The Effect of Non-agriculture Employment on The Rural Household Poverty Alleviation: Evidence from a Deeply Im-poverished County in Southwest China

. Encouraging rural households from deeply impoverished areas to participate in non-agricultural employment has been regarded as an e ﬀ ective way to alleviate rural poverty. China’s targeted poverty alleviation (TPA) project has made signiﬁcant achievements, with its policy to encourage rural households to participate in non-agricultural employments playing an important TPA role. Taking a deeply impoverished county in Southwest China as an example, this paper used an endogenous switching regression (ESR) model under a counterfactual inference framework to evaluate the e ﬀ ects of non-agricultural employment on alleviating household poverty, with the simpliﬁed “Organization for Economic Co-operation and Development (OECD) equivalent scale” formula used to adjust the income to measure household welfare. It was found that non-agricultural employment had reduced participant poverty and greatly improved the welfare of the participating households. However, for the non-participants, the non-agricultural employment income would be lower than the agricultural income, and the transfer of the household labor force to non-agricultural employment would deepen household poverty. This paper concluded with a discussion of the policy options to consolidate the achievements of poverty alleviation in deeply impoverished areas.


Introduction
Poverty alleviation is a global challenge [1], with the international community continuing to find ways to narrow the rural-urban gap and eliminate poverty [2]. Since the implementation of the TPA project in 2013, China's poverty alleviation has achieved remarkable achievements. Even in deeply impoverished areas with slow social development, many farmers have been lifted out of absolute poverty a few years after the implementation of the TPA project. During the implementation of TPA, encouraging rural households to participate in the nonagricultural employment was considered to be an effective poverty alleviation policy and has been widely implemented.
However, in deeply impoverished areas, the impact of participating in non-agricultural employment on poverty reduction for the household was still undetermined, and some rural households lifted out of poverty never participated in non-agricultural employment. Some studies suggested that not all farmers benefit from off-farm employment, such as those with low levels of education who end up losing their jobs or entering the informal economy [3]. The positive effect of participating in non-agricultural employment policy was influenced by household characteristics.
The choice of whether to participate in non-agricultural employment showed the heterogeneity among households. For non-participants, non-agricultural employment may not be an effective means of poverty alleviation. The counterfactual inference of the household's participation in non-agricultural employment will provide a solid theoretical basis for studying the poverty alleviation effect of non-agricultural employment in deeply impoverished areas. This paper used an ESR model to analyze the impacts of non-agricultural employment participation on rural household income and welfare based on a survey conducted in Southwest China. Household welfare was determined by adjusting the household income using a simplified "OECD equivalent scale" formula, which more accurately measured the poverty alleviation effects of the non-agricultural employment.
The remainder of this paper is organized as follows. Section 2 gives an overview of previous literature, Section 3 describes the sampling method and research strategy, Section 4 introduces the empirical model and results, Section 5 assesses the robustness of empirical results, and Section 6 discusses the results and gives the conclusion.

