Implications of Non-Farm Work for Clean Energy Adoption: Evidence from Rural China

Abstract

Due to China’s socioeconomic development, labor force transfer from rural areas has become more common, the income of rural households has increased, and the structure of rural household clean living energy consumption has changed. However, few studies have explored the correlation between non-farm employment and clean energy adoption in rural households. Using survey data from 1175 farmers in 106 villages from a 2018 Survey in Liaoning Province, this study uses a Probit model to analyze the effect of non-farm work on clean energy adoption, as well as an effect decomposition model to examine the specific mechanism of their interaction. Robustness tests were performed using extended regression models (ERMs), propensity score matching (PSM), and variation of the core explanatory variable measures. The results found that: (1) Rural residents’ non-farm work has a significant positive effect on their household clean energy adoption. (2) Increasing rural residents’ household income and promoting the growth of their health knowledge are the main channels through which non-farm work influences their clean energy adoption. (3) Non-farm work has a more positive impact on household clean energy adoption for young or male farmers, those who had a junior high school education or above, and those who had a village head in the family. This study provides an understanding of rural non-farm work and clean energy adoption decisions and provides references for the effective allocation of rural labor resources and the formulation of policies related to rural energy adoption.

1. Introduction

As of 2019, more than 40% of the world population relied primarily on polluting fuels such as biomass, charcoal and coal for cooking [1]. There are significant differences globally between polluting fuel-use in rural and urban populations, with 65.3% of total rural populations and 19.3% of total urban populations using polluting fuels. The same is true for China, the biggest energy consumer in the world. According to a data bulletin from China’s third National Agricultural Census, 45% (103 million) of rural households in China still use firewood, straw, and other biomass as their main cooking fuel. Firewood and coal also account for a higher proportion of living energy fuel sources used. Household solid-fuel (biomass, coal) burning contributes to climate change and is a leading health risk factor [2]. Air pollution from solid-fuel stoves contributes to an estimated 2.8 million premature deaths annually and influences regional and global air quality [3,4,5].

Clean energy adoption is intimately linked to a number of the Sustainable Development Goals (SDGs) that were adopted by all United Nations Member States in 2015. It is particularly relevant to SDG number seven (“Ensure access to affordable, reliable, sustainable and modern energy for all”) and SDG number three (“Ensure healthy lives and promote well-being for all at all ages” (More specifically, SDG 7.1.2 is to increase the proportion of the population with primary reliance on clean fuels and technology by 2030; and SDG 3.9.1 is to substantially reduce mortality rates attributed to household and ambient air pollution by 2030 (UNSD, 2018; Rosenthal et al., 2018) [6,7,8]). The usage of cleaner energy, especially in rural areas, is key to the achievement of SDGs 7 and 3.

Both official and non-official organizations in many developing countries, including China, have been working to address the heavy dependence on polluting fuels in rural areas. Many different strategies have been adopted to this end, for example the integrated rural energy-policy [9], the rural liquefied petroleum gas (LPG) promotion program [10,11], the west African gas pipeline (WAGP) project [12], the Coal-to-gas policy [13], the credit policy [14], and so on. Although these strategies have achieved varying degrees of success, dependence on polluting energy in rural areas remains a serious issue.

Rural residents’ fuel choices are an “old problem” [15]. Existing literature has evaluated the influence of demographic and domestic characteristics (e.g., age, gender, education, household size, and dwelling characteristics), socio-economic factors, local environmental factors (e.g., geography, climate), energy availability, and external factors (e.g., policies, fuel prices, fuel supply) on clean fuel adoption [16,17,18,19]. These can vary in magnitude and direction across different settings [20]. Though academia has paid much attention to rural residents’ energy adoption in China [2,13,21,22,23], few studies have focused on their energy adoption caused by their non-farm work. In recent decades, a large number of rural laborers have entered non-agricultural employment as the rate of urbanization in China has increased. This has been regarded as a remarkable phenomenon in China [24]. Existing literature on non-farm work mainly focuses on its impact on household incomes, poverty, and food security in developing countries [25,26,27,28]. Literature in this field focusing on rural residents’ cleaner energy adoption remains scant. In the existing literature, scholars have made some research findings about non-farm work and factors influencing household energy choices, and have proposed many findings with reference value, which provide a basis for this study. The energy ladder hypothesis suggests that as the economy develops and income increases, household domestic energy use gradually transitions toward clean, efficient and modern energy. According to this theory, income is the most direct and important factor influencing household energy use. In recent years, studies have provided strong evidence to support this theory [22,29]. A study by Rahut et al. [30] found that non-farm work by household heads has a positive and significant effect on per capita gas and electricity expenditures in Bhutan. Ma et al. [22] found that income effects are positive for clean fuel gas and electricity, but not for dirty solid fuels such as coal and biomass. Ma et al. [31] showed that non-farm income significantly increased electricity and gas expenditures, while decreasing coal expenditures among rural households. In contrast, Liao et al. [32] found that if the household head is involved in non-farm work, their use of firewood will decrease by 14–21%, but the effect of income is slightly smaller. Shi et al. [33] found that increased non-farm work opportunities do not promote a transition in rural energy use in poor areas, using data from a remote village in Jiangxi. Despite the strong relationship between non-farm work and energy consumption patterns, few studies have examined the direct impact of non-farm work on household clean energy choices. On this basis, this study uses data from a 2018 sample of 1175 farm households in 106 villages in 12 prefecture-level cities (or prefectures) in Liaoning Province, China to develop a Probit model from the micro perspective of non-farm work to explore the impact of non-farm work on rural household clean energy adoption and further analyze its specific mechanisms. This deepens understanding of rural non-farm work and household clean energy adoption decisions, which in turn provides a reference for rational allocation of labor resources and formulation of rural energy policies.

