The development of agricultural insurance has long been constrained by difficulties in risk identification and pricing, information asymmetry, and high transaction costs. By incorporating satellite remote sensing, meteorological monitoring, and big-data modeling, InsurTech enhances the acquisition and processing of risk information, thereby alleviating information asymmetry between insurance providers and policyholders. At the same time, digital technologies and platform-based applications facilitate the diffusion of agricultural insurance products and risk management technologies. Moreover, InsurTech simplifies underwriting, claims settlement, and regulatory processes, reducing transaction costs in the operation of agricultural insurance.

We employ a Bayesian signal extraction model (Xiang, 2017).

Let the farmer's inherent risk parameter be denoted by θ, θ□N(μθ,σθ2).

With the adoption of InsurTech, the signal observed by the insurer is given by: s = θ+ε,

Here, T represents the level of InsurTech adoption.

The posterior distribution is:

The linear function of each category of indemnity is given by:

By pricing according to the conditional expectation, we obtain:

Where m represents the loading factor, the mean squared error (MSE) of pricing is:

By substituting the linear expected indemnity, we obtain:

This yields:

It can be observed that, as σε2(T) decreases with increasing T, dMSE(T)dT<0, higher levels of technology lead to more accurate pricing and weaker adverse selection. From an economic perspective, the above results not only reflect the mechanism through which InsurTech improves agricultural insurance pricing efficiency by reducing information asymmetry, but also implicitly capture other channels through which InsurTech influences the development of agricultural insurance. Specifically, enhanced capabilities in acquiring and processing risk information facilitate the dissemination and adoption of insurance technologies and products across regions and market participants, reflecting a technology diffusion effect. At the same time, more accurate pricing and automated processes reduce insurers' resource inputs in risk assessment, contract execution, and claims settlement, thereby lowering transaction costs in the operation of agricultural insurance. Therefore, the model characterizes the comprehensive impact of InsurTech on agricultural insurance development from multiple theoretical dimensions.

Therefore, we propose hypotheses:

China's land area is about 9.60 million km², with roughly 4.73 million km² of sea area; its land boundary is approximately 22,000 km and its continental coastline about 18,000 km (Central People's Government of the People's Republic of China, 2024). Provinces differ markedly in topography and climate, as well as in agricultural foundations and economic development. The effect of InsurTech on agricultural insurance is therefore likely to exhibit substantial regional heterogeneity.

To further investigate the potential regional heterogeneity in the impact of InsurTech on agricultural insurance development, this study introduces a regional variable into the Bayesian Signal Extraction Model. Let the regional variable be denoted by r, and the farmer's inherent risk parameter by θ, θ□N(μθr,σθr2).

With the application of InsurTech, the insurer observes the following signal:

Here, T_r represents the level of InsurTech adoption in a given region r.

The posterior distribution is:

The pricing curve is:

Accordingly, the mean squared pricing error (MSE) is:

Let σεr2≡gr(Tr), Differentiating with respect to T_r, we obtain:

Thus, the larger σθr2, the faster the mean squared pricing error (MSE) declines. Likewise, the greater the magnitude of |∂σεr2∂Tr|, the faster the MSE decreases. These results indicate that InsurTech generates greater marginal benefits in regions characterized by high risk volatility and substantial potential for data improvement, while the benefits are smaller in regions with low risk volatility or limited data improvability. However, in the actual agricultural production landscape, there are systematic differences between major grain-producing regions and non-major grain-producing regions in terms of agricultural risk characteristics and data infrastructure conditions. Specifically, major grain-producing regions typically undertake agricultural production tasks of higher intensity and larger scale, with more concentrated agricultural operations. They are more significantly affected by natural disasters and market fluctuations, resulting in overall higher volatility in agricultural risks. At the same time, major grain-producing regions have a better data foundation and greater potential for information improvement in areas such as agricultural inputs, production monitoring, and policy support, making it easier for InsurTech to play a role in risk identification, pricing optimization, and claims efficiency enhancement. In contrast, agricultural production in non-major grain-producing regions is relatively decentralized, with relatively limited risk volatility and data accumulation levels. The marginal effect of InsurTech in improving information structures and enhancing insurance operational efficiency may be relatively weaker in these areas. Therefore, from the perspective of functional zoning for grain production, major grain-producing regions can be seen as the practical embodiment of areas with “high risk volatility and relatively high data improvability.” The role of InsurTech in promoting the development of agricultural insurance is expected to be more significant in such regions.