Literature Review
Moving to urban areas to participate in non-agricultural employment has become one of the most important ways for rural households to increase their incomes [4]. Early economists noticed that non-agricultural employment increased income; for example, [5] concluded that a labor force would continue to move from an agricultural sector with low marginal productivity to urban industrial sectors with high marginal productivity to improve rural household income. However, [6] believed that even if traditional agriculture was poor, efficient farmers could optimally allocate existing production factors, and that the transfer of rural labor to non-agricultural sectors could lead to a reduction in agricultural income, that is, improving human capital and introducing new production factors could more effectively improve agricultural output and household income.
Many studies have analyzed the poverty alleviation effects of non-agricultural employment. For example, [4] studied the data of three waves (2000, 2007 and 2014) of an Indonesian family life survey using Difference in Difference (DID) regression and an ordered logit model, finding that moving out of agriculture significantly increased the welfare of poor rural households, especially between 2000-2007; however, this was not found from 2007-2014. [7] used a multiple regression model and a boosted regression trees method to analyze 338 villages in a traditional agricultural area in northern Jiangsu, China, and found that transportation infrastructure, labor outflow, land transfers, non-agricultural employment, entrepreneurship, and rural labor transfers were the determinants for rural economic development. [8] used quantile regression to analyze data from Jiangxi province, China, and found that labor force transfers had a significant positive impact on the income of poor families. [9] analyzed Ghanaian data using a propensity score matching (PSM) model and inverse probability weighted regression adjustment and found that a diversification in non-agricultural income increased participation in agricultural technology, which significantly contributed to household welfare improvements. [10] used a PSM found that non-agricultural employment played a vital role in improving household's economic farewell. [11] also used PSM to study the poverty problem of Mauritania, and believed that the extra income from non-agricultural employment eased the mobility restrictions, enabling household to obtain better agricultural production inputs. Because the problem of choice bias must be solved when measuring the poverty reduction effect of non-agricultural employment of households in many cases, the PSM method has been used in quite many studies. Still, PSM has defects in the heterogeneity of households. The use of the ESR model can overcome the problem of bias and heterogeneity simultaneously but few studies used the ESR model to study the poverty reduction effect of non-agricultural employment.
Non-agricultural employment has also been found to impact agricultural production because families that have limited land tenure and livestock choose to leave agriculture to earn higher incomes elsewhere, with this being especially true for young people who were often underemployed on their own family farms [12]. The relationship between non-agricultural employment and food production has been found to be an inverted U-shaped; when the nonagricultural labor supply is relatively low, non-agricultural employment increases grain production and when the supply of non-agricultural labor force is relatively high, there is a decrease in grain production. It has also been surmised that agricultural machinery and land scale input factors could also affect the relationship between non-agricultural employment and food production [13].
Studies on the human capital factors that affect farmer participation in non-agricultural employment have been identified as a lack of necessary skills and poor health. Because of the health status of family members, farmers often ignore employment opportunities [12]. The characteristics of farmers who participate in non-agricultural employment have also been found to affect their income. For example, [12] found that when farmland was converted to industrial land in Vietnam, the farmers had difficulty finding employment because of their low education levels, and [14] concluded that the employment substitution effect was only evident for low skilled local workers.
There have also been many studies on the impact of material capital on non-agricultural employment participation, with infrastructure and transportation also having been found to affect rural worker choices. [15] concluded that poverty was directly related to isolation, poor transportation, and poor infrastructure development, that is geographic isolation and a lack of transportation limited the participation of poor farmers in labor and agricultural markets, which hindered their income opportunities. Therefore, it has been suggested that better rural road infrastructure could increase the opportunities for the rural poor by creating jobs, providing basic services, and increasing household mobility [16]. Using state level data from 1970 to 1993, [17] developed a simultaneous equation model to estimate the direct and indirect impacts of different types of infrastructure on agricultural productivity and rural poverty in India, and found that additional investment in rural roads and agricultural research could increase agricultural productivity, create jobs, and increase per capita income. Reputation, ethnic networks, and education levels were also found to be positively significant, and had an impact, with all factors being found to contribute to improving employment opportunities for immigrants [18].

The Survey Sampling
The sample data examined in this paper were taken from a large-scale survey that had been conducted in a deeply impoverished county in Southwest China. The sample area is 2150.79 square kilometers and has a population of 235643 in 227 villages (including 208 villages assessed as impoverished in 2014). Using a stratified random sampling design and structured questionnaires, 822 households from 28 villages were selected. First, based on suggestions from the local authority staff, the townships were divided into rich, ordinary, and poor categories, the ratio for which was around 2:3:5. Then, based on the economic conditions, 10 townships comprising 28 villages were selected using stratified random sampling to ensure the sampled villages represented the overall economic situation in the sample area. Finally, all households in the selected villages were surveyed.
An 80 questions survey was used in the interview, which covered information about the household structures and economic statuses. The survey was carried out with the assistance of local authority staff and the data collected by the research team through face-to-face interviews over one week at the respondents' homes with the household head or household members over 18 years old. Some data were also collected from the local authority database.