Compared with previous studies, the possible marginal contributions of this study are as follows. First, previous studies have focused on the impact of non-farm work on clean energy adoption, but these studies have not deeply dissected the interactions between the two. The present study examines the mechanisms by which non-farm work affects rural residents’ clean domestic energy choices and uses the methods of Gelbach [34] and Gong et al. [35] to quantify the importance of each channel. Second, the study dealt with possible endogeneity issues using a combination of extended regression models (ERMs), PSM methods, and instrumental variable methods. Finally, the conclusions drawn in this paper can be used as a reference in policy formulation to address dependence on polluting energy sources and to promote clean energy adoption among rural residents in China. In addition, they provide a framework for other developing countries to develop relevant policies.

The paper is organized as follows. Section 2 provides an overview of the data sources, descriptive statistics, and econometric models. Section 3 presents the main empirical results, including baseline regression results and robustness test results. The process and results of the mechanism analysis are presented in Section 4. Section 5 includes the heterogeneity analysis, while the final section summarizes the conclusions of the study, presents policy recommendations, and discusses follow-up research possibilities.

2. Methodology

2.1. Data Description

Data was taken from in-person interviews of rural households in Liaoning Province, China conducted in December 2018. A stratified random sampling method—Probability-Proportional-to-Size Sampling (PPS) study was used in the survey. The sampling followed standard practices. The specific process can be described as follows: (1) The sample counties (or districts or county-level cities) were determined. Counties were then divided into three categories based on their per capita GDP in 2017: high, medium, and low. Four counties were randomly selected from each category (Figure 1). (2) The same method of stratified random sampling was then used to identify three sample towns in each sample county, for a total of 36 towns. (3) A total of 106 villages were selected from these sample towns using random sampling. (4) A total of 1180 households were selected at random from the 106 villages. Participants provided verbal consent. Respondents were asked to complete the questionnaire. The village questionnaire collected information on the demographics of the village, infrastructure and characteristics of the village, and the household questionnaire collected information on the demographics of the respondent and their household, household income consumption, and energy consumption.

The target of the village questionnaire is the head or deputy head of the village council. The target of the household questionnaire is the head of the household, because the head of the household has more comprehensive information about the personal information of other household members, household production and consumption, and other related information, and the head of the household usually makes decisions for the household. If the head of the household was not available, the spouse of the head of the household or the son of the head of the household would be contacted to participate in the interview.

The investigators were senior undergraduate or graduate students from the School of Economics and Management of Shenyang Agricultural University. All the investigators attended a training held by Peking University and Shenyang Agricultural University, which focused on how to communicate with farmers, interpret the questions in the questionnaire, and use electronic devices to input data correctly. Before conducting formal research, we gave out a pre-survey with the purpose of familiarizing ourselves with the questionnaire and related equipment. Four rounds of review were conducted to ensure the quality of the survey data. After excluding incomplete questionnaires, a total of 106 village-level questionnaires and 1175 farmer questionnaires were obtained, of which the efficiency rate of the village-level questionnaires was 100% and the farmer questionnaires was 99.58%. Prior to empirical analysis, rigorous and serious data cleaning was carried out by Peking University.

2.2. Variable Selection

The explanatory variable of interest is the adoption of cleaner energy. This variable is measured through the question “What is the main energy source for cooking and boiling water?”. Referring to Ma et al. [31], Carter et al. [2], and Zhou et al. (2022) [36], this article divides energy into two categories: clean energy and non-clean energy. Clean energy is defined as 1 and includes canned liquified gas, pipeline natural gas, electricity, and solar energy. Non-clean energy, which includes straw, wood, and coal, is defined as 0. As shown in the survey results, 483 of the 1175 respondents used clean energy, accounting for 41.1% of the total, while the remaining 692 respondents used non-clean energy for cooking and boiling water. Figure 2. shows the non-clean energy use of the different households in the survey area.This article’s core explanatory variable is non-farm work. Referring to Ma et al. [37] and Zheng et al. [38], this variable is measured by the respondents’ non-agricultural employment status in 2018, which was assessed through the question “Did you have non-farm work in 2018?” From the survey results, 357 respondents had non-farm work in 2018, accounting for 30.4% of the total.

Household income and health knowledge are the two mediating variables in this article. The measurement of health knowledge is based on each respondent’s answers to 10 health knowledge questions. The score for each question is 0 or 1, and the total score ranges from 0 to 10.

Existing literature and available data are referenced to control for variables that may confound the statistical association between non-farm work and clean energy adoption. These variables include gender, age, marital status, and educational attainment. Variables related to household characteristics include household size (family size) and poverty. Some dwelling characteristics are sometimes seen as influencing fuel selection [39]. To account for agricultural production characteristics, existing research also includes farm sizes in fuel use regressions [40,41,42]. For this reason, we also controlled for variables such as housing type and cultivated farming area. As can be seen in

Table 1. T-test Results between Respondents Who Use Clean Energy vs. Non-clean Energy.

