Adoption of Direct Seeding, Yield and Fertilizer Use in Rice Production: Empirical Evidence from China

Abstract

Direct seeding has been widely adopted in rice production due to its advantages of water- and labor-saving characteristics in China and other countries. This study aims to examine the effects of farmers’ adoption of direct seeding on the yield and fertilizer use in rice production in China. Using survey data of 1002 rice farmers in the Yangtze River Basin in China, the endogenous switching regression model is used to address the self-selectivity bias from both observed and unobserved heterogeneity. The results show that about 79% of surveyed farmers adopted direct seeding in rice production. After addressing the self-selectivity bias, the adoption of direct seeding increased rice yield among the adopters by 3.65%, and would have increased rice yield among the non-adopters by 1.54% if direct seeding had been adopted. In addition to its positive effect on rice yield, however, the adoption of direct seeding was also found to increase fertilizer use among the adopters by 19.84%, and would have increased fertilizer use among the non-adopters by 37.75% if direct seeding had been adopted. Moreover, farmers’ adoption of direct seeding exerted heterogeneous effects on rice yield and fertilizer use in terms of farm size and location.

1. Introduction

In many developing countries, agriculture is confronted with the shortages of water and an effective labor force [1,2]. China’s agriculture, for example, consumes more than 60% of total water resources [3], but it is in fiercer competition for water resources with the non-agricultural sector and human consumption [4]. Moreover, the massive rural–urban migration reduces the supply of effective agricultural laborers [5]. In 2020, for example, a total of 169.6 million rural laborers left their hometown and moved into the urban areas, accounting for about 33.3% of the total rural population [6]. Meanwhile, the aging of the agricultural labor force has been repeatedly documented [7]. Note that both the shortage of water and deficiency of effective laborers will hinder the sustainable agricultural development, and thus, it is crucial to adopt some resource-saving technologies to overcome these drawbacks.

Direct seeding has been increasingly adopted in rice production due to its water- and labor-saving characteristics [8]. Note that direct seeding of rice refers to the practice by which farmers directly sow rice seeds into the field without seedling cultivation and transplantation [8], which greatly reduces water and labor inputs. In China, the total rice-sown area was around 30 million hectares (ha) during the past decade, ranking second only to that in India, and the total rice output remained over 200 million tons, accounting for the largest proportion in the world [9]. In recent years, direct seeding has been increasingly adopted in China. For example, the area of rice sown with direct seeding in Heilongjiang, Anhui and Henan provinces in 2017 accounted for 11.8%, 37% and 19.5% of the total rice-sown area in each province, respectively [10]. In the middle reaches of the Yangtze River Basin, the percentage of rice area sown with direct seeding was about 20.5% in 2018 [11]. In addition to China, direct seeding of rice is also widely adopted in other countries, especially in South and the Southeast Asia [8,12,13].

Previous studies from the agronomists have paid considerable attention to the effect of the direct seeding of rice [8]. For example, Tao et al. [14] showed that both the rice yield and water productivity of direct seeding is more than 10% higher than that of transplantation, and direct seeding can also reduce global warming potential by more than 60% in Central China. Xu et al. [15] also concluded that the rice yield of direct seeding is 12% higher than that of transplantation. In Pakistan, Ishfaq et al. [13] also provided evidence that dry direct seeding produces a 13–18% higher yield than transplantation with around 10% less total water input. However, Cabangon et al. [16] found that the rice yield of direct seeding is lower than that of transplantation rice in Malaysia, although it significantly reduces the total water input and increases water productivity. Moreover, Li et al. [17] found that the cumulative methane emissions of direct seeding of rice are 25% higher than that of transplantation. Zhang et al. [18] and Zhou et al. [19] found that direct seeding of rice induces soil erosion, and increases nitrogen and phosphorus runoff losses in China.

The above findings are based on field experiments rather than farmers’ adoption of direct seeding in the context of actual agricultural practices. Note that the effects of farmers’ adoption of direct seeding in the context of actual agricultural practices are not only affected by direct seeding itself but also subject to how farmers manage their rice field after adopting direct seeding. Only a few studies analyzed the effect of farmers’ adoption of direct seeding in the context of actual agricultural practices. Mishra et al. [20] found that farmers’ adoption of direct seeding increases rice yield by 3.7%, and reduces the total cost in rice production by 7.5% in India. Sha et al. [21] showed that rice yield and net rice income among the adopters of direct seeding increase by 3.1% and 106%, respectively, in southern China. However, a recent study from Wang et al. [10] found that the adoption of direct seeding significantly reduces rice yield using survey data of 840 rice farmers.

The effects of farmers’ adoption of direct seeding in the context of actual agricultural practices deserve further analysis. While a few studies mentioned above focused on how farmers’ adoption of direct seeding affects rice yield, production cost and net income, little is known about its effect on fertilizer use. Previous studies have well documented the overuse of fertilizers as well as its adverse environmental effects in China [22,23,24]. China is the largest user of fertilizers, consuming about 30% of the total fertilizers in the world during the past two decades, and the quantity of fertilizer use per unit area of cultivated land is more than three times of the global average [9,25,26]. Rice is the major grain crop in China, and rice production consumes a large quantity of fertilizers. Ji et al. [27] found that about 76% of surveyed farmers overuse fertilizers in rice production. Similar results are also provided by Sun et al. [22]. In addition, van Wesenbeeck et al. [28] found that nitrate and phosphate surpluses can be reduced by over 50% and 75%, respectively, without damaging food self-sufficiency in China. Thus, it has crucial implications when examining the relationship between farmers’ adoption of direct seeding and fertilizer use in addition to the yield in rice production. Using survey data of 1002 rice farmers in China, this study aims to examine how farmers’ adoption of direct seeding affects both the yield and fertilizer use in rice production, addressing the self-selectivity bias arising from both observed and unobserved heterogeneity.