Adjustment to Income
The average income was considered the important index for measuring the poverty level of the households, which comprised income from different sources; transfer income, agricultural income, wage income, and property income. However, using average household income to measure the household living conditions ignores the influence of household size and composition, that is, considering the scale economy effect of the household, traditional household per capita income measurements would underestimate the welfare of its members and overestimate the poverty level of the households [19]. For example, the average family income for a three-person household would only be two-thirds of the average household income for a household of two; however, the gap in living standards would not be as large as the income gap indicated. Previous household welfare studies have used equivalence scales to correct for these deviations [20]. Therefore, based on the simplified "OECD equivalent scale" formula given in [21] to analyze the impact of tax and welfare on household income, this paper adjusted the income of the poor households as follows: AverInc adj = Inc/(a 1 + 0.5a 2 + 0.3a 3 where Inc was the income of the household, a 1 = 1 when there were adults in the household, and 0 otherwise, a 2 equaled the number of extra adults in the household, and a 3 respectively equaled the number of children (0-14 years old). The adjustment formula in the literature was developed based on the European Socioeconomic statement. The elderly in European generally have pensions and rarely live with their children. The "OECD equivalent scale" algorithm did not distinguish the elderly from other adults. However, most of the elderly in deeply impoverished areas have no pension and live with their children. In the sample selected in this paper, the elderly lived alone in only 13 households. According to the actual situation of the household survey, the sharing proportion of the elderly which belonged to extra adults in the house should be closer to that of children. Therefore, based on the existing simplified "OECD equivalent scale" formula, this paper adjusted the weight of the elderly belonging to extra adults in the households to 0.3 according to the actual situation, so as to obtain an equivalent formula suitable for the situation of deeply impoverished areas.

Summary Statistics
Based on the literature review and field research, 9 covariates were then selected (table 1). Household, Student, Old and Young were used to represent the household structure. The variable means classified by participation status (1 = participants, and 0 = non-participants) showed that there were significant differences between the two rural household groups (table 2). The total participant income was only 7% higher than for the non-participants; however, the agricultural income was less than 40% of the non-participant agricultural income, which If the household labor force had completed 9 years of compulsory education, education = 1; otherwise, education = 0 indicated that participation in non-agricultural labor was more likely to motivate some households to give up agricultural production altogether. The adjusted average income showed that the participants' welfare was much higher. There were obvious differences in the household structures between the participant and non-participant groups. The elderly proportion of the participants was very low, the proportion of children was higher, and the family size was larger. The household differences provided evidence for the necessity of using the simplified "OECD equivalent scale" formula to adjust the income levels.
There were no significant differences between the two groups for the variable Disable, which indicated that this variable did not affect the households' choice to participate in nonagricultural employment.
The reason for the poverty of impoverished households was that disability or illness means that some family members of impoverished households had no labor force or weak labor ability, and often need the care of other family members. Disable have been previously highlighted as having a significant impact on household participation in non-agricultural employment [22,23]. On the one hand, the need to take care of household members with disabilities or severe illnesses will reduce the willingness of other household members to go out to work. On the one hand, the need to take care of household members with disabilities or severe illnesses will reduce the willingness of other household members to go out for work. On the other hand, the need to compensate for the decline in income and rising ex- penditures brought about by family members without labor and weak labor will make other healthy household members to more actively seek ways to improve their income. In addition, according to the division of the central government, the causes of poverty are divided into 11 categories. All the samples selected in this paper are impoverished households. Therefore, households who were not impoverished due to illness or disability have other causes of poverty such as suffering from disasters.