Variables	Non-Clean Energy Use Sample Size = 692		Clean Energy Use Sample Size = 483		T-Test (1–3)
	Mean (1)	SD (2)	Mean (3)	SD (4)
Non-farm Work	0.249	0.432	0.383	0.487	−0.134 ***
Household Income	10.173	1.212	10.756	1.179	−0.582 ***
Health Knowledge Score	6.467	1.360	6.938	1.559	−0.471 ***
Gender	0.806	0.395	0.783	0.413	0.024
Age	58.022	9.570	54.613	10.311	3.409 ***
Marital Status	0.923	0.266	0.936	0.245	−0.012
Highest Education Level	1.605	0.605	1.961	0.763	−0.355 ***
Household Size	3.127	1.336	3.176	1.264	−0.049
Housing Type	0.039	0.194	0.143	0.350	−0.104 ***
Poor Family	0.087	0.282	0.019	0.135	0.068 ***
Cultivated Farmland (area)	23.900	90.485	30.813	161.848	−6.914
Village Traffic Conditions	0.322	0.468	0.358	0.480	−0.036
Competitive Sanitation Activities	0.425	0.495	0.491	0.500	−0.066 **
Number of Private Enterprises in Village	2.013	3.751	2.861	5.160	−0.848 ***
Poor Village	0.227	0.419	0.155	0.363	0.072 ***
Straw-burning Prohibition	0.257	0.437	0.284	0.451	−0.026
Proportion of Hardened Roads	60.464	35.312	76.085	32.274	−15.621 ***

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

, respondents living in dwelling units were more likely to use clean energy. In addition, because traffic conditions, village regulations, and other factors could affect the availability and accessibility of fuels, we controlled for some variables at the village level as well. These included village traffic conditions, the number of private enterprises in the village, the number of hardened roads in the village, whether it was forbidden to burn stalks at home (yes or no), whether a family health evaluation program was conducted in the village (yes or no), and whether the village was poor. In poor villages, respondents were more likely to use non-clean energy, as shown in Table 1.

The definitions and descriptive statistics of the above variables are shown in

Table A1. Summary Statistics.

Variables	Description	Mean	SD	Min	Max
Cleaner Energy Adoption	1 = yes, 0 = no	0.411	0.492	0	1
Non-Farm Work	Non-Farm Work in 2018? 1 = yes, 0 = no	0.304	0.460	0	1
Household Income	continuous variable (taking the logarithm)	10.413	1.232	8.458	12.055
Health Knowledge Score	Scores 0–10	6.660	1.463	2	10
Gender	1 = male, 0 = female	0.797	0.403	0	1
Age	Continuous variable (year)	56.620	10.019	27	84
Marital Status	1 = married, 0 = unmarried	0.929	0.258	0	1
Highest Education Level	1 = primary school and below, 2 = junior high school, 3 = senior high school/vocational high school/technical secondary school, 4 = junior college and above	1.751	0.696	1	4
Household Size	discrete variable (people)	3.147	1.306	1	9
Housing Type	Live in a dwelling unit? 1 = yes, 0 = no	0.082	0.274	0	1
Poor Family	1 = yes, 0 = no	0.059	0.235	0	1
Cultivated Farmland (area)	Households’ actual cultivated farm area in 2018 (mu)	26.742	124.841	0	3.200
Village Traffic Conditions	Is the village accessible by bus? 1 = yes, 0 = no	0.337	0.473	0	1
Competitive Sanitation Activities	Has the village conducted a family sanitation competition before? 1 = yes, 0 = no	0.452	0.498	0	1
Number of Private Enterprises in Village	discrete variable (individual)	2.362	4.403	0	32
Poor Village	1 = yes, 0 = no	0.197	0.398	0	1
Straw-burning Prohibition	It is forbidden to burn stalks at home. 1 = yes, 0 = no	0.268	0.443	0	1
Proportion of Hardened Roads	Proportion of hardened roads in the village (%)	66.885	34.938	0	100

Note: A dwelling unit generally refers to a self-contained independent house composed of a kitchen, bathroom, and rooms with relatively complete facilities, equivalent to a western apartment.

(in the Appendix A).

2.3. Empirical Strategy

2.3.1. Probit Models

A Probit Regression was conducted to analyze the impact of non-farm work on respondents’ clean energy adoption. The specific model is:

$$Y_i={\beta }_0+{\beta }_{1}nonfar{m}_i+{\beta }_2{X}_i+u_i$$

(1)

In Formula (1), the dependent variable $Y_i$ is the dependent variable indicates whether the respondents use clean energy; the explanatory variable $nonfar{m}_i$ takes a value of 1 if the respondent had non-farm work in 2018, and 0 otherwise; ${\beta }_1$ is the coefficient to be evaluated; $X_i$ represents the control variables, including individual, family and village characteristics; $u_i$ is the random disturbance term.

2.3.2. Extended Regression Models (ERMs)

Although we have controlled for the related determinants of clean energy use to the extent that data are available, there may still exist potential endogeneity problems caused by unobservable factors. Failing to address this problem can lead to biased estimates of the effects of non-farm work on the adoption of clean energy.

