Promoting fruit and vegetable consumption in rural China: Does off-farm work play a role?

Although off-farm work plays a significant role in facilitating agricultural production and rural develop- ment and improving household welfare, little is known about whether off-farm work can promote fruit and vegetable consumption in rural areas of developing countries. This paper sheds new insights by estimating the impact of off-farm work on fruit and vegetable consumption, measured by purchasing frequencies and consumption expenditures. We employ a two-stage residual inclusion estimator to ad- dress the self-selection bias and analyze data collected from 558 rural households in China. The results show that household heads’ off-farm work promotes rural households’ fruit and vegetable consump- tion by significantly increasing purchasing frequencies and expenditures. Further analysis confirms that household heads’ off-farm work participation, rather than all household members, plays a prominent role in promoting household fruit and vegetable consumption. We also find that farmers’ behaviours of growing fruits and vegetables appear to substitute their purchasing behaviours.


Introduction
Fruits and vegetables are essential parts of a healthy diet. Their inadequate daily consumption would result in micronutrient deficiencies and a wide range of diseases (e.g. heart diseases, cancer, and diabetes). It is reported that low fruit and vegetable consumption leads to around 1.7 million (2.8 per cent) deaths worldwide (WHO-FAO 2004). Specifically, insufficient fruit and vegetable consumption has caused appropriately 9 per cent of stroke deaths, 11 per cent of ischaemic heart disease deaths, and 14 per cent of gastrointestinal cancer deaths (WHO-FAO 2004).
The World Health Organization (WHO) recommends that fruit and vegetable consumption (excluding potatoes and other starchy tubers) should be no less than 146 kg/year/capita (equivalent to 400 g/day/capita) (WHO-FAO 2004;Choudhury et al. 2020). However, fruit and vegetable consumption is generally low worldwide, especially in rural regions of many developing countries (Dehghan, Akhtar-Danesh, and Merchant 2011;Kalmpourtzidou, Eilander, and Talsma 2020). In their meta-analysis for the vegetable intake of 162 countries, Kalmpourtzidou et al. (2020) found that vegetable intake in 88 per cent of the investigated countries is below the WHO recommended level. The significant benefits of fruit and vegetable consumption emphasize that it is vital to promote their consumption for shitting towards healthier and more sustainable diets. Ensuring sustainable consumption patterns of fruits and vegetables also contributes to the achievements of the Sustainable Development Goals of the United Nations regarding 'sustainable consumption and production' (Goal 12).
Agricultural production can affect the nutrition intake of rural households as sources of food and income (Gillespie, Harris, and Kadiyala 2012;Takeshima et al. 2020). Crops cultivated by the households can be translated into food for consumption, directly influencing the nutrition intake of rural people. Agriculture can also generate income through the marketed sales of food produced, facilitating rural households to purchase other food items such as fruits and vegetables for nutrition intake. However, relying solely on agriculture may be insufficient to improve dietary diversity and nutrition outcomes because, in reality, each household is cultivating limited land. Besides, income generated from agriculture is unstable due to, for example, market volatility and seasonality and uncertainties of weather conditions. As land is usually scarce for many rural households, diversifying household income is a more realistic strategy to diversify nutrition intake.
Low income is one of the most important barriers restricting people's adequate food consumption in general and fruit and vegetable intake in particular (Lallukka et al. 2010;Li et al. 2017;Kehoe et al. 2019;Choudhury et al. 2020;Kasprzak et al. 2020). Hall et al. (2009) investigated fruit and vegetable consumption of 196,737 adults from fifty-two lowand middle-income countries and found that around 77.6 per cent of men and 78.4 per cent of women have consumed fruits and vegetables below the recommended minimum level. Restrepo, Rabbitt, and Gregory (2021) reported that loss of income due to coronavirusinduced unemployment reduces household food expenditure in the United States. Lower income reduces consumers' purchasing power and their consumption diversity (Chai, Rohde, and Silber 2015). From sustainable development perspectives, there is a need to identify appropriate practical strategies that help increase incomes to boost fruit and vegetable intake.