Our contribution is threefold. First, this study may be the first to examine the effects of farmers’ adoption of direct seeding on rice yield and fertilizer use in the context of actual agricultural practices, which adds more in-depth discussion to the effects of the adoption of direct seeding of rice. It should be noted that lack of analysis on the relationship of farmers’ adoption of direct seeding of rice with fertilizer use means that the environmental effects of direct seeding are ignored. Second, we employ the endogenous switching regression model to address the potential self-selectivity issue of farmers’ adoption of direct seeding, which can provide unbiased and consistent results. Note that all farmers are free to self-select whether to adopt direct seeding of rice or not, and their decision to adopt or not is subject to both observed and unobserved heterogeneity. Third, we further analyze the heterogeneity in the effect of farmers’ adoption of direct seeding on fertilizer use in rice production in terms of farm size and location, which has been ignored in previous studies. Note that those heterogeneous analyses have crucial implications for raising effective policies to improve farmers’ adoption of direct seeding in rice production.

The remaining part of this article proceeds as follows. Section 2 develops the conceptual framework and empirical strategy. Section 3 presents the data source and descriptive analysis. The results and discussion are shown in Section 4, including the determinants of farmers’ adoption of direct seeding, and the effects of the adoption of direct seeding and other factors on the yield and fertilizer use in rice production. The robustness checks and heterogeneity analysis are also reported. The final section concludes with policy implications.

2. Materials and Methods

2.1. Conceptual Framework

We develop a conceptual framework to explain farmers’ decision to adopt direct seeding and how their adoption of direct seeding affects the yield and fertilizer use in rice production. Farmers’ adoption of direct seeding of rice may be influenced by observed and unobserved heterogeneity [20,21]. Let U_i¹ and U_i⁰ denote the expected utility of adopting and not adopting direct seeding, respectively. It is reasonable to assume that a farmer would adopt direct seeding when U_i¹ > U_i⁰, and thus, the indicator variable DS_i is defined to equal one. A farmer would not adopt direct seeding when U_i¹ ≤ U_i⁰, and thus, the indicator variable DS_i is defined to equal zero.

Overall, the effect of farmers’ adoption of direct seeding on rice yield is theoretically ambiguous. In other words, farmers’ adoption of direct seeding may have both positive and negative effects on rice yield in the context of actual agricultural practices. On the one hand, the tiller and joint stages of direct-seeded rice are relatively later than those of transplanted rice, and direct seeding may also aggravate weed infestation, soil erosion and runoff losses of nutrients [18,19,29,30]. Due to these factors, rice yield of farmers’ adoption of direct seeding may be lower than that of transplantation. On the other hand, however, direct seeding largely prevents the root systems of rice seedlings being harmed, and has higher planting density and more rice ears [20], which is conducive to the increase in rice yield. In fact, previous studies have not reached a consistent conclusion about the yield effect of farmers’ adoption of direct seeding of rice [10,20,21].

In addition, farmers’ adoption of direct seeding may affect fertilizer use in rice production through three channels. First, direct-seeded rice has shallower root system than transplanted rice [31], which is detrimental to absorbing necessary nutrients from the soil. In this context, farmers may have to increase their use of fertilizers to ensure that direct-seeded rice can absorb sufficient nutrients. Second, direct seeding often induces more severe weed infestation than transplantation [29,30], and more weeds provide more intense competition for nutrients in direct-seeded rice. In this context, farmers would also have to use more fertilizers in rice production to avoid potential yield loss. Third, direct seeding of rice is more likely to cause soil erosion and runoff losses of nutrients [18,19], which would also increase the demand for fertilizers in rice production. To summarize, the hypothesis to be tested in this study is whether farmers’ adoption of direct seeding tends to increase fertilizer use in rice production.

2.2. Endogenous Switching Regression Model

Note that farmers’ decision to adopt direct seeding may be influenced by observed and unobserved heterogeneity [20,21]. The neglect of the self-selectivity bias may produce biased results. While the propensity score method can address the self-selectivity bias arising from observed heterogeneity, it is unable to address the self-selectivity bias arising from unobserved heterogeneity [32]. As an alternative, the endogenous switching regression model has been widely employed to address for the self-selectivity bias arising from both observed and unobserved heterogeneity [33,34]. The endogenous switching regression model has a selection equation and two outcome equations. Specifically, the selection equation describes the factors affecting farmers’ decision to adopt direct seeding. In this context, the selection equation describing the influencing factors of farmers’ decision to adopt direct seeding can be specified as:

$$D{S}_i^{\ast }=\alpha X_i+u_i\mathrm{with}D{S}_i=\{\begin{array}{c}1\mathrm{if}D{S}_i^{\ast }=U_i^1-U_i^0>0\\ 0\mathrm{if}D{S}_i^{\ast }=U_i^1-U_i^0\le 0\end{array},$$

(1)

where DS_i^* denotes the difference in the expected utility between adopting and not adopting direct seeding. X_i denotes a vector of exogenous variables affecting farmers’ decision to adopt direct seeding. u_i ~ N(0, σ_u²) indicates an error term. α is a vector of coefficients to be estimated.