Therefore, we propose Hypothesis:

Farm income shapes farmers' risk tolerance and capacity for reproduction, and indirectly reflects their access to resources. High-income farmers typically face lower liquidity constraints and may demand more insurance; at the same time, they can partially self-insure or manage risk through non-insurance instruments. By contrast, low-income farmers often face tighter liquidity constraints and limited ability to pay for insurance. With premium subsidies and InsurTech, however, more precise underwriting can reduce effective premiums, improving affordability and enabling participation by lower-income farmers.

To analyze the effect of farmers' income on the demand for agricultural insurance, this study constructs a premium model and a demand function in order to explore the potential threshold effect of income. In general, farmers' demand for insurance exhibits a decreasing relationship with the premium rate. We further develop a demand function for agricultural insurance that incorporates both farmers' income and InsurTech investment. Let P(y) denote the effective premium actually paid by farmers:

Here, a denotes the fixed charge or base premium; b is the coefficient reflecting the effect of farmers' income on the premium, capturing the sensitivity of premiums to income (typically b > 0); y represents farmers' income; T denotes InsurTech investment; and is the coefficient associated with T.

If InsurTech improves the accuracy of risk assessment and reduces measurement error, thereby affecting the pricing of insurance products, then c < 0. This implies that while farmers' insurance expenditures increase with income, InsurTech investment (T) reduces farmers' actual insurance costs through more precise pricing.

Let D(y) denote farmers' demand for agricultural insurance. Assuming that demand is a function of farmers' income and the insurance premium rate, the demand function can be expressed as follows:

where y denotes farmers' income and P(y) is the premium.

Substituting P(y) into the demand function yields:

Farmers' demand for agricultural insurance is positively related to income. However, when insurance costs are excessively high, low-income farmers may be unable to afford expensive premiums and thus reduce their participation.

Suppose the threshold income level is y_q, with the corresponding premium P(y_q), When farmers' income reaches or exceeds y_d, they are willing to purchase insurance; if income falls below y_d, they are unwilling or unable to do so. With increasing InsurTech investment, premiums are reduced, which can improve affordability and expand participation. Nevertheless, not all farmers benefit equally from InsurTech. Only when farmers' income reaches a certain threshold can the effects of InsurTech materialize, thereby enhancing participation in agricultural insurance. The existence of an income threshold effect implies that increases in income can reduce the exclusion of low-income farmers from the insurance market and raise their level of participation. As InsurTech investment T increases, reductions in agricultural insurance premiums may enhance affordability for a broader group of farmers, thereby stimulating demand for agricultural insurance across different income levels. However, owing to income-related differences in risk exposure and technological absorption capacity, the promotional effect of InsurTech is likely to be more pronounced among higher-income farmers. Specifically, low-income farmers often face more severe liquidity constraints, making their decisions to participate in agricultural insurance highly sensitive to premium levels. By contrast, although high-income farmers possess a certain degree of self-insurance capacity, they typically operate at a larger production scale and are exposed to higher levels of agricultural risk, which sustains a relatively strong marginal demand for formal risk management instruments. In the presence of government premium subsidies and the application of InsurTech, improvements in risk assessment accuracy and underwriting precision help to reduce the effective premiums paid by farmers. This, in turn, alleviates the financial constraints faced by low-income farmers and enables their participation in agricultural insurance. Consequently, InsurTech does not exclusively benefit high-income farmers but contributes to the development of agricultural insurance across all income groups. Nevertheless, because high-income farmers generally exhibit stronger capabilities in information acquisition, technological adoption, and risk management, InsurTech is more readily translated into higher participation rates and more comprehensive insurance coverage within this group. As a result, the positive impact of InsurTech on agricultural insurance development may increase with rising income levels, ultimately manifesting as an income-based threshold effect.

Therefore, we propose Hypothesis:

H3. The impact of InsurTech on the development of agricultural insurance exhibits a threshold effect across different income levels, and its promotional effect gradually strengthens as farmers' income levels increase.

This study conducts an empirical analysis using provincial panel data covering 31 provinces in China from 2011 to 2023. The data were primarily drawn from multiple authoritative sources, including: China Insurance Yearbook (2012–2024), the Peking University Digital Inclusive Finance Index Database (data released in 2024), China Population and Employment Statistical Yearbook (2012–2024), China Rural Statistical Yearbook (2012–2024), China Finance Yearbook (2012–2023), the 2024 Budget Implementation Reports of individual provinces, and various provincial statistical yearbooks. The definitions of variables, computation methods, and corresponding data sources are presented in Table 1.