Off-farm work is such a strategy. The importance of off-farm work in improving agricultural production, increasing rural incomes, and facilitating rural development has been well documented (e.g. Mathenge, Smale, and Tschirley 2015;Van den Broeck and Maertens 2017;Ma, Renwick, and Grafton 2018;Anang, Nkrumah-Ennin, and Nyaaba 2020;Zhou et al. 2020;Duong, Thanh, and Ancev 2021;Janvry and Sadoulet 2001;Zheng et al. 2021). There exist two strands of literature regarding the nexus between off-farm work and food consumption and security. The first strand of literature focuses on the impact of off-farm work/income on food expenditure (Chang and Mishra 2008;Chang and Yen 2010;Mishra, Mottaleb, and Mohanty 2015;Zereyesus et al. 2017;Liu, Renwick, and Fu 2019). Using a nationwide farm household survey in the United States, Chang and Mishra (2008) showed that the operator's off-farm work participation increases food expenditure, while the spouse's off-farm work participation decreases food expenditure. The study for Ghana Promoting fruit and vegetable consumption in rural China 3 by Zereyesus et al. (2017) reveals that participation in off-farm work significantly increases the future expected food consumption, thereby alleviating households' vulnerability to food poverty. Liu et al. (2019) found an increase in off-farm income increases per capita food expenditure in rural China.
The second focuses on the impact of off-farm work on food security and nutrition (e.g. Owusu, Abdulai, and Abdul-Rahman 2011;Tsiboe, Zereyesus, and Osei 2016;Zereyesus et al. 2017;Kuwornu et al. 2018;Dzanku 2019;Pritchard, Rammohan, and Vicol 2019;Rahman and Mishra 2020;Schmidt, Mueller, and Rosenbach 2020). For example, Babatunde and Qaim (2010) revealed that off-farm income positively impacts food security and nutrition in Nigeria. By estimating data collected from six African countries (Ghana, Kenya, Malawi, Mozambique, Tanzania, and Zambia), Dzanku (2019) found that off-farm income has a positive and significant impact on food security, and such impacts on female-headed and poor region households are larger than it has on male-headed and rich region households in most countries. The study for rural Vietnam by Duong et al. (2021) reveals that off-farm employment ensures food security and contributes to poverty alleviation.
Despite the existence of numerous studies on off-farm work effects, there is still a lack of research exploring whether off-farm work can help promote fruit and vegetable consumption. Therefore, the present study contributes to the literature by furthering our understanding of potential strategies for promoting fruit and vegetable consumption, focusing on offfarm work. We consider two important consumption dimensions, including fruit and vegetable purchasing frequencies and consumption expenditures. The off-farm work variable is potentially endogenous because of its non-random nature (Owusu et al. 2011;Duong et al. 2021). Therefore, we employ an innovative two-stage residual inclusion (2SRI) approach to address the endogeneity issue of off-farm work. The data analyzed in this study are collected from rural households in the Shandong, Henan, and Anhui provinces of China. As a further contribution, we estimate how the off-farm income share and the off-farm worker share in a household affect fruit and vegetable consumption. As an additional understanding, we also provide evidence to show whether farmers' behaviours of growing fruits and vegetables substitute or complement their purchasing behaviours.
China is an interesting example. Due to the increased production of processed food, rapid urbanization, and changing lifestyles, people consume more foods with high energy, free sugars, salt but fewer fruits and vegetables. A survey showed that, in 2018, the fruits and vegetables consumed by rural households were, respectively, 36.3 and 85.6 kg/capita (121.9 kg/capita in total) (NBS 2020), which are lower than the WHO recommended level (146 kg/year/capita). WHO (2020) reports that people in China consume excessive amounts of cooking oil and more than double the recommended daily salt intake. An unhealthy diet leads to a rising prevalence of obesity and overweight and other body diseases such as cancers, diabetes, and cardiovascular diseases. China is now facing a high prevalence of diet-related non-communicable diseases (e.g. hypertension, stroke, and coronary heart disease) and micronutrient deficiency among adults and children, increasing the burden of healthcare costs (Yang et al. 2008;Wong et al. 2013;Chen et al. 2020;Huang et al. 2020). Given the importance of consuming more fruits and vegetables in a healthy diet, it is significant to promote their consumption in China. This study's findings would provide significant evidence that helps design programs and strategies for promoting fruit and vegetable consumption in China and other countries to improve individual and public health.