The ESR Model
Generally, the choice of whether to participate in non-agricultural employment is mainly based on the spontaneous choice of household, and the resource endowment of the household will affect their choice of participating in non-agricultural employment. In this case, the self-selection of the household is the source of endogeneity. For example, households with sufficient labor may be more inclined to participate in non-agricultural employment. If the self-selection bias was not controlled, it was likely to overestimate the income increase effect of non-agricultural employment. During the implementation of the targeted poverty alleviation policy, the choice of whether to participate in non-agricultural employment might also be influenced by the encouragement of local government staff. Government staff would spend more effort on helping impoverished households increase their income. Since participation in non-agricultural employment was considered an effective way to increase income, government staff would be more inclined to help households with lower incomes to participate in non-agricultural employment. If the selection of government staff was not controlled, it might underestimate the income increase effect of non-agricultural employment. The selection bias can be controlled through the simultaneous selection model and income model. In addition, it should be considered that the impact of covariates in the income model was different between participants and non-participants. For example, households who participate in non-agricultural employment can benefit more from education, so it is necessary to separate the income functions of participants and non-participants. The use of the ESR model can meet the above requirements well.
To account for both endogeneity and sample selection, the ESR model, which was employed to evaluate the poverty alleviation effect of non-agricultural employment, was designed as a two-stage model. A probit model was used in the first stage model to estimate the possibility of different household types participating in non-agricultural employment, with the estimation results being added to the second stage model to eliminate any selection bias. The first stage model was as follows: where C * i was a latent variable that captured the expected benefits of the participation choices with respect to not participating, Z i was a variable matrix that affected whether the rural households participated in non-agricultural employment, α was a vector of the coefficients to be estimated, and η i was a residual term.
The second stage model was the household income model. The income function for whether to participate in non-agricultural employment in each case is estimated separately as follow: where y i was the average income of the poor households, X i were the variables affecting the income, and ε ji , ( j = 0, 1) was the error term.
The ESR model parameters were obtained using maximum likelihood estimation. In the ESR model, ρ j denoted the correlation coefficient between the error term ε ji of the second stage model and the error term η i of the first stage model [24]. The estimated results for index ρ j measured the endogenous transformation results of the second stage model. ρ 1 < 0 implied there was a positive selection bias, which indicated that the households with income higher than average were more likely to participate in non-agricultural employment; otherwise, it indicated a negative selection bias [25]. The ESR model parameters were estimated using the complete information maximum likelihood estimation method.
The average treatment effect (ATT) of the participants and the average treatment effect of the untreated (ATU) non-participants was calculated to estimate the participants' nonparticipation and the non-participants' participation under a counterfactual framework, which was calculated as follows: where AverInc 1i was the per capita income of the participating households, AverInc 0i was the income of the non-participating households, with C i = 1 indicating participation in nonagricultural employment, and C i = 0 indicating non-participation in non-agricultural employment.

Identification Variables
The identification of the ESR model requires that the first-stage model contains at least one variable that does not directly affect the second-stage dependent variable. Although the model may be identified by construction through non-linearities generated in the selection equation, it is important for the identification variables in the selection model to contain an instrument for a more robust identification [25].
Compulsory education was selected as an identification variable in the model of the effect of non-agricultural employment on per income and agricultural income. Due to the backwardness of education in deeply impoverished areas, many rural residents were uneducated. In the sample, more than 46% of households without household members completed the nineyear compulsory education, and only about 2% of households with household members have a high school degree or above. For non-agricultural employment, there was no difference in the employment choice of the residents with high school education or below, most of them engaged in jobs that did not require education. In terms of agricultural production, since traditional farming methods were still widely used in deeply impoverished areas, relatively good education have little effect on improving rural productivity [26]. Although the completion of compulsory education itself did not directly affect income, the residents completed compulsory education would have the ability to read, write and calculate, which made them easier to adapt to urban life. On the other hand, education would broaden residents' horizons and directly increase their willingness to leave the countryside. Therefore, residents completed compulsory education were more inclined to participate in non-agricultural employment, whether they had completed compulsory education can be considered as an identifying variable. When the dependent variable is agricultural income, it is obvious that the transfer income brought by the subsistence allowance has no direct relationship with agricultural income. So, SA can also be used as an identification variable in the agriculture income model. After adjusting income, due to the elimination of the scale effect, there was no longer a direct correlation between per capita income and the number of households. Therefore, in the model for measuring the impact of non-agricultural employment on households' welfare, the number of households' members could also be used as an identification variable.
The admissibility of the identification variable can be established by performing a simple falsification test [25], that is, through basic regression analysis to determine the relationship between identification variables and selection variables and income. The results showed that (table 3) the regression coefficient of Education was statistically significant regression only in the selection function, while Household was not statistically significant when the dependent variable was adjusted income. When the dependent variable was agricultural income, the regression coefficient of Education was statistically significantly negative, but this did not mean that completing the nine-year compulsory education would harm agricultural production. The Education affected agricultural income by influencing the choice of participating in non-agricultural labor. When the mediating effect of Education and C was controlled, the p-value of Education was 0.142, which meant that completion of compulsory education did not directly affect agricultural income. In addition, identification variables have the nature of instrumental variables, so identification variables must have a strong correlation with endogenous explanatory variables. This paper used cragg-donald test to make sure the selected instrumental variables were not weak instrumental variables. The F value was greater than 10 for all the variables (table 4), so it can be considered that the instrumental variable was not a weak instrumental variable. In summary, the selection of identification variables in this paper was effective.
Firstly, Education was selected as the identification variable of the total income model and agricultural income model. Education and Household were selected as the identification variables of the welfare model.  Significance level: * 10%, * * 5%, * * * 1%, for the t-test and χ 2 test, respectively. Continuous data are presented as mean (standard deviations), and discrete data are presented as frequencies with percentages.