Note that the endogenous switching regression model can explain the yield and fertilizer use for the adopters and non-adopters of direct seeding separately. Two outcome equations are specified as:

$$\mathrm{Regime}1:Y_{1i}={\beta }_1{Z}_i+v_{1i}\mathrm{if}D{S}_i=1,$$

(2)

$$\mathrm{Regime}2:Y_{2i}={\beta }_2{Z}_i+v_{2i}\mathrm{if}D{S}_i=0,$$

(3)

where Y_1i indicates the variable of rice yield and fertilizer use separately. Z_i is a vector of independent variables affecting farmers’ rice yield and fertilizer use. v_1i and v_2i are error terms with zero mean. Both β₁ and β₂ are coefficients to be estimated. Note that the error terms of the selection and outcome equations must follow a trivariate normal distribution [35]. Hence, the covariance matrix is represented as:

$$\mathsf{\Sigma }=\left[\begin{array}{ccc}{\sigma }_u^2& {\sigma }_{u1}& {\sigma }_{u2}\\ {\sigma }_{u1}& {\sigma }_1^2& {\sigma }_{12}\\ {\sigma }_{u2}& {\sigma }_{12}& {\sigma }_2^2\end{array}\right],$$

(4)

where σ_u², σ₁², and σ₂² denote the variances of u_i, v_1i, and v_2i, respectively. σ_u1, σ_u2, and σ₁₂ denote the covariance between u_i and v_1i, covariance between u_i and v_2i, and covariance between v_1i and v_2i, respectively. To examine the effect of farmers’ adoption of direct seeding, the counterfactual situations are constructed for those adopting direct seeding and those not adopting direct seeding. The conditional expected rice yield and fertilizer use are calculated as:

Adopters with adoption of direct seeding (factual situation):

$$E(Y_{1i}|D{S}_i=1)={\beta }_1{Z}_i+{\sigma }_{u1}{\lambda }_{1i},$$

(5)

Adopters without adoption of direct seeding (counterfactual situation):

$$E(Y_{2i}|D{S}_i=1)={\beta }_2{Z}_i+{\sigma }_{u2}{\lambda }_{1i},$$

(6)

Non-adopters with adoption of direct seeding (counterfactual situation):

$$E(Y_{1i}|D{S}_i=0)={\beta }_1{Z}_i+{\sigma }_{u1}{\lambda }_{2i},$$

(7)

Non-adopters without adoption of direct seeding (factual situation):

$$E(Y_{2i}|D{S}_i=0)={\beta }_2{Z}_i+{\sigma }_{u2}{\lambda }_{2i},$$

(8)

where λ_1i and λ_2i are the inverse Mills ratios for addressing the potential self-selectivity bias [36]. Note that these two inverse Mills ratios can be calculated as:

$${\lambda }_{1i}=\frac{E(u_i>-\alpha X_i)}{{\sigma }_{u1}}=\frac{\phi (\alpha X_i)}{\mathsf{\Phi }(\alpha X_i)},$$

(9)

$${\lambda }_{2i}=\frac{E(u_i\le -\alpha X_i)}{{\sigma }_{u2}}=\frac{-\phi (\alpha X_i)}{1-\mathsf{\Phi }(\alpha X_i)},$$

(10)

where φ(•) indicates the probability density function of the standard normal distribution, and Φ(•) indicates the cumulative distribution function of the standard normal distribution [33]. Using the conditional expected rice yield and fertilizer use in four situations, we further calculate the average treatment effects on the treated (ATT) for the adopters of direct seeding, and the average treatment effects on the untreated (ATU) for the non-adopters as:

$$\mathrm{ATT}=({\beta }_1-{\beta }_2)Z_i+({\sigma }_{u1}-{\sigma }_{u2}){\lambda }_{1i},$$

(11)

$$\mathrm{ATU}=({\beta }_1-{\beta }_2)Z_i+({\sigma }_{u1}-{\sigma }_{u2}){\lambda }_{2i}$$

(12)

To estimate the endogenous switching regression model, a valid instrumental variable included in X_i but not included in Z_i is required. The instrumental variable should affect farmers’ decision to adopt direct seeding, but not have direct effect on rice yield and fertilizer use except through its effect on farmers’ adoption of direct seeding. Following the previous studies [36,37,38,39], we construct the instrumental variable from a peer-effect perspective. We reasonably assume that whether a farmer adopts direct seeding or not is likely to be affected by his neighbors’ adoption of direct seeding, and his neighbors’ adoption of direct seeding will not directly affect the farmer’s rice yield and fertilizer use. Thus, the village-level adoption rate of direct seeding except the farmer themself is used as the instrumental variable in this study.

2.3. Data and Descriptive Analysis

In this study, data were collected from a cross-sectional survey of rice farmers in the middle and lower reaches of the Yangtze River Basin in China. It should be noted that the middle and lower reaches of the Yangtze River Basin is one of the most important major rice-producing regions where both the sown area and output of rice account for more than 50% of the totals in China [40]. In particular, the area sown with direct seeding in the middle and lower reaches of the Yangtze River Basin accounts for about 75% of the total area sown with direct seeding in China. The survey was conducted in Hubei, Jiangxi and Jiangsu provinces in the region. We used a multistage sampling procedure to choose the surveyed farmers. Specifically, we randomly chose a rice-producing county in each province, and then randomly chose four townships in each sampled county. In each sampled township, four villages were randomly chosen, and in each sampled village, about 20–25 rice farmers were chosen following a random sampling method. After excluding a few farmers who failed to provide complete information, a total of 1002 rice farmers remained in this study. It should be noted that sample size in this study was mainly subject to the research fund. Nevertheless, our sample size is appropriate, comparable with or even larger than that in many previous studies [20,21].

A cross-sectional survey was conducted in 2018 to collect information. We used a multi-criteria decision-making method to design the survey questionnaire. While the survey collected a wide range of information, this study used five parts of survey data, including the yield, fertilizer use, pesticide use, and labor input in rice production, adoption of direct seeding, individual and household characteristics, planting characteristics, and prices of fertilizers and rice. All enumerators, consisting of undergraduates and graduate students from several universities in China, were recruited to conduct the survey. These enumerators were about 20 years old. To ensure their qualification and the accuracy of survey data, we trained them prior to the survey with a focus on how to conduct an effective farmer survey.