To overcome any potential endogeneity, we use the IV and employ Extended Regression Models (ERMs) as suggested by Stata 15 [43]. From the variables included in the survey data, a relatively appropriate IV can be constructed using the normalization method. Referring to Ma et al. [37], and Zhou et al. [36], the proportion of non-farm work in the village is used as an instrumental variable for the non-farm work-in dummy in Model (1). Theoretically, the proportion of residents in the village who perform non-farm work will affect fellow residents’ choice of farm or non-farm work; in other words, the more migrant workers there are in the village, the more likely a given resident is to choose non-farm work. This represents the presence of peer effects. For the exogeneity and exclusion restriction of the IV, the proportion of non-farm work in the village is a variable on the village level and the adoption of clean energy is a variable on the individual level. Therefore, the dependent variable cannot affect the IV in reverse and the IV is considerably exogenous to the dependent variable. Furthermore, the IVs are unlikely to directly affect rural residents’ clean energy adoption by influencing other explanatory variables.

Because the endogenous variable in Model (1) is a binary indicator, Stata’s traditional Probit Regression commands that have endogeneity concerns (such as “ivprobit”) cannot estimate the model parameters. For this reason, ERMs are used instead. ERMs are regression models that account for continuous, binary, or ordinal endogenous covariates in the models through continuous, interval-measured and censored, binary and ordinal outcomes by employing maximum likelihood estimation [43]. Since the dependent variable in this paper is also a binary variable, we use the command “eprobit” to create Probit regression models with instrumental variables.

2.3.3. Propensity Score Matching (PSM)

Considering that it is difficult to find an absolute exogenous IV for non-farm work within the scope of the survey data, the PSM method was employed to further test the robustness of the results. PSM can effectively reduce selection bias through mediating differences between treatment and control groups [44]. Additionally, PSM uses nonparametric estimates without making assumptions about the relationship between the explained and explanatory variables, so it is widely used in economics [45,46,47,48,49].

3. Results and Discussion

3.1. Baseline Results

Before conducting the empirical test, firstly, all variables were tested for multicollinearity using the Stata16 software, and the results showed that MaxVIF = 1.29, MinVIF = 1.01, and MeanVIF = 1.12, and the maximum VIF was significantly less than 10, which shows that there is no serious multicollinearity problem among the variables. Second, the correlation coefficients between the explanatory variables were tested and the results showed that the maximum value of the absolute value of the correlation coefficient between the explanatory variables was 0.295 and the minimum value was 0.001, which indicates that the explanatory variables are all independent of each other.

Table 2. Baseline Regression Results.

Variables	Clean Energy Adoption
	(1)	(2)	(3)
Non-farm Work	0.150 ***	0.041 ***	0.041 ***
	(0.028)	(0.008)	(0.006)
Gender		0.016	0.027
		(0.073)	(0.073)
Age		−0.005 ***	−0.005 ***
		(0.002)	(0.002)
Age2		0.005	0.002
		(0.003)	(0.003)
Marital Status		0.014	0.021
		(0.031)	(0.031)
Highest Education Level		0.121 ***	0.116 ***
		(0.009)	(0.010)
Household Size		−0.012 ***	−0.008 *
		(0.004)	(0.004)
Housing Type		0.226 ***	0.217 ***
		(0.072)	(0.075)
Poor Family		−0.211 ***	−0.199 ***
		(0.065)	(0.042)
Cultivated Farmland (area)		0.000	0.000
		(0.000)	(0.000)
Village Traffic Conditions		0.021 ***	0.062 ***
		(0.005)	(0.010)
Competitive Sanitation Activities		0.031 ***	0.031 ***
		(0.005)	(0.000)
Number of Private Enterprises in Village		0.011	0.006
		(0.008)	(0.007)
Poor Village		−0.089	−0.122 **
		(0.057)	(0.055)
Straw-burning Prohibition		0.041 **	0.006
		(0.020)	(0.029)
Proportion of Hardened Roads		0.003 ***	0.003 ***
		(0.000)	(0.000)
Area Fixed Effect	No	No	Yes
Observation	1175	1175	1175

Note: Robust standard errors are reported in parentheses. Significance: * p < 0.10, ** p < 0.05, *** p < 0.01.

reports the marginal effects of the explanatory variables based on the estimated parameters of the Probit model. As shown in Table 2, Column (1), the coefficient for the non-farm work is positive and statistically significant. After controlling for individual, household, and village characteristics, the effect is found to be consistently positive and statistically significant at the 1 percent level, as shown in Column (2). After adding the area fixed effect, the empirical result in Column (3) is in line with expectations–that is, the non-farm work has significant marginal effects on the rural residents’ adoption of clean energy.