Factors Affecting the Participation
The first stage model regression results indicated that the coefficients of educational status (Educational), availability of Subsistence allowance receiving (SA), and road status (Road) were all positive (table 5). Subsistence allowances and access road conditions reflected the local government's selection effect on households' participation in non-agricultural employment. Households with subsistence allowances would receive more attention from the local government. Terrible access road conditions also meant that the household was more impoverished than other residents. The government would give priority to helping more impoverished households, and encouraging households to participate in non-agricultural labor was one of the main poverty alleviation policies.    The coefficients of variables related to household structure showed that households with a high proportion of the elderly and children are more inclined to not participate in nonagricultural employment, while the Student has a positive coefficient which is not statistically significant. It should be noted that there is no multi-collinearity problem between Young and Student. Removing Young or Student from the model alone has very little impact on the estimation of the remaining variables. The expansion of household size will increase the willingness of households member to participate in non-agricultural employment (table 5). The elderly and children in the household need the care of adults, and participating in non-agricultural employment will make it difficult for adults to return home frequently to deal with trivial matters related to the elderly and children. For rural residents living in deeply impoverished areas, the long-distance and inconvenient transportation usually requires a couple of hours from home to the nearby county. After controlling the proportion of the elderly and children, a household with more members means that there are more adults. Even if some members go out for work, other adults who stay at home can take care of the elderly and children.
In addition, there was a positive tendency in the regression coefficient of Disabled, Farmland and Student, but it was not statistically significant (table 5).