In this study, we utilized different indicators to measure farmers’ fertilizer use in rice production. We calculated the quantity of fertilizers used per unit area as an indicator measuring farmers’ fertilizer use. In addition, since nitrogen nutrients generally account for the largest proportion of total fertilizers used by farmers [22], we also calculated the quantity of nitrogen nutrients used per unit area as an alternative indicator.

Table 1. Definition and descriptive statistics of variables.

Variable	Definition or Description	Mean	SD
Yield	Quantity of rice output per ha (tonnes/ha)	9.16	1.64
Fertilizer use	Quantity of total fertilizers per ha (kg/ha)	426.62	177.24
Nitrogen use	Quantity of nitrogen nutrients per ha (kg/ha)	247.38	129.87
Adoption of direct seeding	1 if farmer adopts direct seeding, 0 otherwise	0.79	0.41
Age	Age of farmer (year)	56.06	8.11
Education	Formal education of farmer (year)	6.94	3.49
Cadre	1 if farmer is a village leader, 0 otherwise	0.19	0.39
Laborers	Number of household laborers (person)	3.19	1.14
Assets	Total household fixed assets (1000 CNY)	401.95	504.53
Farm size	Sown area of rice (ha)	2.93	9.57
Hybrid	1 if farmer adopts hybrid seed, 0 otherwise	0.63	0.48
Late	1 if farmer plants late-season rice, 0 otherwise	0.08	0.27
Training	1 if farmer attends technical training, 0 otherwise	0.22	0.41
Price ratio	Price ratio of rice-to-fertilizer	46.93	10.58
Labor input	Labor input per ha (days/ha)	265.92	222.47
Pesticide use	Quantity of active ingredients of pesticide per ha (kg/ha)	2.50	2.23
IV	Village-level adoption rate of direct seeding except the farmer themself	0.75	0.22

Notes: The IV equals the ratio of the number of adopters of direct seeding except the farmer themself in a village to the total number of surveyed farmers in the same village. As a result, the value of IV is the same for each farmer in a village. CNY denotes Chinese Yuan, and SD denotes standard deviations. Authors’ survey.

presents the definition and descriptive statistics of main variables.

Table 2. Mean differences between the adopters and non-adopters of direct seeding.

Variable	Adopters (n = 793)	Non-Adopters (n = 209)	Mean Difference
Yield	9.14 (1.59)	9.25 (1.80)	0.11
Fertilizer use	439.39 (183.86)	378.19 (139.59)	61.20 ***
Nitrogen use	257.59 (133.68)	208.64 (105.90)	48.96 ***
Age	55.87 (8.20)	56.79 (7.71)	−0.92
Education	6.95 (3.51)	6.92 (3.44)	0.02
Village leader	0.19 (0.39)	0.19 (0.39)	−0.00
Laborers	3.13 (1.14)	3.39 (1.13)	−0.25 ***
Assets	403.15 (497.17)	397.38 (532.74)	5.76
Farm size	3.34 (10.70)	1.34 (1.22)	2.00 ***
Hybrid	0.58 (0.49)	0.82 (0.38)	−0.24 ***
Late	0.07 (0.25)	0.13 (0.34)	−0.07 ***
Training	0.22 (0.42)	0.19 (0.39)	0.03
Price ratio	47.70 (10.87)	44.00 (8.85)	3.70 ***
Labor input	264.82 (230.86)	270.12 (187.70)	−5.31
Pesticide use	2.65 (2.36)	1.91 (1.45)	0.75 ***
IV	0.81 (0.16)	0.53 (0.28)	0.29 ***

Notes: Standard deviations are in parentheses. *** p < 0.01.

shows the mean differences between the adopters and non-adopters. The comparison of the mean difference between the adopters and non-adopters of direct seeding shows significant heterogeneity in many variables. For example, farmers adopting direct seeding use more fertilizers and nitrogen nutrients per unit area than those not adopting direct seeding. In terms of other variables, the adopters of direct seeding have fewer household laborers and larger farm size of rice production. Farmers adopting direct seeding are less likely to choose hybrid seed, and plant late-season rice. Moreover, the price ratio of rice-to-fertilizer among the adopters is significantly higher than that among the non-adopters. Note that the significant difference in those variables implies the presence of potential self-selectivity issue of farmers’ adoption of direct seeding.

3. Results and Discussion

The results of the endogenous switching regression models estimated using the full information maximum likelihood method are presented in

Table 3. Estimated results of the effects on the yield and fertilizer use in rice production.