The rest of this paper is structured as follows. Section 2 introduces the estimation strategy, followed by Section 3, which introduces data and descriptive statistics. We present and discuss the empirical results in Section 4. The final section concludes and proposes implications.

Model selection
Rural farmers themselves decide whether or not to participate in off-farm work. Their participation decisions are affected by both observed factors (e.g. age, education, farm size, and household size) and unobserved factors (e.g. motivations to work and innate abilities) (Owusu et al. 2011;Kuwornu et al. 2018;Ma, Abdulai, and Ma 2018;Anang et al. 2020;Zhou et al. 2020). The fact indicates that off-farm work is an endogenous variable in the models estimating off-farm work impacts on fruit and vegetable consumption. Failing to address such an endogeneity issue would cause biased estimates.
The existing literature has applied different approaches to address the endogeneity issue of the treatment variable (off-farm work in this study). These include the propensity score matching (PSM) model (Owusu et al. 2011;Khachatryan, Baghdasaryan, and Hartarska 2019), difference-in-difference (DID) model (Azam and Saing 2017), PSM-DID model (Liu, Tan, and Zhang 2021), two-stage predictor substitution (2SPS) approach , and the 2SRI approach (Kumar et al. 2020;. The PSM model cannot identify the impacts of control variables on outcome variables due to its non-parametric nature, and it cannot consider the selection bias arising from unobserved characteristics. Both DID model and PSM-DID model require panel data for estimations, so they are not applicable as this study uses cross-sectional data. In comparison, two parametric methods, including the 2SPS approach and the 2SRI approach, can overcome the PSM approach's drawbacks and estimate cross-sectional data. As argued by Terza, Basu, and Rathouz (2008) and Ying, Xu, and Murphy (2019), the 2SRI can provide consistent estimates of structural parameters and average partial/marginal effects while the 2SPS may be inconsistent. Thus, the estimation of the 2SRI model is usually more reliable than that of the 2SPS model. Several studies have applied the 2SRI model in their estimations (Kumar et al. 2020;Ma and Zhu 2020;Tesfaye and Tirivayi 2020;. For example, by applying the 2SRI approach, Kumar et al. (2020) estimated the impact of food safety measure adoption on milk yield and profitability of smallholder dairy farms in India, while Ma and Zhu (2020) examined the impact of the existence of agricultural cooperatives in rural villages on cropland abandonment in China. In this study, we also employ the 2SRI model to estimate the impact of off-farm work on fruit and vegetable consumption.

2SRI model
The 2SRI model estimation involves two stages. In the first stage, it estimates farmers' off-farm work participation equation. This study focuses on off-farm wage work because we do not identify any household heads having self-employed off-farm work in our sample. Following previous studies (Fernandez-Cornejo et al. 2007;Owusu et al. 2011;, we assume that farmers compare the anticipated off-farm wage (W O ) with the reservation wage (W R ) when making off-farm work participation decisions. Here, reservation wage refers to the minimum wage that a farmer requires to participate in the labour markets. Because farmers are assumed to be rational in decisionmaking, they may choose to participate in the off-farm work if the anticipated off-farm wage is higher than the reservation wage. Let OF * be the difference between the anticipated off-farm wage and the reservation wage, that is OF * = W O − W R , farmers' decisions to participate in off-farm work can be modelled in a latent variable function as follows: where OF * represents the probabilities of off-farm work participation. It is identified by an observed variable OF i , indicating whether a respondent i participated in any off-farm work (OF i = 1 for off-farm work participants and OF i = 0 for non-participants); X i is a vector of variables that are expected to affect a respondent i's decision to participate in off-farm work; IV i is a vector of the instrumental variables used for model identification; and ε i is an error term. Equation (1) can be estimated by a dichotomous model (Logit model in this study). We select two variables to serve as instrument variables (IV) in Equation (1). The first (IV1) is a dichotomous variable measuring whether households' friends can offer help when finding off-farm work (1 = yes and 0 = no). Similarly, the second (IV2) identifies whether households can obtain assistance from neighbours on the off-farm job search (1 = yes and 0 = no). The choice of the two instruments is inspired by the significant role of social networks in labour allocation and agricultural production decisions, which has been highlighted in the existing studies (Ma and Wang 2020;Tesfaye and Tirivayi 2020;Zhou et al. 2020;Cole and Chen 2021). The assistance from friends and neighbours is expected to affect household heads' off-farm work participation, but it does not directly affect households' fruit and vegetable consumption. We have applied the Hansen J-tests to examine the issues of overidentification restrictions, and the results confirm the validity and effectiveness of the selected IVs.