To examine whether InsurTech promotes the development of agricultural insurance, this study employs a two-way fixed effects model. The regression model is specified as follows:

where the subscript i denotes province (i = 1, 2, …, 31), and t denotes year (t = 2011, 2012, …, 2023). y_it represents the level of agricultural insurance in province i at time t. α is the constant term (intercept), representing the baseline level of agricultural insurance development when all explanatory and control variables are equal to zero. x_it is the core explanatory variable, and β is its estimated coefficient. z_ijt denotes the control variables, with γ_j as their corresponding coefficients, represents the jth control variable. λ_i and μ_t are province and year fixed effects, respectively, included to control for unobserved heterogeneity across provinces and years. Finally, ε_it is the random disturbance term.

The dependent variable is the level of agricultural insurance development. Prior studies commonly proxy this construct using premium income (Ruan et al., 2024), claims expenditure (Li, 2016), composite index scores derived from indicator systems (Liu and Yang, 2024), insurance density (Ye and Chen, 2022), or insurance depth (Wang et al., 2024). However, absolute measures such as premium income and claims expenditure can be confounded by cross-regional differences in agricultural scale and development, which undermines comparability. Composite indices may capture multiple dimensions more comprehensively, but their construction is often constrained by methodological choices and may include variables that also serve as explanatory or control variables in the main regressions, thereby risking estimation bias.

Accordingly, we use agricultural insurance depth (y1) as the proxy for the level of agricultural insurance development. This indicator aligns well with regional realities and offers stronger interpretability relative to alternatives. For robustness and endogeneity checks, we also employ agricultural insurance density (y2) as an alternative dependent variable.

The explanatory variable is the level of InsurTech development. Following the approach in Wanyan and Suo (2019) and Ye (2024), we use the insurance subindex of the Peking University Digital Inclusive Finance Index (hereafter, the “Digital Inclusive Finance Insurance Index”) as a proxy for InsurTech development. In the literature, some studies use this subindex directly, whereas others weight it by provincial premium income (Chen, 2024). We adopt the direct measure because our dependent variables are constructed from agricultural insurance premium income; weighting the subindex by premiums could introduce multicollinearity and thereby distort the regression results.

The Digital Inclusive Finance Insurance Index is compiled by the Peking University Digital Finance Research Center using Ant Financial's transaction-account data. The insurance-business subindex is constructed from three indicators: the number of insured users per 10,000 Alipay users, the average number of insurance transactions per user, and the average insurance amount per user. Each indicator is normalized, and the weights are determined by combining the analytic hierarchy process (AHP) with the coefficient-of-variation method. The weighted aggregation yields the insurance-business subindex (Guo et al., 2020).

We account for several covariates. For farmer characteristics, we use years of schooling among rural residents (z1). To capture regional agricultural structure and development, we include the share of primary industry in GDP (z2), the share of sown grain area (z6), and per capita grain output (z10). Production conditions are proxied by total agricultural machinery power per unit area (z3), pesticide use per unit area (z4), and chemical fertilizer application per unit area (z5). Regional production efficiency is represented by agricultural labor productivity (z9). Fiscal support for agriculture is measured by the ratio of local budgetary expenditure on agriculture–forestry–water affairs to total general public budget expenditure (z7). The level of rural financial development (z8) is measured by the ratio of agriculture-related loans to the value added of the primary industry.

It should be noted that the variable share of sown grain area (z6) is used to measure the relative scale of grain cultivation within the agricultural production structure of a region. We focus on grain crops because this sector faces the highest level of exposure to natural risks and represents the primary coverage area of China's policy-led agricultural insurance. To more accurately capture pre-planting decisions, we use sown area rather than yield as a more exogenous and predetermined measure. Furthermore, by employing a proportional measure rather than absolute area, we control for provincial scale differences, thereby enabling a more precise reflection of the internal structure of agricultural production. Data for this indicator are sourced from official statistics, which ensures consistency in measurement standards and strong continuity across years.

We use farmers' per capita disposable income (k) as the threshold variable. This indicator captures household heterogeneity, economic endowments, and risk-coping capacity. Section 3.3 provides the theoretical rationale for a potential income threshold effect.

