Grain Areas
$\ln \left(\mathrm{Digital}\mathrm{finance}\right)\times \mathrm{Capital}\mathrm{misallocation}$	−13.0215 ***	−1.1751
	(3.1383)	(1.3469)
$\ln \left(\mathrm{Digital}\mathrm{finance}\right)$	−1.4603	0.0038
	(1.6098)	(0.3072)
Capital misallocation	9.0230 ***	−1.2170
	(0.1917)	(1.2795)
Controls	Yes	Yes
City FE	Yes	Yes
Year FE	Yes	Yes
Observations	815	1600
$R^2$	0.7812	0.8925

Note: *** $p<0.01$.

reports the results for major and non-major grain-producing regions. The coefficient on the interaction term between digital finance and capital misallocation is significantly negative at the 1% level only in non-major grain-producing areas, suggesting that these regions are particularly exposed to the adverse effects of digital finance. By contrast, the interaction effect is insignificant in major grain-producing areas, consistent with the notion that stronger natural resource endowments and policy support buffer the negative impact. Figure 2 shows these results.

Finally, cultivation scale is another source of heterogeneity in the impact of digital finance on agricultural TFP. Small-scale farms, constrained by limited information and weaker links to formal financial institutions, receive less attention from financial intermediaries and are thus more vulnerable to capital misallocation and crowding-out effects. By contrast, larger-scale farms—better integrated into financial networks and able to demonstrate stable returns—are relatively insulated from such distortions. Columns (1) and (2) of

Table 8. Heterogeneity Test III: Planting scale.

	(1) Small	(2) Large
$\ln \left(\mathrm{Digital}\mathrm{finance}\right)\times \mathrm{Capital}\mathrm{misallocation}$	−11.3480 ***	0.5476
	(2.2910)	(2.6277)
$\ln \left(\mathrm{Digital}\mathrm{finance}\right)$	−3.1173 ***	−1.0142 **
	(0.0861)	(0.5076)
Capital misallocation	8.5981 ***	−2.1096
	(1.7789)	(1.6266)
Controls	Yes	Yes
City FE	Yes	Yes
Year FE	Yes	Yes
Observations	1138	1277
$R^2$	0.7846	0.8515

Note: ** $p<0.05$, *** $p<0.01$.

report results based on the median per capita cultivation scale (crop planting area per agricultural worker). The interaction between digital finance and capital misallocation significantly reduces agricultural TFP in small-scale farming regions, whereas the effect is statistically insignificant in large-scale regions. Figure 3 shows these results.

5. Robustness Check

5.1. Instrumental Variables

To address potential endogeneity concerns, we employ two instrumental variables following [47]. Instrumental Variable 1 (IV1) is constructed as the product of the one-period lagged digital finance index and the first-order difference of the national digital finance index, following a Bartik-style approach. This instrument is strongly correlated with local digital finance development, as current adoption is influenced by both past local financial infrastructure and national trends. At the same time, it is plausibly exogenous to local agricultural TFP because the national index aggregates data from over 200 cities and is unlikely to be materially affected by the agricultural performance of any single city.

Instrumental Variable 2 (IV2) is defined as the average provincial digital finance index excluding the focal city. This instrument captures broader regional trends in digital finance that influence local digital finance development while remaining independent of city-specific agricultural productivity shocks. By excluding the city itself, IV2 mitigates potential reverse causality, ensuring that the instrument affects local agricultural TFP solely through its impact on digital finance.

Table 9. Instrumental variable regressions.

	IV1		IV2
	First Stage	Second Stage	First Stage	Second Stage
$\ln \left(\mathrm{Digital}\mathrm{finance}\right)$		−5.1919 ***		−22.2355 ***
		(1.4501)		(3.9199)
Bartik instrument	0.0009 ***
	(0.0000)
Average provincial digital finance index			0.7035 ***
			(0.0874)
Controls	Yes	Yes	Yes	Yes
City FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
Observations	2164	2164	2372	2372
$R^2$	0.4180	0.4180	0.9624	0.9624

Note: *** $p<0.01$. The instruments used are as follows: (i) Bartik instrument, defined as the product of the lagged digital finance index and the first-difference of the national digital finance index; (ii) the average provincial digital finance index excluding the focal city. Both instruments are strongly correlated with local digital finance while plausibly exogenous to city-level agricultural productivity.

shows that both IVs are strongly valid, with first-stage coefficients significantly positive at the 1% level. The second-stage estimates confirm that the digital finance index retains a significantly negative effect on agricultural TFP at the 1% level, reinforcing the robustness of our main findings. We provide some discussion for the instrumental variables.