Certain control variables also provide valuable information. Existing research indicates that gender is a debated factor as to its influence on farmers’ energy use [20]. In our research (see Table 2), we found no significant correlation between gender and clean energy use. This finding is in line with several studies [50,51]. When controlling for area fixed effects, age was negatively, but not significantly, associated with the adoption of clean energy. The squared term for age was also not significant. There is a U-shaped relationship between age and clean energy adoption. This result may be affected by the relatively high average age of the sample data, and the possibility that rural residents will adopt less clean energy as they age. Existing literature has confirmed that education also plays an important role in energy use decision-making [52,53]. Education level is found to be positively correlated with clean energy use in our sample, a finding that is consistent with Heltberg [54], who conducted a survey in eight developing countries and found that there was a significant positive relationship between the use of modern fuels and education level. Household size also plays an important role in energy choices. However, the effect of household size on energy use is still ambiguous [20]. This study finds that there is a significant negative correlation between household size and clean energy use; that is, households are more likely to use solid fuels when household size is large. This finding is consistent with those of Rao and Reddy [55] and Ozcan et al. [56].

The probability that farmers will use clean energy is higher in villages with better traffic conditions and higher proportions of hardened roads. Farmers living in dwelling units are more likely to use clean energy. In addition, villages that have competitive sanitation activities increase the probability that rural residents adopt clean energy. Therefore, the use of clean energy depends on both personal insight and institutional incentives and is limited by wealth and infrastructure.

3.2. Robustness Check

3.2.1. Results Using Instrumental Variables

Table 3. Estimation Results Based on the ERM Regression.

Variables	Cleaner Energy Adoption	Non-Farm Work
	(1)	(2)
Non-farm Work	2.193 ***
	(0.096)
Instrumental Variable		0.242 ***
		(0.081)
Controls	Yes	Yes
Area Fixed Effect	Yes	Yes
Correlation between Non-farm Work and Clean Energy Adoption	−0.935 ***
	(0.044)
Observation	1175	1175

Note: Control variables include gender, age, age², marital status, highest education level, household size, housing type, family poverty, cultivated farmland (area), village traffic conditions, competitive sanitation activities, number of private enterprises in village, village poverty, straw-burning prohibition, and proportion of hardened roads. Robust standard errors are reported in parentheses. Significance: *** p < 0.01.

, Column (1) reports the parameter estimates using the ERM. Table 3, Column (2) presents the first-stage regression results. The correlations between the error terms in the equations for non-farm work and clean energy adoption were analyzed to determine whether non-farm work and adoption of clean energy are endogenous. The results show that the correlation between error terms is significantly different from 0 (Column (1), row 5). This indicates that unobserved factors affect non-farm work and adoption of clean energy simultaneously, and that non-farm work is endogenous. Additionally, this indicates that the instrumental variable works well. In villages where the proportion of migrant workers is higher, rural residents tend to have non-farm work. The coefficient of the instrumental variable is also significant at the 0.01 level. The value of the LR chi-squared test for the instrumental variable is 167.39. In addition, this paper refers to the method proposed by Andrews [57] of observing the minimum characteristic root to make a judgment; the minimum characteristic root is equivalent to the F statistic, and the larger the value, the less likely it is that a weak instrumental variable appears. An F value of 10.94 exceeds the empirical value of 10, so the original hypothesis that the proportion of non-farm work in the village is a weak instrumental variable can be empirically rejected.

Compared with the Probit estimation results in Table 2, Column (3), the ERM estimation coefficient is increased, indicating that possible endogeneity does make the Probit results underestimated to a certain extent. The results of the ERM estimation confirm the robustness of the benchmark regression conclusions.

3.2.2. Results Using the PSM Method

Table 4. Estimation Results Based on the PSM Method.

Matching Method	K-Nearest Neighbor Matching (k = 1)	K-Nearest Neighbor Matching (k = 4)	Radius Matching (k = 3 and r = 0.09)
ATT	0.070 **	0.068 **	0.062 *
S.E.	0.040	0.032	0.032
Z-test	1.74	2.12	1.91

Note: k-nearest neighbor matching uses the “teffects nnmatch” command and the Z-test, while radius matching uses the “psmatch2” command and the Z-test. Significance: * p < 0.10, ** p < 0.05.

shows the estimation results using the PSM method. In particular, “psmatch2” provides many matching methods, but has the disadvantage of obtaining incorrect standard errors. That is, the standard error provided by “psmatch2” does not take into account that the propensity score is estimated, which may lead to biased estimates [58]. Stata 13 introduced a new “teffects” command for estimating treatment effects, which gives the correct standard errors as proposed by Abadie and Imbens (called “AI Robust Standard Errors”). Therefore, the “teffects nnmatch” command was used to perform k-nearest neighbor matching, allowing the correct standard error to be obtained. Additionally, the “psmatch2” command was used to perform radius matching to check the robustness of the results. In addition, the results of the balance test show no significant difference in mean values among of all covariables after matches are made between the control group and the treatment group. The results of the test are shown in

Table A2. Balance Test.

Variables	Mean		% Bias	t-Test
	Treated	Control		t	p > |t|
Gender	0.813	0.822	−2.2	−0.30	0.766
Age	53.428	53.65	−2.3	−0.32	0.747
Marital Status	2935.2	2961.9	−2.5	−0.36	0.715
Highest Education Level	0.920	0.929	−3.6	−0.47	0.640
Household Size	1.960	1.910	7.0	0.93	0.351
Housing type	3.259	3.256	0.2	0.02	0.981
Poor Family	0.106	0.101	1.7	0.21	0.832
Cultivated Farmland (area)	0.029	0.031	−0.8	−0.14	0.888
Village Traffic Conditions	27.228	30.618	−2.4	−0.32	0.752
Competitive Sanitation Activities	0.330	0.342	−2.5	−0.33	0.740
Number of Private Enterprises in Village	0.503	0.503	0.1	0.01	0.994
Poor Village	2.603	2.671	−1.4	−0.19	0.848
Straw-burning Prohibition	0.178	0.180	−0.4	−0.05	0.962
Proportion of Hardened Roads	0.241	0.241	0.0	0.00	0.996

.