Table 2 reports the descriptive statistics for the variables used in the empirical analysis. The dependent variable—agricultural insurance depth (y1)—has a maximum of 7.08092%, a minimum of 0.74685%, and a median of 0.4701%, indicating substantial dispersion and pronounced differences in agricultural insurance development across years and provinces. The core explanatory variable—the Digital Inclusive Finance Insurance Index (x1)—ranges from 0.2500 to 989.1700, with a median of 530.5000, likewise exhibiting considerable variation. To mitigate potential heteroskedasticity, the empirical estimations apply logarithmic transformations to the dependent variable (including its alternative used in robustness checks) and to the core explanatory variable.

Building on the two-way fixed-effects framework, we estimate the empirical model of InsurTech's impact on agricultural insurance development (see Equation 7). Table 3 reports the results from Equation 7, presenting specifications without and with control variables. Specifically, columns (1) and (2) in Table 3 summarize the effects of InsurTech under these two settings. The estimates indicate a significant positive association between InsurTech and the development of agricultural insurance. The coefficient on InsurTech remains positive and statistically significant at the 1% level after adding controls, which supports Hypothesis H1 and suggests that the application of InsurTech is conducive to the development of agricultural insurance.

Regarding the controls, we discuss only those that are significant at the 10% level. The coefficient on z2 (the share of primary industry in regional GDP) is positive, implying that as agriculture's relative scale within the economy expands, the facilitating role of InsurTech becomes more apparent. The coefficient on z7 (the ratio of budgetary expenditure on agriculture–forestry–water affairs to total general public budget expenditure) is also positive, consistent with leverage and signaling effects of public spending in this domain. It is noteworthy that, as shown in Table 3, this study finds that both pesticide use per unit area (z4) and fertilizer application per unit area (z5) exhibit a significantly negative impact on the development level of agricultural insurance. This result may reflect two types of mechanisms: first, a moral hazard mechanism, whereby agricultural insurance coverage may incentivize farmers to increase chemical inputs in pursuit of short-term output, thereby raising the risk level of the production system; second, an environmental and sustainability mechanism, as over-reliance on pesticides and fertilizers is often associated with unsustainable agricultural production practices, which may exacerbate soil degradation and ecological risks, thereby hindering the long-term healthy development of agricultural insurance. This suggests that policymakers should, while promoting agricultural insurance, focus on guiding green production behaviors and fostering synergy between insurance mechanisms and sustainable agricultural policies. Finally, z9 (labor productivity) and z10 (per capita grain output) show negative coefficients, which may reflect structural adjustments in agriculture—for example, shifts away from grain production—that influence observed grain output.

We replace the original dependent variable with agricultural insurance density (y2) and re-estimate Equation 7. The regression results are reported in Table 4, columns (1) and (2), corresponding to the specifications without and with control variables, respectively. The findings show that, irrespective of whether controls are included, InsurTech maintains a statistically significant positive effect on the development of agricultural insurance.

Building on Zhu's (2022) research, we use the one-period lag of the explanatory variable as an instrument and estimate the model with a panel two-stage least squares (2SLS) approach. Table 4, columns (3) and (4), report the first- and second-stage results, respectively. In the first stage, the coefficient on the lagged Digital Inclusive Finance–Insurance Index, L.ln(x), is positive and statistically significant at the 1% level, indicating that the instrument has strong explanatory power for the endogenous regressor. In the second stage, InsurTech continues to exert a significant positive effect on agricultural insurance. In addition, tests of the null hypothesis of “insufficient instrument identification,” based on the Kleibergen–Paap Wald F-statistic and the Cragg–Donald Wald F-statistic, show that the instrument passes weak-instrument diagnostics, implying that the IV estimates are valid. These findings indicate that, after accounting for endogeneity, the paper's main conclusion remains intact.

Given that the COVID-19 pandemic (2020–2022) constituted an exogenous shock to the agricultural economy and could affect the robustness of the estimates, we conduct a restricted-period check by removing observations from 2020 to 2022 and re-estimating the model. As reported in Table 4, column (5), the results remain consistent with the baseline findings, further corroborating the robustness of the paper's conclusions.

To further address potential endogeneity, we apply propensity score matching (PSM) to the full sample. Provinces are divided into a treatment group (x > sample mean) of the Digital Inclusive Finance–Insurance Index and a control group (x ≤ sample mean). We implement nearest-neighbor matching with replacement at a 1:2 ratio and a caliper radius of 0.01. To reflect the specific channels through which InsurTech may operate, all control variables from the baseline regression are included in the propensity-score model. The specific steps are as follows:

Step 1: Definition of the treatment and control groups. The Digital Inclusive Finance–Insurance Index is used as the treatment variable. Based on its sample mean, observations are divided into a treatment group (index above the mean) and a control group (index at or below the mean).