Variables	Adoption of Direct Seeding	Ln (Yield)		Direct Seeding	Ln (Fertilizer Use)
		Adopters	Non-Adopters		Adopters	Non-Adopters
Age	−0.01	−0.002 ***	−0.00	−0.00	0.00	0.00
	(0.01)	(0.00)	(0.00)	(0.01)	(0.00)	(0.00)
Education	0.02	0.00	−0.01 **	0.02	−0.00	−0.01
	(0.02)	(0.00)	(0.00)	(0.02)	(0.00)	(0.01)
Village leader	−0.07	0.01	−0.03 ***	−0.03	−0.05	−0.13 **
	(0.06)	(0.01)	(0.00)	(0.06)	(0.03)	(0.06)
Laborers	0.01	−0.01	−0.00	0.02	−0.01	−0.02
	(0.02)	(0.01)	(0.00)	(0.03)	(0.01)	(0.02)
Ln(Assets)	0.01	0.02 ***	−0.00	−0.00	0.01	0.01 ***
	(0.05)	(0.01)	(0.02)	(0.07)	(0.01)	(0.00)
Ln(Farm size)	0.18 ***	−0.01 **	−0.06 **	0.18 ***	0.02	−0.04
	(0.05)	(0.00)	(0.03)	(0.05)	(0.02)	(0.06)
Hybrid	−0.42 ***	0.03 ***	−0.02	−0.38 ***	0.04	0.13 **
	(0.09)	(0.00)	(0.02)	(0.09)	(0.04)	(0.06)
Late	−0.32 ***	−0.02 ***	−0.11 ***	−0.34 ***	0.05 ***	−0.08 ***
	(0.02)	(0.00)	(0.00)	(0.02)	(0.01)	(0.01)
Training	−0.14	0.01	−0.01	−0.16	0.04	−0.09 *
	(0.10)	(0.01)	(0.01)	(0.14)	(0.03)	(0.05)
Price ratio	0.004 *	0.0002 *	−0.003 ***	0.002 *	0.003 **	0.01 ***
	(0.00)	(0.00)	(0.00)	(0.00)	(0.00)	(0.00)
IV	2.52 ***			2.51 ***
	(0.17)			(0.02)
Township dummies	Yes	Yes	Yes	Yes	Yes	Yes
Constant	−0.62	9.15 ***	9.47 ***	−0.61	6.04 ***	5.58 ***
	(1.02)	(0.02)	(0.16)	(0.96)	(0.12)	(0.17)
Independent equations (χ2)	70.59 ***			26.00 ***
ρu1		0.04			−0.55 ***
		(0.08)			(0.09)
ρu2			−0.13			−0.32 **
			(0.31)			(0.14)
n		1002			1002

Notes: Standard errors clustered at the provincial level are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01.

. The estimated χ² statistics for both the yield and fertilizer use models are positive and significant at the 1% level, which indicates that farmers’ decision to adopt direct seeding are correlated with the yield and fertilizer use in rice production. Note that ρ_u1 = σ_u1/(σ_uσ₁) denotes the correlation coefficient between u_i and v_1i, and ρ_u2 = σ_u2/(σ_uσ₂) denotes the correlation coefficient between u_i and v_2i. While neither ρ_u1 nor ρ_u2 is significant for the yield model, both ρ_u1 and ρ_u2 are negative and significant for the fertilizer use model, indicating a negative selection in the adopters and a positive selection in the non-adopters for fertilizer use in rice production. These results illustrate the fact that employing the endogenous switching regression models to address the self-selectivity bias is appropriate.

To verify the validity of the instrumental variable, we conducted falsification tests following Di Falco et al. [41], Kumar et al. [42], and Liu et al. [43]. The estimated results of the falsification tests on the instrumental variable are presented in

Table 4. Results of the falsification test for the instrumental variable.

Variable	Adoption of Direct Seeding	Ln (Yield)	Ln (Yield)	Ln (Fertilizer Use)	Ln (Nitrogen Use)
IV	2.52 ***	0.02	0.01	0.21	0.24
	(0.16)	(0.06)	(0.06)	(0.12)	(0.17)
Age	−0.01	−0.00	−0.00	0.00	−0.00
	(0.01)	(0.00)	(0.00)	(0.00)	(0.00)
Education	0.02	−0.00	−0.00	−0.01	−0.01
	(0.02)	(0.00)	(0.00)	(0.01)	(0.01)
Village leader	−0.07	−0.03 **	−0.03	−0.14	−0.16
	(0.06)	(0.00)	(0.01)	(0.06)	(0.07)
Laborers	0.01	−0.00	0.00	−0.02	−0.03
	(0.02)	(0.00)	(0.00)	(0.02)	(0.02)
Ln(Assets)	0.01	−0.00	−0.01	0.01	0.00
	(0.05)	(0.02)	(0.02)	(0.01)	(0.02)
Ln(Farm size)	0.18 ***	−0.06	−0.06	−0.02	−0.02
	(0.05)	(0.03)	(0.03)	(0.07)	(0.08)
Hybrid	−0.42 ***	−0.03	−0.04	0.10	0.08
	(0.09)	(0.04)	(0.03)	(0.05)	(0.03)
Late	−0.32 ***	−0.11 ***	−0.11 **	−0.10 **	−0.10 **
	(0.02)	(0.01)	(0.01)	(0.02)	(0.02)
Training	−0.14	−0.01	−0.01	−0.10	−0.14
	(0.11)	(0.01)	(0.01)	(0.06)	(0.06)
Price ratio	0.004 *	−0.003 *	−0.003 *	0.01 **	0.01 **
	(0.00)	(0.00)	(0.00)	(0.00)	(0.00)
Ln(Labor input)			0.04
			(0.04)
Ln(Pesticide use)			0.02
			(0.01)
Ln(Fertilizer use)			−0.01
			(0.03)
Township dummies	Yes	Yes	Yes	Yes	Yes
Constant	−0.62	9.48 ***	9.29 ***	5.59 ***	5.59 ***
	(1.02)	(0.16)	(0.46)	(0.11)	(0.24)
Pseudo/Adjusted R2	0.28	0.38	0.38	0.29	0.45
n	1002	209	209	209	209

Notes: Standard errors clustered at the provincial level are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01.

. It provides evidence that the instrumental variable has a significant relationship with farmers’ adoption of direct seeding, but has no significant relationship with the yield and fertilizer use in rice production, confirming the validity of the instrumental variable.