As shown in Table 4, after k-nearest neighbor matching, having non-farm work significantly contributes to the probability that rural residents will adopt clean energy. The result for radius matching was similar, indicating that the results of the benchmark regression are robust.

3.2.3. Changing the Measure of Variables

A common practice for robustness testing is to replace the core variables, so the core explanatory variable in this article, non-farm work, was replaced with the proportion of migrant workers across the total number of families, and their non-farm income [31]. The Equation (1) was then re-estimated. These results still confirm that non-farm work has significant marginal effects on rural residents’ adoption of clean energy. The coefficients of the other control variables are consistent with Table 2. See

Table A3. Robustness Checks: Proportion of Migrant Workers and Non-Farm Income.

Variables	Clean Energy Adoption
	(1)	(2)
Proportion of Migrant Workers in the Total Number of Families	0.095 ***
	(0.006)
Non-Farm Income		0.012 ***
		(0.002)
Gender	0.037	0.040
	(0.074)	(0.073)
Age	−0.011 **	−0.011 *
	(0.005)	(0.006)
Age2	0.005 *	0.005
	(0.003)	(0.004)
Marital Status	0.032	0.027
	(0.037)	(0.035)
Highest Education Level	0.114 ***	0.114 ***
	(0.009)	(0.007)
Household Size	−0.011 ***	−0.021 ***
	(0.002)	(0.003)
Housing Type	0.214 ***	0.214 ***
	(0.076)	(0.076)
Poor Family	−0.195 ***	−0.193 ***
	(0.043)	(0.043)
Cultivated Farmland (area)	0.000	0.000
	(0.000)	(0.000)
Village Traffic Conditions	0.062 ***	0.066 ***
	(0.012)	(0.012)
Competitive Sanitation Activities	0.030 ***	0.031 ***
	(0.002)	(0.001)
Number of Private Enterprises in Village	0.006	0.006
	(0.007)	(0.007)
Poor Village	−0.120 **	−0.121 **
	(0.051)	(0.055)
Straw-burning Prohibition	0.005	0.003
	(0.029)	(0.023)
Proportion of Hardened Roads	0.003 ***	0.003 ***
	(0.000)	(0.000)
Area Fixed Effect	Yes	Yes
Observation	1175	1175

Note: Robust standard errors are reported in parentheses. Significance: * p < 0.10, ** p < 0.05, *** p < 0.01.

in the Appendix A for details.

Additionally, this study estimates the effect non-farm work had on the residents’ use of straw, wood, coal, liquefied gas, pipeline natural gas, and electricity individually. The results, which are presented in

Table A4. Robustness Checks: Different Types of Energy.

Variables	Straw	Wood	Coal	Liquefied Gas	Natural Gas	Electricity
	(1)	(2)	(3)	(4)	(5)	(6)
Non-Farm Work	−0.010	−0.022 *	−0.020	0.002 **	0.037	−0.000
	(0.006)	(0.013)	(0.000)	(0.001)	(0.000)	(0.007)
Control Variables	Yes	Yes	Yes	Yes	Yes	Yes
Area Fixed Effect	Yes	Yes	Yes	Yes	Yes	Yes
Observation	1175	1175	1175	1175	1175	1175

Note: Control variables include gender, age, age², marital status, highest education level, household size, housing type, family poverty, cultivated farmland (area), village traffic conditions, competitive sanitation activities, number of private enterprises in village, village poverty, straw-burning prohibition, and proportion of hardened roads. Robust standard errors are reported in parentheses. Significance: * p < 0.10, ** p < 0.05.

in the Appendix A, demonstrate that non-farm work has a positive impact on the use of liquified gas and pipeline natural gas, while it affects straw, wood, and coal use negatively. These findings further confirm the benchmark regression conclusions.

4. Mechanism Analysis

In economic terms, non-farm work is a very important production activity for rural residents, as it profoundly affects their production and consumption decisions [38]. In the above analysis, significant positive impact was found on rural residents’ adoption of clean energy from non-farm work. This section explores potential mechanisms, particularly focusing on the residents’ household income (income effect) and health knowledge (human capital accumulation effect). It is difficult to rule out the possibility that there are other mechanisms at play. Accordingly, a decomposition analysis is employed to show that these two mechanisms explain a great deal of the effect of non-farm work on rural residents’ clean energy adoption.

Firstly, non-farm work affects rural residents’ household incomes, which influences their adoption of clean energy. Studies have addressed the effect of non-farm work on household incomes and shown that employment in the non-farm sector can exert positive and significant impacts on household income [28,59,60]. Furthermore, numerous studies posit income as the major driver of increased use of modern fuels and indicate that transitioning from fuelwood to kerosene, natural gas, and electricity occurs alongside increases in income. Household wealth is a significant negative determinant of level of fuelwood consumption; moreover, reduction of fuelwood consumption may halt at a certain income level [39,42,51]. As shown in

Table 5. Mechanism: Household Income and Health Knowledge Score.