In the second stage, the 2SRI approach estimates the impacts of off-farm work on fruit and vegetable consumption, captured by purchasing frequency and consumption expenditure. The models can be specified as follows: where PF i refers to fruit purchasing frequency or vegetable purchasing frequency; E * i is a latent variable representing fruit expenditure or vegetable expenditure, captured by the observed value of E i . In the present study, we define expenditure (E * i ) as a censored variable because some households in our sample did not purchase fruits or vegetables. OF i refers to off-farm work participation status; X i is defined earlier; Residual i refers to a residual term predicted after estimating the Equation (1), and it is included in Equations (2a) and (2b) to address the endogeneity issue of the off-farm work variable; γ i , δ i , η i , ϑ i , ω ii , and θ i are parameters to be estimated; μ i and ψ i are two error terms.
The dependent variable PF i in Equation (2a) is a count variable, and therefore, we employ a Poisson regression model to estimate it. Because the coefficient estimations of the Poisson regression cannot be explained as magnitudes, we follow previous studies (Erdogdu 2013;Ma and Wang 2020;Zhang et al. 2020) and calculate the incidence rate ratios (IRRs) of the independent variables to enhance our explanation and understanding. In particular, the IRR is calculated by exponentiating the count model regression coefficients, that is IRR = exp (coefficient). For example, the IRR of the off-farm work variable in Equation (2a) is calculated as e γi . The IRR estimates indicate the multiplicative effect of an independent variable on the dependent variable, telling us that the dependent variable is 'IRR' times its original value when one unit is added to an independent variable. Alternatively, one can say the incidence rate of the dependent variable increases by (IRR − 1) × 100/100 if IRR > 1 or decreases by (1 − IRR) × 100/100 if 0 < IRR < 1 when one unit is added to an independent variable. Moreover, the expenditure variable E i in Equation (2b) is censoring from zero as not all households purchase fruits or vegetables from the markets. Therefore, the Tobit model is employed to estimate Equation (2b).

Study context
Data collection was organized in the Shandong, Henan, and Anhui provinces of China between June and July 2019. Shandong is a coastal province of China and is part of the East China region. The province has a land area of 157,100 km 2 . Shandong is the second-most populous province of China, after Guangdong, with a population of 100.7 million by 2019 (NBS 2020). Shandong has a temperate climate, lying in the transition between the humid subtropical and humid continental zones with four distinct seasons. Annual precipitation is between 550 and 950 mm. Henan is located in the central part of China, which is a landlocked province of the country. The province has a land area of 167,000 km 2 . At the end of 2019, Henan has a population size of 96.4 million (NBS 2020). Henan has a diverse landscape with mountains in the west and floodplains in the east. The province has a temperate climate that is humid, subtropical to the south of the Yellow River, and bordering on humid continental to the north. Anhui province is part of the Central China region. The province has a land area of 140,100 km 2 and a population size of 63.66 million by 2019 (NBS 2020). Anhui is quite diverse topographically. The north of the province is part of the North China Plain, while the north-central areas are part of the Huai River watershed. As with topography, the province differs in climate from north to south. The north is more temperate and has more clear-cut seasons.
Shandong, Henan, and Anhui provinces differ significantly in terms of economic conditions, disposal incomes, and consumption levels. In 2019, the per capita GDPs were 70,129 Yuan/capita in Shandong, 55,825 Yuan/capita in Henan, and 58,072 Yuan/capita in Anhui (NBS 2020). The per capita disposable incomes were 31,597 Yuan in Shandong, 23,903 Yuan in Henan, and 26,415 Yuan in Anhui in 2019. Among the disposable income, the wage income (mainly from off-farm work) contributed to 57 per cent of disposable income in Shandong, 50 per cent in Henan, and 53 per cent in Anhui (CSY 2020). The net operating revenue (mainly from farm work) contributed to 21-23 per cent of disposable income in these three provinces. The consumption expenditures of urban and rural residents in Shandong were 26,731 and 12,309 Yuan/capita in Shandong, 21,972 and 11,546 Yuan/capita in Henan, and 23,782 and 14,546 Yuan/capita in Anhui (NBS 2020). In 2019, rural people employed in farm and off-farm sectors in Shandong, Henan, and Anhui provinces were 28.91, 44.17, and 29.69 million, respectively.