Impact on Per Capita Total Income and Agricultural Income
For the total income model, ρ 1 was significantly positive and ρ 0 was significantly negative, which implied that the hypothesis that there was no sample selection bias might not be rejected. For participated households, The expansion of household size had a negative effect on income, which might be caused by not taking into account the factors of economic sharing within the household. The increase in the ratio of elderly to children also had a negative impact on household income. Households with students had a lower income because students at school cannot provide income at all. Most of the coefficients in the non-participating group were not significant (table 6).
For the agriculture model, both ρ 0 and ρ 0 were not significant, which indicates that the selection deviation may not exist. The main factor affecting agricultural income, cultivated land area, was not significant in the selection model, which implied the absence of selection bias. However, for the convenience of comparison, we still used the ESR model to evaluate the impact of participating in non-agricultural employment on agricultural income. The increase of cultivated land area and the improvement of access road conditions had a significant positive impact on the agricultural income of participated households, and the coefficient of Disabled was also significantly positive. The estimation results of most other variable coefficients were not significant. Similar to the average income model, most of the regression coefficients of the non-participated group were not significant. Only the increase of the proportion of the elderly had a perceptible negative effect on agricultural income. The household size of the non-participants was smaller, and the increase of the elderly led to a severer problem of labor shortage (table 6).
If the non-participants participated in non-agricultural employment, their per capita income would reduce by 2551 CNY and their agricultural income would reduce by 2352 CNY (table 7). The roughly equal reductions in these incomes indicated that even if they tried to participate in non-agricultural employment, they would find it difficult to find jobs.
The decision to participate in non-agricultural employment increased the per capita income of participating households by about 30% ; however, the per capita agricultural income decreased by about 70% (table 7). After seeking non-agricultural employment, rural households would tend to spend less time and investment in agricultural production as the nonagricultural employment income would replace the agricultural production income to some extent. Generally, the substitution of the low agricultural incomes with non-agricultural employment income mostly depended on the rural labor characteristics, which was also the reason that households seldom considered their cultivated land area when choosing to participate in non-agricultural employment.
Comparing the total income and agricultural income, it could be found that agricultural income accounted for only half of the total income of non-participants, while the agricultural income and non-agricultural employment income of participants accounted for most of the total income. Because the other operating income and property income of impoverished households were very small, it could be considered that the income of the household was mainly composed of non-agricultural employment income, agricultural income and transfer income. The transfer income of non-participated households accounted for nearly half of the total income, which showed that government financial subsidies and assistance from relatives were very important to non-participated households.
figure adjusted using the simplified "OECD equivalent scale" formula to represent household welfare, the non-agricultural employment participation also significantly increased the welfare of the participating households (table 8). The difference between the estimated results before and after the adjustment was caused by the structural household participant characteristics (table 9).
Participants' household benefits rose by about 23.5%, slightly less than the increase in income. In the case of not participating in non-agricultural work, the household welfare of the participants is slightly higher than that of the non-participants. This is because participants had twice as many household members on average, twice as many children, and 15% of the elderly compared to non-participants. Income adjustments reveal the true contribution of non-agriculture employment to participant welfare, that is, a group of people who were living worse off life cannot improve their living conditions by participating in non-agriculture employment.

Robustness Check
The adjusted income model can be regarded as the robustness test of the results of the per capita gross income model. To test the robustness of the results of other models, this paper compared the results of the two equivalent scale formulas, and replaced the identification variables of the agricultural income model and adjusted income model.

Comparison of Equivalent Formulas
Since the OECD launched the equivalent scale formula in the 1980s, two modifications have been made [27]. This paper used the formula after the first modification. The original formula, also known as the Oxford scale, assigned 1 for the first household member, 0.7 for each  additional adult, and 0.5 for each child. The third edition formula is called the square root scale, which uses a scale that divides household income by the square root of household size. In general, there was no accepted method for determining equivalence scales, so comparing the impact of three equivalent formulas on the results was necessary.
The use of the Oxford scale for adjustment underestimated the effect of non-agricultural employment on welfare. Using the square root scale will overestimate the damage of nonagricultural employment on the welfare of the non-participating group (table 10). But generally, the conclusions obtained using three different methods were consistent.

Replacement of Identification Variable
SA could also be used as the identification variable of the agricultural income model. The identification variables of the agricultural income model were changed to SA and SA + Education, and named model1 and model2 respectively. The estimation results of ATT and ATU were still roughly consistent with the above (table 11). Similarly, for the adjusted model, this paper also calculated the case where only Education was used as the identification variable, which was named model3. Similarly, the estimation results of ATT and ATU were still roughly consistent with the above (table 11). Given all of that, The results of this paper can be considered robust.