Variables	Household Income	Health Knowledge Score
	(1)	(2)
Non-Farm Work	0.382 *	0.141 *
	(0.040)	(0.013)
Control Variables	Yes	Yes
Area Fixed Effect	Yes	Yes
Observations	1175	1175
R-squared	0.391	0.069

Note: Control variables include gender, age, age², marital status, highest education level, household size, housing type, family poverty, cultivated farmland (area), village traffic conditions, competitive sanitation activities, number of private enterprises in village, village poverty, straw-burning prohibition, and proportion of hardened roads. Robust standard errors are reported in parentheses. Significance: * p < 0.10.

, Column (1), non-farm work has a positive association with rural residents’ household income. Our findings echo prior studies.

A second possible mechanism is that non-farm work affects farmer’s health knowledge, which influences their adoption of clean energy. The theory of human capital accumulation states that different personal experiences—such as education/training experience, work experience, and life experience—may affect a person’s accumulation of knowledge and skills. These personal experiences may also change a person’s thoughts and behaviors [61,62]. Out-migrating for work is an important way to accumulate human capital. Work experience in a more developed county or city can expand a rural resident’s horizons. Furthermore, the awareness of environmental protection and healthy behaviors brought about by a more modern life will subtly impact rural residents’ traditional living habits, in turn promoting a higher level of health and hygiene knowledge in this population [63,64]. This improvement in health knowledge will affect rural residents’ decisions on the use of clean energy. Migrant workers who leave rural areas for non-farm work often move to more developed cities and gain experience with the use of clean energy for cooking and bathing, making them more willing to use clean energy after returning to their homes in rural areas. Column (2) in Table 5 shows the results of the estimation, which indicate that non-farm work is significantly positively correlated with rural residents’ health knowledge.

Our findings show that non-farm work may influence rural residents’ household incomes and health knowledge, in turn influencing their choice to adopt clean energy. To further understand the extent to which each channel explains the impact of non-farm work and the mechanisms’ total explanatory power, we draw on Gelbach [34] and Gong et al. [35] in employing a decomposition method. Specifically, $M_i^j$ is denoted as the mechanism variable j and the following estimation equation is considered:

$$M_i^j={\alpha }_1^{j}nonfar{m}_i+{\beta }_2{X}_i+u_i$$

(2)

The following specification is considered including all relevant mechanism variables in Equation (1):

$$Y_i={\theta }_{1}nonfar{m}_i+{\beta }_2{X}_i+{{\displaystyle \sum }}_j{\gamma }^j{M}_i^j+{ϵ}_i$$

(3)

Gelbach (2016) shows that:

$$\widehat{{\beta }_1}=\widehat{{\theta }_1}+{{\displaystyle \sum }}_j{\widehat{\gamma }}^j{\widehat{{\alpha }_1}}^j$$

(4)

This equation suggests that mechanism j’s component is ${\widehat{\gamma }}^j{\widehat{{\alpha }_1}}^j$, and the remaining unexplained part is $\widehat{{\theta }_1}$. For each mechanism, explanatory power is computed for impact of non-farm work by ${\widehat{\gamma }}^j{\widehat{{\alpha }_1}}^j/\widehat{{\beta }_1}$ (Note that if unmeasured mechanisms are associated with observed mechanisms and/or if the observed mechanisms have measurement errors, ${\gamma }^j$ may be biased. Therefore, caution must be used when interpreting the decomposition results).

Figure 3 plots the estimated decomposition of the impact of non-farm work on clean energy adoption into household income, health knowledge, and other factors. In regard to the effect on the adoption of clean energy, household income is found to explain approximately 62.0% of this effect, and health knowledge explains around 9.9%. Taken together, they explain 71.9% of the impact of non-farm work on clean energy adoption amongst rural residents. The remainder is unexplained by these mechanisms.

5. Heterogeneity Analysis

Finally, this study explores variations in the effect of non-farm work based on individual characteristics—i.e., based on gender, educational attainment, involvement in village cadres, and age as shown in Figure 4 and

Table 6. Heterogeneity analysis: Age.

Variables	Clean Energy Adoption
	38 Years Old and Younger	39–60 Years Old	61 Years and Older
Non-farm Work	0.173 ***	0.053 *	0.039
	(0.029)	(0.029)	(0.075)
Controls	Yes	Yes	Yes
Area Fixed Effect	Yes	Yes	Yes
Observation	51	647	476

Note: Control variables include gender, marital status, highest education level, household size, housing type, family poverty, cultivated farmland (area), village traffic conditions, competitive sanitation activities, number of private enterprises in village, village poverty, straw-burning prohibition, and proportion of hardened roads. Robust standard errors are reported in parentheses. Significance: * p < 0.10, *** p < 0.01.

.