Sampling method
A multi-stage sampling technique was utilized to select rural households. The survey was conducted in Shandong, Henan, and Anhui provinces in China. After purposely selecting the three provinces, two cities from each province were randomly selected in the second stage. These include Linyi and Zaozhuang in Shandong, Xinyang and Zhumadian in Henan, and Suzhou and Huaibei in Anhui. Third, around two to three towns in each selected city were randomly chosen. Fourth, appropriately two to three villages were selected from each selected town. Finally, around ten to thirty farm households were randomly selected in each selected village. This sampling procedure helped collect a total sample of 558 rural households. Among them, 164 household heads participated in off-farm work in 2018, and the rest 394 did not.
Well-trained enumerators carried out a face-to-face interview with the respondents. The enumerators are all postgraduate students hired from an agricultural-based university in China. The questionnaire used for data collection covered information on fruit and vegetable consumption frequencies and expenditures, individual and household-level characteristics, socioeconomic characteristics, and various institutional and contextual variables. The survey also required respondents to provide comprehensive information on the off-farm work participation status of household heads and other household members, as well as total off-farm income. This helps enrich our understanding of the relationship between off-farm work and fruit and vegetable consumption from different dimensions. Rural households usually consume different types of fruits and vegetables without paying attention to how many grams they have eaten each day. Therefore, we focus on their purchasing frequency and total consumption expenditure rather than consumption quantity and types to simplify our analysis.  In our sample, the average expenditures on fruits and vegetables are 89 Yuan/month (1 Yuan = 0.15 USD) and 113 Yuan/month, respectively. Around 29 per cent of household heads had worked off the farm in 2018. Among control variables, we show that the average age of household heads was 56.5 years, and 60 per cent of them were male. There are approximately five members within sampled households. The number of elder members and child members accounts for 22 per cent and 16 per cent of total household size, respectively. Figs 1 and 2 depict the histogram distributions of fruit and vegetable purchasing frequencies, respectively. They show notable differences in fruit and vegetable purchasing frequencies between households with heads working off the farm (n = 164) and households without heads working off the farm (n = 394). Specifically, around 59 per cent of households with heads working off the farm purchased fruits between 2 and 6 times per month (Fig. 1a), while only 51 per cent of their counterparts without heads working off the farm do it in that way (Fig. 1b). Fig. 2 shows that around 6.7 per cent of households with heads working off the farm purchased vegetables only 1 or 2 times per month, while this value is almost doubled, which is 12.2 per cent, for households without heads working off the farm. In general, the information presented in Figs 1 and 2 suggests that household heads' off-farm work participation may have impacts on fruit and vegetable purchasing frequencies. Table 2 reports the mean differences of the dependent variables and control variables between off-farm work participants and non-participants. In general, we show that relative to non-participants, off-farm work participants have relatively higher frequencies of purchasing fruits and vegetables. They also spend more on both fruits and vegetables. The findings tend to indicate that off-farm work participation increases fruit and vegetable consumption in rural China. However, the mean comparison results are not rigorous because they do not consider other factors that may affect farmers' decisions to participate in off-farm work and their fruit and vegetable consumption decisions. Table 2 clearly shows that off-farm work participants and non-participants are systematically different in age, gender, and education of household heads, and elder member ratio. These observable differences highlight the necessity of employing a rigorous econometric approach such as the 2SRI model to analyze the impact of off-farm work on fruit and vegetable consumption.

Empirical results
This section discusses the empirical results estimated based on the 2SRI approach. Specifically, Section 4.1 presents and discusses the results estimated for the first stage of the 2SRI approach, which reports the factors affecting off-farm work participation. Sections 4.2 and 4.3 discuss the results estimated for the second-stage of the 2SRI approach, focusing on the impacts of off-farm work and other control variables on fruit and vegetable purchasing frequencies and consumption expenditure, respectively. Finally, we present and discuss some additional analysis results to improve our understanding of the nexus between off-farm work and fruit and vegetable consumption. Note: * < 0.10 and ** < 0.01. Standard deviation is presented in parentheses.