It is possible that regional or national macro-trends embedded in the national first-difference could directly influence agricultural TFP and that broader provincial digital-finance trends may proxy for regional development levels that affect TFP through non-digital channels. We acknowledge these risks and view our IV strategy as partially mitigating—not eliminating—them. Our approach, however, partially addresses these issues on several fronts. First, the Bartik-style instrument exploits predetermined city exposure (lagged digital finance) interacted with the national first-difference, identifying the differential impact of a common national expansion across cities after absorbing nationwide shocks with year fixed effects and conditioning on rich time-varying controls; in robustness, province×year fixed effects further net out regional macro- or policy shocks. Second, the leave-one-out provincial mean shifts a city’s digital finance via diffusion from outside the city itself, and we limit direct channels by the same fixed-effects saturation and controls; results are stable when excluding provincial capitals/platform hubs. Third, diagnostics support credibility: strong first stages, over-identification tests that do not reject, lead/placebo checks showing no pre-trends, and sensitivity analyses (alternative controls, trends, outcome normalizations) that leave the sign and magnitude intact; Oster-type bounds indicate that unobservables would need implausibly large selection to overturn the core effect.

5.2. Further Robustness Test

We conduct further robustness checks by modifying the dependent variable. Specifically, we use the price-adjusted value added of the agricultural sector at the prefecture level as the output variable, with labor and capital per unit of land as inputs. Following the methodology outlined in Section 2, agricultural TFP is first estimated using the stochastic frontier approach, and the resulting TFP measures are then employed in the regressions. Columns (1) and (2) of

Table 10. Robustness checks.

	Alternative TFP Measure (Value Added)		Samples Before and After 2015
	(1)	(2)	(3) Before 2015	(4) After 2015
$\ln \left(\mathrm{Digital}\mathrm{finance}\right)$	−2.8025 **	−2.6514 *	−0.3267 **	−22.0379 ***
	(1.1016)	(1.3686)	(0.1353)	(5.0490)
Controls	No	Yes	Yes	Yes
City FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
Observations	2909	2415	1163	1252
$R^2$	0.1669	0.2825	0.7482	0.7594

Note: Columns (1) and (2) replace the dependent variable with the price-adjusted value added of the agricultural sector (TFP estimated via stochastic frontier). Columns (3) and (4) split the sample into pre- and post-2015 periods. * $p<0.1$, ** $p<0.05$, *** $p<0.01$.

show that the main findings remain robust under this alternative specification.

We further examine the heterogeneity of the effect by splitting the sample into pre- and post-2015 periods. In 2015, a landmark Internet personal credit product—Huabei—was launched, and it has since become widely adopted among Chinese consumers. As digital finance became more widespread in cities after 2015, it is more likely to exacerbate capital misallocation by diverting resources away from agriculture and into non-agricultural sectors. Column (3) reports the regression results for the pre-2015 sample, while column (4) reports the results for the post-2015 sample. Consistent with this mechanism, the negative effect of digital finance on agricultural TFP is stronger in the post-2015 period, as shown in column (4). Notice that the 2015 contrast is illustrative only and that our inference does not depend on a DID design based on this 2015 shock.

6. Conclusions

This study demonstrates that the expansion of digital finance significantly reduces agricultural TFP in China. The adverse effect operates mainly through capital misallocation, as financial resources are diverted away from agriculture, and is most pronounced in regions with pre-existing capital market distortions. The heterogeneity analysis further shows that labor-intensive regions, non-major grain-producing areas, and small-scale farming regions are particularly vulnerable, reflecting their higher sensitivity to capital shortages and weaker connections to formal financial networks.

These results highlight the unintended consequences of rapid digital financial development for the agricultural sector and point to the need for targeted policy interventions. To foster inclusive and sustainable productivity growth, policymakers should improve capital allocation efficiency in rural areas, strengthen financial access for smallholders, and tailor digital finance initiatives to local structural characteristics. Overall, this study provides novel evidence on the complex interaction between digital finance and agricultural productivity in emerging economies.