Step 2: Estimation of propensity scores. A Logit model is employed to estimate the probability (propensity score) that each observation is assigned to the treatment group. All control variables used in the baseline regression are included in the model to ensure comparability.

Step 3: Selection of the matching method. Based on the estimated propensity scores, this study applies a 1:2 nearest-neighbor matching with replacement and sets a caliper radius of 0.01 to improve matching quality and reduce extreme matching errors.

Step 4: Assessment of matching quality. Standardized bias tests, common support checks, and kernel density comparisons are conducted to evaluate the balance and comparability between the treatment and control groups before and after matching.

Step 5: Post-matching regression analysis. The regression model is re-estimated using the matched sample to examine whether the impact of InsurTech on agricultural insurance development remains significant after controlling for sample selection bias.

Figure 2 shows that standardized biases for most covariates decrease after matching. Figure 3 indicates that the majority of observations lie on common support. Figure 4 compares kernel density distributions before and after matching and shows closer alignment between the treatment and control groups post-matching, satisfying the common-support assumption.

Table 5 reports balance tests for the matched covariates. Post-matching standardized differences are substantially reduced and remain below 10% for all controls; two-sample t-tests fail to reject the null hypothesis of no systematic differences between the two groups. These results indicate that the matched samples are well balanced.

Table 6 shows that, after PSM, the coefficient on the core explanatory variable (InsurTech level) is positive and statistically significant at the 1% level. Accounting for sample-selection bias therefore leaves the main finding intact, providing additional support for Hypothesis 1.

To further examine regional heterogeneity in the effect of insurance technology (InsurTech) on the development of agricultural insurance, we partition the sample into major grain-producing regions and non-major grain-producing regions. We then estimate the same two-way fixed effects model as in the baseline specification, given by Equation 7.

Table 7 reports the heterogeneity results. Columns (1)–(2) present estimates for major grain-producing regions without and with the full set of control variables, respectively; columns (3)–(4) report the corresponding estimates for non-major grain-producing regions.

Across both subsamples, InsurTech exerts a positive and statistically significant effect on the development of agricultural insurance, with coefficients significant at the 1% level. Notably, the coefficient in the non-major grain-producing regions is larger than that in the major grain-producing regions, indicating a stronger marginal impact of InsurTech outside the core grain areas.

A plausible explanation is that InsurTech alleviates key frictions that hinder agricultural insurance development—frictions that tend to be more pronounced in non-major grain-producing regions. As a result, the adoption of InsurTech yields comparatively larger gains in these areas. These regional heterogeneity findings further corroborate Hypothesis H2.

To further examine whether the impact of InsurTech on agricultural insurance varies across rural income levels, we follow Hansen's (1999) methodology and specify the following threshold regression model:

Here, k_it is the threshold variable—rural residents' disposable income; π_n denotes the threshold value(s); φ₁, φ₂, φ_n, φ_n+1 are the regime-specific slope coefficients in the segmented threshold model; I is an indicator function that equals 1 when the stated condition holds and 0 otherwise; z_ijt represents control variables, λ_i and μ_t are time and individual fixed effects, respectively, and ε_it is the error term.

We first test whether a threshold effect exists—that is, whether rural disposable income serves as a valid threshold variable. Using bootstrap resampling (300 replications), we compute the F-statistics and corresponding p-values and conduct sequential tests for triple-, double-, and single-threshold specifications to assess statistical significance.

Table 8 reports the results. When rural disposable income is used as the threshold variable, the null of no threshold is rejected at the 10% significance level for both the single- and double-threshold tests. This indicates the presence of two threshold values and thereby corroborates Hypothesis H3. Accordingly, we estimate the double-threshold model in Equation 8.

The double-threshold regression results are presented in Table 9. The first threshold is 6,007.5000 yuan (CNY) and the second is 12,841.9004 yuan. Figure 5 displays the profile of the likelihood-ratio (LR) statistic for the double-threshold case. The estimates indicate that:

All three coefficients are significant at the 1% level. These findings imply that as rural disposable income increases, the promotional effect of InsurTech on the development of agricultural insurance strengthens; in other words, InsurTech exhibits larger marginal effects among higher-income rural households.