Factors affecting off-farm work participation of household heads
The results that report the factors determining farmers' decisions to work off the farm are presented in the second column of Table 3. As the coefficients of the variables are not straightforward in interpretation, we calculate and present the marginal effects in the last column of Table 3 to facilitate our explanation. In general, our estimates show that household heads' decisions to work off the farm are positively affected by their gender and education but negatively influenced by their age and elder and child member ratios. For example, the significant marginal effect of the gender variable suggests that male farmers are 8.7 percentage points more likely to participate in off-farm work relative to their female counterparts. In developing countries like China, men are more likely to migrate to urban regions to seek better-salary opportunities in off-farm sectors, while women are usually left at home to undertake household activities.
The age variable has a negative and statistically significant marginal effect. The finding suggests that farmers are 1.1 percentage points less likely to work off the farm if their age increases one year. Compared with younger farmers, older ones are usually less competitive
in the off-farm work markets due to unfavourable health conditions and working skills. For this reason, they are less motivated to work off the farm. Our results echo the findings of Cole and Chen (2021), who also reported a negative relationship between age and off-farm employment in New England. The positive and statistically significant marginal effects of the variables representing elder member ratio and child member ratio suggest that one unit increase in elder member ratio and child member ratio would decrease the probabilities of off-farm work participation by 15.8 and 41.5 percentage points, respectively. The findings are expected. Higher ratios of elder and child members mean a potential shortage of available labour force in a household, limiting the household heads' ability and motivation to leave home for off-farm works. Ahmed and Melesse (2018) provided similar results that members living in households without abundant labour are less likely to engage themselves in off-farm activities.  Table 4) to facilitate our discussion because the coefficient estimates of the Poisson regression model (columns 2 and 3) do not reflect the magnitudes. The residual of the off-farm work variable, which is predicted from the first-stage estimation, is included as an additional regressor to account for the endogeneity issue. The residual term's coefficient  Note: * < 0.10, ** < 0.05, and *** < 0.01. The reference province is Anhui. Standard errors are presented in parentheses.

Factors affecting fruit and vegetable purchasing frequencies
in column 2 of Table 4 is negative and statistically significant, suggesting the presence of endogeneity issues stemming from unobserved heterogeneities (Terza et al. 2008;Ying et al. 2019;.
The results show that off-farm work significantly increases fruit purchasing frequency by a factor of 1.430 and vegetable purchasing frequency by a factor of 1.312. The findings suggest that off-farm work participants purchase fruits and vegetables each month 1.430 and 1.312 times more than their non-participant counterparts, given the other variables are held constant in the model. Participation in off-farm work allows household heads to send remittance to their families for regularly purchasing fresh fruits and vegetables.
We show that rural households' fruit and vegetable purchasing frequencies are also affected by other factors. For example, the coefficients of the age variable are negative and statistically significant. The IRR estimates suggest that if the age of a household head were to increase by one year, the rate ratios for purchasing fruits and vegetables would be expected to decrease by 0.9 per cent [i.e. (1-0.991) × 100/100] and 1.9 per cent [i.e. (1-0.981) × 100/100], respectively, while holding all other variables in the model constant. By investigating nationally representative data of 170,847 adults from China, Li et al. (2017) also found that old age is associated with low fruit and vegetable consumption. The household Note: * < 0.10, ** < 0.05, and *** < 0.01. The reference province is Anhui. Standard errors are presented in parentheses. Fruit consumption expenditure and vegetable consumption expenditure are measured at 100 Yuan/month. size variable appears to have a positive and significant impact on fruit and vegetable purchasing frequencies. The IRR estimates show that an additional member increase in a household would increase the incidence rates of purchasing fruits by 5.7 per cent and purchasing vegetables by 5.2 per cent. The coefficient of elder member ratio variable is negative and statistically significant in the second column of Table 4. The corresponding IRR estimate suggests that one unit increase in the elder member ratio would result in a 31.7 per cent reduction in fruit purchasing incidence rate, holding other variables in the model constant. This is possibly because older people may prefer consuming their self-produced agricultural products rather than purchasing them from the markets. Table 5 demonstrates the results for the impacts of off-farm work and control variables on expenditures on fruits and vegetables, which are estimated by the Tobit model and Equation (2b). The significant coefficient of the off-farm work residual term in column 2 of Table 5 again confirms the existence of the endogeneity issue and the necessity of employing the 2SRI model to address it. The results show that the off-farm work variable's coefficients are positive and statistically significant, suggesting that monthly household expenditure on fruits and vegetables would increase if household heads participate in off-farm work. In particular, our estimates suggest that, on average, households with heads participating in off-farm work spend 96 Yuan more on fruits and 104 Yuan more on vegetables than households whose heads did not work off the farm. Off-farm work participation of household heads increases household income, allowing the households to purchase more fruits and vegetables to enrich diets. The findings are largely supported by the income-increasing effects of off-farm work, showing a positive relationship between off-farm work/income and household consumption (Mishra et al. 2015;Al-Amin and Hossain 2019;Hossain and Al-Amin 2019).