We find that non-farm work contributes significantly to clean energy use in men, but it does not have a significant effect on clean energy use in women. This is perhaps because men generally earn more than women from non-farm employment, which would magnify the income effect. Furthermore, in rural areas men and women have unequal family statuses, and economic decisions (such as purchasing clean energy) are mostly made by men, which weakens the impact of women’s non-farm work. This is consistent with McLean [65], who found no significant association between female employment and solid fuel use in Peru. However, in the developing world, women are traditionally responsible for cooking, and therefore have a strong interest in cleaner and more convenient energy sources [30]. As a result, when women’s income from non-farm work increases, or when they head their households, their preferences are more likely to be realized. This is verified in our study. As shown in

Table 7. Marginal Effects of Women’s Non-Farm Income.

Variables	Low Non-Farm Income	High Non-Farm Income
	(1)	(2)
Non-Farm Work	0.078	0.095 ***
	(0.084)	(0.023)
Control Variables	Yes	Yes
Area Fixed Effect	Yes	Yes
Observations	121	118

Note: Control variables include age, age², marital status, highest education level, household size, housing type, family poverty, cultivated farmland (area), village traffic conditions, competitive sanitation activities, number of private enterprises in village, village poverty, straw-burning prohibition, and proportion of hardened roads. Robust standard errors are reported in parentheses. Significance: *** p < 0.01.

, when women’s income from non-farm work increases, their non-farm work has a significantly positive impact on their adoption of clean energy (=0.095, p < 0.01). This impact is greater than that seen with increases in men’s non-farm income (=0.078, p > 0.1).

Education is a strong determinant of switching to cleaner fuel sources [66,67]. This paper finds that non-farm work positively impacts clean energy adoption in those with a junior high school education or above. However, non-farm work is found to have a negative impact on clean energy adoption in participants with a primary school education or below. In addition, there is evidence that family member village cadre-involvement significantly increases the probability of household clean energy adoption. As representatives of collective organizations in the village, members of village cadres are obligated to set an example in environmental protection and are therefore more receptive to suggestions from family members about clean energy adoption. This is a possible explanation for these findings. Age is also a significant factor in household clean energy adoption [2,68]. Non-farm work has a significant positive effect on household clean energy adoption for those 38 years old and younger, while non-farm work has a non-significant effect on household clean energy adoption for those age 61 and older. Possible explanations are that older people may be less likely to use clean energy because they want to save money, stick to old habits, avoid using clean stoves, or because they have a harder time accepting new things.

6. Conclusions and Policy Recommendations

Promoting the adoption of clean energy in rural households is not only beneficial to the health of rural residents, but also advantageous for the protection of rural environments. In this study, a representative survey of rural residents was used to investigate the effect of non-farm work on their clean energy adoption. Extended Regression Models were combined with the instrumental variables method to address endogeneity issues.

Our study shows that non-farm work significantly raises the probability of clean energy adoption among rural households. By exploring potential mechanisms, evidence was found that non-farm work significantly increases total household income and health knowledge, which in turn influences their clean energy adoption. These mechanisms explain many of the identified effects. Further heterogeneity analysis showed that non-farm work has a more positive impact on household clean energy adoption for male and young farmers, those with a junior high school education or above, and those with a village head in the family.

Rural residents’ non-farm work contributes to the well-being of their family, especially in the context of the large economic and infrastructure gaps, between urban and rural China. These results have strong policy implications. Firstly, the government should recognize the income effect brought about by non-farm work. Moreover, the government should provide further opportunity for non-farm work for these communities and promote the growth of their non-farm income, so as to reduce the economic burden they face in the process of clean energy adoption. Specifically, the government should pay attention to changes in the demand for non-farm work from rural laborers and improve the service level and market environment for the non-farm work market. This study found that non-farm work affects clean energy adoption for rural residents with an education level above junior high school, while it had a negative effect on those with an education level below primary school. Therefore, attention should be paid to improving the human capital and vocational skills of the non-farm labor force, providing more development opportunities for rural residents, promoting opportunities for them to become better qualified for higher-level jobs, and further increasing the income effect of non-farm work on the clean energy adoption.

Secondly, the government should increase its efforts to promote health literacy in rural areas and raise awareness of clean energy adoption. The majority of China’s rural population lacks general health knowledge. Policymakers should use news media such as radio, television, and the Internet to enhance health education for rural residents to understand that clean energy adoption is beneficial to environmental and public health. For example, health education events could be held regularly in villages to show educational videos related to health and environmental protection.

Finally, the findings of this study also suggest that rural infrastructure (such as the proportion of hardened roads) and positive institutional arrangements (such as competitive sanitation activities in the village) also promote clean energy adoption. Therefore, the government should strengthen the construction of rural infrastructure, vigorously develop solar and wind power generation, improve the environment and conditions for clean energy, and design a reasonable system to promote the transformation and upgrading of energy for rural life.

There are shortcomings in this paper: first, women are becoming more important in household decision making, and the influence of women on household domestic energy use needs to be further explored. Second, the data used in this paper are cross-sectional data, which cannot fully control for unobservable factors, and further use of panel data is needed to explore the influence mechanism between non-farm work and household clean energy adoption. Due to the design of the questionnaire, the surveyed data only have the main energy sources used for cooking and boiling water and lack a detailed description of the total energy consumption and energy consumption structure of rural households. The fuel accumulation behavior of Chinese rural households needs to be explored in depth in the future. Finally, further research using more representative data from different countries and regions is necessary to test whether the reported relationships hold in other settings and to better understand the determinants of clean energy adoption in developing countries.