3.1. Determinants of Adoption of Direct Seeding

The estimated results of the selection equations for the yield and fertilizer use are highly consistent (Table 3). The coefficient of logarithmic farm size is significant and positive, indicating that farmers with larger farm size are more likely to adopt direct seeding. The significant and negative coefficient of hybrid seeding implies that farmers who adopt hybrid seeding are less likely to adopt direct seeding. The coefficient of late-season rice is significant and negative, indicating that farmers planting late-season rice have a larger probability of adopting direct seeding. The price ratio of rice-to-fertilizer has a positive relationship with farmers’ adoption of direct seeding. In addition, the coefficient of instrumental variable is significant and positive, illustrating that a farmer is more likely to adopt direct seeding when there is a larger adoption rate of direct seeding in his village.

3.2. Average Treatment Effects of Adoption of Direct Seeding on Rice Yield and Fertilizer Use

The estimated ATT and ATU of farmers’ adoption of direct seeding on the yield and fertilizer use in rice production are presented in

Table 5. Average treatment effects of the adoption of direct seeding in rice production.

Groups	Mean Outcomes		Treatment Effects	Change (%)
	To Adopt	Not to Adopt
Yield (tonnes/ha)
Adopters	9.04 (0.92)	8.72 (1.34)	ATT: 0.32 ***	3.65
Non-adopters	9.29 (0.94)	9.15 (1.17)	ATU: 0.14 **	1.54
Fertilizer use (kg/ha)
Adopters	413.74 (73.40)	345.24 (81.84)	ATT: 68.50 ***	19.84
Non-adopters	500.05 (115.99)	363.02 (82.71)	ATU: 137.03 ***	37.75

Notes: ATT and ATU denote the average treatment effects on the adopters and non-adopters, respectively. Standard deviations are presented in parentheses. ** p < 0.05, *** p < 0.01.

. Note that the ATT and ATU calculated based on the endogenous switching regression models addressed the self-selectivity bias arising from observed and unobserved heterogeneity. It has been shown that the adoption of direct seeding significantly increases the yield and fertilizer use in rice production. Specifically, the adoption of direct seeding increases rice yield and fertilizer use among the adopters by 0.32 tonnes/ha (3.65%) and 68.50 kg/ha (19.84%), respectively. Meanwhile, rice yield and fertilizer use among the non-adopters would have also increased by 0.14 tonnes/ha (1.54%) and 137.03 kg/ha (37.75%), respectively, if they had adopted direct seeding. These results confirm that farmers’ adoption of direct seeding increases the yield and fertilizer use in rice production, which is logically consistent with Ishfaq et al. [13], Tao et al. [14], Xu et al. [15], Zhang et al. [18], and Zhou et al. [19], who found that direct seeding of rice has a significantly positive effect on rice yield, but increases soil erosion and runoff losses of nitrogen and phosphorus nutrients.

3.3. Effect of Other Factors on Rice Yield and Fertilizer Use

The estimated results presented in Table 3 also show the effect of other factors on rice yield. Overall, factors affecting rice yield among the adopters and non-adopters of direct seeding are different. Among the adopters of direct seeding, each yearly increase in farmers’ age is associated with a 0.2% reduction in rice yield. Farmers’ education fails to exert any significant effect on rice yield among the adopters of direct seeding, but has a negative effect on that among the non-adopters of direct seeding. Compared with that of village leaders among the non-adopters of direct seeding, the rice yield of village leaders among the adopters of direct seeding is 3% lower. In addition, each 1% increase in farmers’ total household fixed assets induces a 0.02% increase in rice yield among the adopters of direct seeding. Moreover, there is a negative relationship between farm size and rice yield. Specifically, each 1% increase in farm size is associated with a 0.01% reduction in rice yield among the adopters of direct seeding and a 0.06 reduction in rice yield among the non-adopters. Compared with that of non-hybrid seeding, the yield of hybrid seeding is 3% higher among the adopters of direct seeding. The yield of late-season rice is 2% and 11% lower among the adopters and non-adopters of direct seeding, respectively. Each unit increase in the price ratio of rice-to-fertilizer increases rice yield among the adopters of direct seeding by 0.02%, but reduces that among the non-adopters by 0.3%.

Table 3 also reports the estimated effect of other factors on fertilizer use in rice production. The village leaders not adopting direct seeding use 13% fewer fertilizers in rice production than common non-adopters. Each 1% increase in farmers’ total household fixed assets induces a 0.01% increase in fertilizer use among the non-adopters of direct seeding, confirming that a higher level of household wealth can enable farmers to buy more fertilizers, as argued by Tefera et al. [44] and Guo et al. [45]. As for the non-adopters of direct seeding, fertilizer use in rice production when farmers adopt hybrid seeding is 13% higher than that when farmers adopt non-hybrid seeding. Moreover, the adopters of direct seeding use 5% more fertilizers in the production of late-season rice, while the non-adopters of direct seeding use 8% fewer fertilizers in the production of late-season rice. Participation in technical training activities is found to reduce the quantity of fertilizers per unit area among the non-adopters of direct seeding by 9%, which is consistent with the findings of Huang et al. [46] and Pan et al. [47]. Each unit increase in the price ratio of rice-to-fertilizer increases fertilizer use by 0.3% among the adopters of direct seeding, and by 1% among the non-adopters of direct seeding. This finding is reasonable because higher fertilizer price may lead farmers to reduce fertilizer use, while higher rice price may encourage farmers to increase fertilizer use.

3.4. Robustness Checks

We check the robustness of the main results by adding variables of factor inputs for the yield model, and replacing fertilizer use with nitrogen use for the fertilizer use model. The estimated results of the endogenous switching regression models are presented in

Table 6. Estimated results for the robustness checks.