Factors affecting fruit and vegetable consumption expenditures
Other factors also affect fruit and vegetable consumption of rural households in China. For example, the variable representing the child member ratio has positive and statistically significant coefficients in Table 5. The findings suggest that households with more children younger than 15 years old spend more on fruits and vegetables, which are largely consistent with the results in the existing literature (Blanchette and Brug 2005;Qian et al. 2016). Fruit and vegetable eating habits initiated in childhood would persist into adulthood, and starting the habit from childhood could provide life-long benefits for health (Blanchette and Brug 2005;Hirvonen and Hoddinott 2017). A poor fruit and vegetable diet in childhood may sow the seeds of serious health problems associated with older adults, while early intervention by consuming more fruits and vegetables can help reap the maximum health benefits. Ownership of assets (farm machine in this study) has a negative and significant impact on vegetable consumption expenditure. As essential production equipment, farm machinery improves farmers' efficiency when undertaking farm activities (e.g. land preparation and pesticide application), motivating farmers' intention to plant vegetables themselves but reducing farmers' interests to purchase vegetables from the markets. This fact leads to reduced expenditure on vegetables.

Impacts of off-farm work intensity
Our analyses in Sections 4.2 and 4.3 have focused on household heads' binary decision of off-farm work participation. As an additional understanding, we investigate how off-farm work intensity affects fruit and vegetable consumption, using the 2SRI approach. The offfarm working intensity refers to the total time (measured in months) allocated to off-farm work activities by household heads. For the sake of simplicity, we only present the results of the off-farm work intensity variable. The results [see Estimation (1) in Table A1 in Appendix A] show that off-farm work intensity affects fruit and vegetable purchasing frequencies positively and significantly. In comparison, off-farm work intensity is only significantly associated with a higher vegetable consumption expenditure.

Impacts of off-farm income ratio and off-farm worker ratio
Previous analyses have exclusively focused on household heads. To enrich our analyses, we have also estimated the impacts of the off-farm income ratio and off-farm worker ratio at the household level on fruit and vegetable consumption, using the 2SRI approach. In particular, the off-farm income ratio refers to the ratio of off-farm income to total household income. The off-farm worker ratio refers to the ratio of the number of off-farm work participants in a household to household size. These two variables capture the different dimensions of off-farm work participation within a household. For the sake of simplicity, in Table A1 in Appendix A, we only present the results of the key variables estimated from the second-stage estimations of the 2SRI model.
Our estimates [see Estimation (2) in Table A1] show that a higher off-farm income ratio in a household has a significantly positive relationship with vegetable purchasing frequency and fruit consumption expenditure. Its impacts on fruit purchasing frequency and vegetable consumption expenditure are also positive though not statistically significant.
Off-farm work of household members would generate additional income and contribute to increased household income, empowering households' ability to purchase fruits and vegetables. Besides, a higher off-farm worker ratio is only positively and significantly (at a 10 per cent level) associated with higher fruit consumption expenditure [see Estimation (3) in Table A1]. By comparing the results in Tables 4 and 5 with the results in Table A1, we can conclude that the off-farm work participation of household heads, rather than all household members, plays a prominent role in promoting household fruit and vegetable consumption. For optimal intra-household resource allocation, households may choose to allocate off-farm income generated by the heads to food consumption and the off-farm income generated by other members to other household activities such as building houses, purchasing durable goods, and business investment.