Heterogeneity of Rural Households
Comparing the average income of participated households and non-participated households, the average income of participants was only about 8 % higher than that of non-participated households, but the income composition of the two groups was quite different. The nonparticipant showed a high dependence on transfer payment. The agricultural income model showed that even if they did not participate in non-agricultural employment, the agricultural income of the participating group would be higher than that of the non-participating group. Similarly, the non-participating group had no ability to participate in non-agricultural employment, which meant that on the one hand, these impoverished households needed continuous subsidies from the government. On the other hand, As subsidies depend entirely on the policy decisions of the high-level government, households' subsidies can be regarded as exogenous for a long time, which indicates that the subsidies received by households will not change much. Households in deeply impoverished areas were different from ordinary poor regions. In deeply impoverished areas, subsidy income did not form capital but was mainly used to increase expenditure. Therefore, it made a relatively weak contribution to enhancing household livelihood capital and a rather vital contribution to improving household welfare.
Stabilizing the employment of participants was the top priority to consolidating the achievement of poverty alleviation. For participants, their income has been greatly increased by the participation of non-agricultural employment, and non-agricultural employment income had become the most important source of income for them. The low agricultural income of participants was mainly due to the transfer of the labor force to non-agricultural employment. Increasing the income of non-agricultural employment was the practical way to further improve the income. Specifically, increasing the number of non-agricultural workers and learning labor skills were feasible paths to further increase income.

Education and Labor Skills Training
Promoting education and labor skills training for farmers is an effective means of increasing income for most households. Education and workforce skills training have been found to be key determinants of improving the status of employed farmers [28]. Rural residents who received education and workforce skills training have the opportunity to find higher-income jobs. The selection model showed that some uninsured households may choose to participate in non-agricultural employment after education and training to increase their income. However, due to the low education level of the adult population in deeply impoverished areas, it is difficult to improve the human capital of rural households, and it may be difficult to achieve significant income growth in a short time. However, in the long run, improving the education of children in deeply impoverished areas can have a significant effect on resolving intergenerational poverty.

The Elderly-caring and Children-rearing
The results of the selection model showed that solving the problems of the caring of the elderly and the rearing of children would allow more impoverished households to participate in non-agricultural employment. Making children migrate with their parents is an effective way to solve the problem of child-rearing, but there are some problems in practice. For example, migrant children got the same education and support as urban children were difficult [29], migrant children were easy targets of discrimination [30]. Improving the social service system for the care of left-behind children is also a way to solve the problem of children-rearing.
The support of the government is crucial to solving the elderly-caring problem. The social service and security system for the elderly in deeply impoverished areas is not developed well. Enhancing the accessibility of community services requires government input. Rural elderly residents are at higher risk of depressive symptoms. The promotion of rural elderly's mental health should be a priority, especially by allocating more public resources to help the elderly establish social connections and cultivate their positive attitudes towards aging [31]. Pensions provided by governments can increase the income of the elderly while enhancing community cohesion [32]. Providing elderly care services by the government or non-profit organizations will effectively solve the worries of rural households participating in non-agricultural employment.

Conclusion
This paper empirically examined the impacts of non-agricultural employment on household income and welfare in a deeply impoverished area in China. To reduce the selection bias, an ESR model was employed to analyze the poverty alleviation effects of participating in non-agricultural employment on per capita total income, agricultural income, and household welfare under a counterfactual inference framework.
To correct for the underestimation of the poverty reduction effectiveness caused by direct income analysis, a simplified "OECD equivalent scale" formula was applied to adjust the per capita income and estimate the overall household welfare. Non-agricultural employment was found to effectively increase the income of participants, alleviate participant poverty, and reduce agricultural income. However, for the non-participants, who were limited by their labor force quality, giving up agriculture and participating in non-agricultural employment was found to reduce total income and deepen household poverty.
Ensuring existing non-agricultural employment and maintaining subsidies to impoverished households are necessary means to consolidate the effectiveness of poverty alleviation for the moment. In the medium term, providing skill training for rural residents, solving the problems of supporting the elderly and children, and introducing advanced agricultural technology are effective ways for the further development of deeply impoverished areas. In the longer term, the government should focus on developing educational facilities in deeply impoverished areas, improving the quality of the rural labor force to eliminating poverty.
Finally, caution should be taken in interpreting these findings because of the data limitations. First, the use of cross-sectional data made it impossible to capture the household structure and income dynamics. Second, the lack of distinction between the number of people who participated in non-agricultural employment and the lack of focus on the specific employment sectors prevented any further analysis on the effect of non-agricultural employment on rural poverty alleviation; therefore, these considerations need to be included in future work.