Variables	Adoption of Direct Seeding	Ln (Yield)		Direct Seeding	Ln (Nitrogen Use)
		Adopters	Non-Adopters		Adopters	Non-Adopters
Age	−0.00	−0.002 ***	−0.00	−0.00	0.00	−0.00
	(0.01)	(0.00)	(0.00)	(0.01)	(0.00)	(0.00)
Education	0.02	0.00	−0.00	0.02	−0.00	−0.01
	(0.02)	(0.00)	(0.00)	(0.02)	(0.00)	(0.01)
Village leader	−0.08 ***	0.01	−0.03 **	−0.05	−0.03	−0.15 **
	(0.02)	(0.01)	(0.01)	(0.07)	(0.04)	(0.07)
Laborers	0.01	−0.01	0.00	0.02	−0.00	−0.03
	(0.03)	(0.01)	(0.00)	(0.02)	(0.00)	(0.02)
Ln(Assets)	0.01	0.02 ***	−0.01	0.01	−0.00	0.00
	(0.05)	(0.01)	(0.02)	(0.06)	(0.01)	(0.02)
Ln(Farm size)	0.17 **	−0.01	−0.06 **	0.16 ***	0.01	−0.04
	(0.07)	(0.01)	(0.03)	(0.04)	(0.02)	(0.06)
Hybrid	−0.42 ***	0.03 ***	−0.04 **	−0.38 ***	0.06	0.11 **
	(0.06)	(0.00)	(0.02)	(0.09)	(0.04)	(0.05)
Late	−0.31 ***	−0.02 ***	−0.10 ***	−0.37 ***	0.11 ***	−0.07 ***
	(0.05)	(0.01)	(0.01)	(0.03)	(0.02)	(0.02)
Training	−0.14	0.01	−0.01	−0.16	0.02	−0.12 **
	(0.12)	(0.01)	(0.01)	(0.15)	(0.04)	(0.05)
Price ratio	0.00	0.0002 **	−0.003 ***	0.002 *	0.003 ***	0.01 ***
	(0.00)	(0.00)	(0.00)	(0.00)	(0.00)	(0.00)
Ln(Labor input)	−0.05	0.01	−0.01
	(0.25)	(0.01)	(0.03)
Ln(Pesticide use)	0.08 ***	−0.01 **	0.02
	(0.02)	(0.00)	(0.01)
Ln(Fertilizer use)	0.04	0.00	0.04
	(0.18)	(0.02)	(0.03)
IV	2.49 ***			2.41 ***
	(0.15)			(0.03)
Township dummies	Yes	Yes	Yes	Yes	Yes	Yes
Constant	−0.68	9.12 ***	9.28 ***	−0.66	5.65 ***	5.58 ***
	(3.06)	(0.15)	(0.42)	(0.97)	(0.04)	(0.22)
Independent equations (χ2)	157.51 ***			62.52 ***
ρu1		0.03			−0.65 ***
		(0.08)			(0.08)
ρu2			−0.08			−0.34 *
			(0.31)			(0.17)
n		1002			1002

Notes: Standard errors clustered at the provincial level are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01.

, and the calculated ATT and ATU are presented in

Table 7. Average treatment effects for the robustness checks.

Groups	Mean Outcomes		Treatment Effects	Change (%)
	To Adopt	Not to Adopt
Yield (tonnes/ha)
Adopters	9.04 (0.93)	8.87 (1.37)	ATT: 0.17 ***	1.93
Non-adopters	9.35 (0.96)	9.15 (1.18)	ATU: 0.19 ***	2.09
Nitrogen use (kg/ha)
Adopters	239.88 (80.00)	207.08 (87.52)	ATT: 32.80 ***	15.84
Non-adopters	297.76 (97.80)	199.52 (79.77)	ATU: 98.24 ***	49.24

Notes: ATT and ATU denote the average treatment effects on the adopters and non-adopters, respectively. Standard deviations are presented in parentheses. *** p < 0.01.

. As for the adopters, the adoption of direct seeding increases the yield and nitrogen use in rice production by 0.17 tonnes/ha (1.93%) and 32.80 kg/ha (15.84%), respectively. By comparison, the positive ATU show that the positive effect of farmers’ adoption of direct seeding on rice yield and the quantity of nitrogen nutrients is apparently larger among the non-adopters than the adopters of direct seeding. If the non-adopters had adopted direct seeding in rice production, the yield and nitrogen use would have increased by 0.19 tonnes/ha (2.09) and 98.24 kg/ha (49.24%), respectively. Thus, the main results about the positive effect of farmers’ adoption of direct seeding on the yield and fertilizer use in rice production are highly robust.

3.5. Heterogeneity Analysis

To improve the understanding of the effects of the adoption of direct seeding on rice yield and fertilizer use for different farmers, we further conducted several kinds of heterogeneity analysis. Specifically, we classified the surveyed farmers into different groups in terms of their farm size and location. Since the endogenous switching regression models for several sub-samples fail to converge due to insufficient observations after classifying the surveyed farmers into different groups, we employ the treatment effects model to estimate the heterogeneity in the effects of farmers’ adoption of direct seeding on rice yield and fertilizer use. Note that the treatment effects model can also effectively address the self-selectivity issue of direct seeding arising from observed and unobserved heterogeneity [48].

The estimated results of the treatment effects models are presented in

Table 8. Estimated results for the heterogeneity analysis.