The nexus between fruit and vegetable growing and household purchasing
Some surveyed farm households have grown fruits (e.g. peach, pear, watermelon, grape, and apricot) and vegetables (e.g. potato, tomato, bean, eggplant, chilli, and cucumber) by themselves either for commercial purposes or self-consumption. Thus, it is interesting to understand how (substitutively or complementarily) farmers' fruit and vegetable growing behaviours affect their purchasing behaviour. We have created two dummy variables, including fruit growing and vegetable growing, to identify the growing status. The fruit growing variable is given a value of one if a farm household produces any fruits and zero otherwise. Similarly, the vegetable growing variable is given a value of one if a farm household produces any vegetables and zero otherwise.
Afterwards, we estimate the impact of fruit and vegetable growing status on fruit and vegetable purchasing frequencies and expenditures. The results (Table A2 in Appendix A) show that farmers' fruit and vegetable growing behaviours affect their purchasing behaviours negatively, and the influences are almost all statistically significant (except for the impacts on fruit consumption expenditure). The findings suggest that farmers' fruit and vegetable growing behaviours and their purchasing behaviours are substitutes.

Conclusion
In essence, off-farm work helps diversify household income and relax capital constraints, enabling people to consume more food, including fruits and vegetables. However, to date, the empirical evidence on the association between off-farm work and fruit and vegetable consumption is rare. This study addressed this research gap by empirically analyzing household survey data collected from rural China. We employed the 2SRI model to address the endogeneity issue associated with off-farm work.
The empirical results showed that off-farm work participation of household heads significantly increases fruit and vegetable purchasing frequencies and consumption expenditures. We found that rural households' fruit and vegetable purchasing frequencies are also positively affected by household size and child member ratio but are negatively influenced by the age of household heads and elder member ratio. Consumption expenditure on fruits was found to be positively associated with household size and child member ratio, and that on vegetables was mainly determined by child member ratio and asset ownership. The additional analyses revealed that a higher off-farm income ratio is positively and significantly associated with an increased vegetable purchasing frequency and higher expenditure on fruits. A higher off-farm worker ratio is exclusively correlated with higher expenditure on fruits. Finally, we showed that farmers' behaviours of growing fruits and vegetables appear to substitute their purchasing behaviours.
Our findings highlight the importance of encouraging rural labourers, especially household heads, to participate in off-farm works in promoting household consumption of fruits and vegetables. By doing this, the government may consider further improving rural labourers' accessibility to off-farm work information by virtue of modern information and communication technologies such as smartphones. The findings that older and female household heads are less likely to participate in off-farm work suggest that designing targeted supporting policies and training programs for these farmers is essential to ensure they obtain an off-farm job when there is a need. Education enhances off-farm work participation. The finding highlights the importance of improving literacy education in rural areas.
This study is subject to two limitations. First, it is a cross-section analysis based on data collected from 558 households in three provinces in China. Thus, it is not possible to capture the time-variant characteristics that may affect peoples' consumption behaviours. Second, we have only focused on purchasing frequencies and expenditures when capturing rural households' intake of fruits and vegetables. It might be interesting for future studies to look at how off-farm work impacts nutrition intake (e.g. protein, fat, and fibre) as a higher quantity of fruit and vegetable intake is not necessarily associated with a nutritious diet. Future studies may also expand our analysis by investigating how the majority Han Chinese people's and the ethnic minority people's off-farm work participation influences their fruit and vegetable consumption. Traditionally, these two groups of people have different dietary habit structures. Table A1. Impact of off-farm income ratio and off-farm worker ratio at the household level, and off-farm work intensity of household heads on fruit and vegetable consumption. Note: * < 0.10, ** < 0.05, and *** < 0.01. Standard errors are presented in parentheses. Fruit consumption expenditure and vegetable consumption expenditure are measured at 100 Yuan/month. For the sake of simplicity, we only present the coefficients of the key explanatory variables for Estimations (1)-(3). Note: * < 0.10, ** < 0.05, and *** < 0.01. The reference province is Anhui. Standard errors are presented in parentheses. Fruit consumption expenditure and vegetable consumption expenditure are measured at 100 Yuan/month.