Variables	Total Size		Province
	>2 ha	≤2 ha	Hubei	Jiangsu	Jiangxi
Ln(Yield)
Adoption of direct seeding	0.12 ***	0.02	0.07	−0.16 ***	0.13 **
	(0.03)	(0.05)	(0.06)	(0.05)	(0.07)
Control variables	Yes	Yes	Yes	Yes	Yes
Township dummies	Yes	Yes	Yes	Yes	Yes
Constant	9.17 ***	9.16 ***	9.20 ***	9.40 ***	8.99 ***
	(0.10)	(0.06)	(0.12)	(0.09)	(0.11)
Independent equations (χ2)	0.14	0.18	0.21	7.80 ***	2.24
ρ	−0.10	−0.05	−0.09	0.44 ***	−0.42
n	244	758	350	326	326
Ln(Fertilizer use)
Adoption of direct seeding	0.53 ***	0.25	0.04	−0.26 ***	0.43 ***
	(0.17)	(0.19)	(0.10)	(0.10)	(0.08)
Control variables	Yes	Yes	Yes	Yes	Yes
Township dummies	Yes	Yes	Yes	Yes	Yes
Constant	5.32 ***	5.79 ***	5.51 ***	6.52 ***	5.59 ***
	(0.45)	(0.17)	(0.17)	(0.24)	(0.27)
Independent equations (χ2)	4.44 **	2.15	0.43	3.24 *	21.05 ***
ρ	−0.77 **	−0.45	−0.14	0.25 *	−0.49 ***
n	244	758	350	326	326

Notes: Only the estimated results of the outcome equations for the treatment effects models are reported. Control variables and township dummies are the same as shown in Table 3. Standard errors clustered at the provincial level are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01.

. Overall, the results show that the adoption of direct seeding exerts heterogeneous effects on rice yield among farmers with different farm sizes. Specifically, the adoption of direct seeding increases rice yield among farmers with farm size larger than 2 ha by 12%, but fails to exert significant effect on rice yield among farmers with farm size not larger than 2 ha. In addition, the results also show that the effects of farmers’ adoption of direct seeding on rice yield differ across provinces. While the adoption of direct seeding significantly causes a 16% reduction in rice yield among farmers in Jiangsu, it induces a 13% increase in rice yield among farmers in Jiangxi. Meanwhile, the adoption of direct seeding seems not to result in a significant effect on rice yield among farmers in Hubei.

Table 8 also shows heterogeneous effects of farmers’ adoption of direct seeding on fertilizer use in rice production. Specifically, the adoption of direct seeding increases the quantity of fertilizers used in rice production among farmers with farm size larger than 2 ha by 53%, but has no significant effect on fertilizer use in rice production among farmers with farm size not larger than 2 ha. Note that smaller farm size tends to increase the probability of precise tillage and farmland management, which can effectively reduce soil erosion and runoff losses of fertilizers associated with direct seeding. As for the surveyed farmers in different provinces, while the adoption of direct seeding causes a significantly negative effect on fertilizer use among farmers in Jiangsu, its effect on fertilizer use among farmers in Jiangxi is significantly positive. Meanwhile, the adoption of direct seeding is not found to exert any significant effect on fertilizer use among farmers in Hubei. Note that the heterogeneous effects of farmers’ adoption of direct seeding on fertilizer use in terms of their location are related to different natural conditions in different provinces.

4. Conclusions and Policy Implications

This study investigates the effects of farmers’ adoption of direct seeding on the yield and fertilizer use in rice production using survey data of 1002 rice farmers in the middle and lower reaches of the Yangtze River Basin in China. The endogenous switching regression model is employed to address the self-selectivity bias arising from observed and unobserved heterogeneity. The treatment effects model is utilized for heterogeneity analysis in terms of farmers’ farm size and location.

The results show that 79% of surveyed farmers adopt direct seeding in rice production. Farmers’ adoption of direct seeding in rice production is positively affected by their farm size and the price ratio of rice-to-fertilizer, but negatively affected by their adoption of hybrid seeding and production of late-season rice. The average treatment effects on the treated (i.e., the adopters of direct seeding) show that the adoption of direct seeding increases the yield and fertilizer use in rice production by 3.65% and 19.84%, respectively. Meanwhile, the yield and fertilizer use in rice production would have been 1.54% and 37.75% higher, respectively, if direct seeding had been adopted by the non-adopters. These findings are confirmed when more variables of factor inputs were added into the yield model, and fertilizer use was replaced by nitrogen use. Moreover, this study also shows considerable heterogeneity in the effect of the adoption of direct seeding on the yield and fertilizer use in rice production in terms of farmers’ farm size and location.

Our findings have several important policy implications. First, both the economic and environmental effects of farmers’ adoption of direct seeding should be taken into account when spreading direct seeding in rice production. While previous studies showed that direct seeding could reduce production cost and increase rice yield [8,13,14,15], we found that farmers’ adoption of direct seeding not only increases rice yield but also increases fertilizer use in rice production in the context of actual agricultural practices. Note that China has been making great efforts to promulgate the Action Plan for the Zero-Growth in Fertilizer Use since 2015, and thus, it is crucial to acknowledge the positive effect of farmers’ adoption of direct seeding on fertilizer use as well as its adverse environmental effects. Second, the spread of direct seeding should also meet the specific needs of different types of farmers. The heterogeneity analysis shows that the adoption of direct seeding significantly increases rice yield and fertilizer use among larger-farm-size farmers rather than smaller-farm-size farmers. Note that physical conditions make it harder for larger-farm-size farmers to conduct precise tillage and farmland management, and thus, larger-farm-size farmers tend to increase fertilizer use to overcome the shortcomings of direct seeding when they want to obtain a higher rice yield. Third, our findings also have implications for other developing countries, especially those in South and Southeast Asia. In recent years, direct seeding of rice has been increasingly adopted in these regions [8,12,13]. Thus, our findings may provide some valuable reference points. Note that different countries and regions have specific conditions and needs, and thus, wider and more in-depth evaluation of both the economic and environmental effects of direct seeding in these countries and regions are warranted.