Persistent Agricultural Shocks and Child Poverty

Abstract This study shows how persistent agricultural shocks in Ethiopia affect education, health and labor outcomes through a time-use study of young people aged 5-22. Leveraging five rounds of the Young Lives Study from 2002-2016, we use dynamic panel instrumental variable regressions to account for the unobserved heterogeneity and serial correlation in the estimation. Agricultural shocks significantly reduce schooling participation and time spent in schooling, deteriorate health, and increase both labor force participation and labor time. Household wealth acts as a buffer and mitigates the adverse effects of shocks on schooling. Interestingly, children from wealthier households have a higher likelihood of joining agricultural labor during shocks, but their intensity of child labor is significantly lower compared to poorer households.


Introduction
The adverse impact of agricultural shocks and their threat to household and child welfare in the developing world has been documented for over four decades. 1 However, the persistent nature of agricultural shocks (floods, droughts, pests, soil erosion, and frosts, among others) has raised questions about the estimation techniques from previous research, highlighting issues of endogeneity and serial correlations.Households living in areas with recurring agricultural shocks might be adjusting their consumption behavior and time-use accordingly.For instance, in areas prone to recurring agricultural shocks, households may be stockpiling more and selling fewer yields, saving more and spending less on consumption and child education, and members may have higher migration rates to seek better economic opportunities (Giles & Yoo, 2007).These adaptive behaviors are likely to alter the time-use patterns of household members.In this context, a robust longitudinal study of how recurring agricultural shocks affect child education, health, labor, and time use dynamics has remained elusive.
In low-income countries, a large proportion of the population depends on subsistence agriculture, cash crop production, or hired farm labor as primary means of economic support.
Therein, rainfall and agricultural shocks have kept agricultural yield and demand for agricultural labor low, thus negatively affecting the economic conditions of agricultural households.Ethiopia is one such country, where about 85% of the population live in rain-fed rural areas and have been unable to protect themselves from rainfall failure that occurs, on average, every five years (Porter, 2012).Ethiopia is heavily reliant on small-scale rain-fed agricultural systems, highly localized agricultural markets, erratic climatic conditions, weak storage capacity, and high-yield transportation costs.These structural and systematic failures are argued to be the causes of recurrent agricultural shocks and food insecurity (Durevall, Loening, & Birru, 2013;Miller, 2017;Rijkers & S€ oderbom, 2013;Shah & Steinberg, 2017).Despite Ethiopia's steady progress in enhancing development indicators, there remains a persistent concern regarding vulnerability to severe weather conditions, such as the El Nino from 2015-2016 (Koo, Thurlow, ElDidi, Ringler, & De Pinto, 2019).This is especially true for individuals who rely on agriculture as their primary source of income.In addition, more recently, the growing scarcity of land for agricultural production has restricted younger generations from investing their time in agriculture and has disincentivized knowledge creation and transmission for improving agricultural output in Ethiopia (Bezu & Holden, 2014).
To the extent that agricultural households are often geographically isolated in regions with few off-farm employment opportunities or other mechanisms for coping with agricultural risk, precautionary under-investments in education and health may reinforce poverty traps (Giles & Yoo, 2007;Shah & Steinberg, 2017).Observers have also concluded that poor children suffer from "double jeopardy" as they are both more likely to suffer negative education and health shocks, and are less likely to recover from them (Becker & Tomes, 1986;Currie & Hyson, 1999).Early research on household consumption responses to agricultural risks has not allowed for the updating of risk perceptions with changes in household wealth due to the lack of dynamic panel data, or has introduced the impact of changes in expected wealth in an ad-hoc manner using constant absolute risk preferences or short panel surveys that do not allow for dynamic risk updating (Campbell, 1987;Jalan & Ravallion, 1999).More recent studies have shown that wealth indicators such as land size, assets and livestock holding play a significant role in explaining changes in household consumption and child labor (Amare, Jensen, Shiferaw, & Ciss e, 2018;Bhalotra & Heady, 2003).
Often neglected from the public domain and policy discussions, a recent report by the ILO and UNICEF in 2020 2 finds that since 2016, global child labor has increased for the first time in two decades.According to the report, the number of children in child labor has risen to 160 million worldwidean increase of 8.4 million children in the last four years.The majority of these children are engaged in agricultural labor and predominantly work on family farms (Bhalotra & Heady, 2003).In this context, examining how children reallocate their time and responsibilities in the event of agricultural shocks can provide us with a clearer picture of the costs of recurring agricultural shocks on human development, as schooling lost in formative years has been found to be a significant deterrent to quality of life in later years (Das, Singh, & Chang, 2022;Huebner, Hills, DA Ll, & Gilman, 2014).
Risk exposures and smoothing behaviors could evolve with changes in household wealth.In the context of certain agricultural-based economies in Africa, the impact of household land ownership on children's education has been found to be fairly positive (Charles & Hurst, 2003;Jensen, 2000).Nonetheless, in developing economies such as Ghana, Pakistan and Burkina Faso, having more agricultural land has led to increased child labor in the agricultural sector (more agricultural land leading to more child agricultural labor force participation), creating what has been regarded as the "wealth paradox" (Basu, Das, & Dutta, 2010;Bhalotra & Heady, 2003;Dumas, 2006).However, thus far, the analysis of the impact of wealth on child outcomes has focused on land ownership as an indicator of wealth but not on wealth in generalhousing quality, consumer durables and access to services.Focusing solely on land as natural capital has geographical limits, therefore, there is a need to examine the impact of physical capital as a buffer against agricultural shocks to inform more sustainable and resilient development.This is Persistent agricultural shocks and child poverty 31 especially important in the Ethiopian context, where the state controls the distribution of land, and the allocation of land for agricultural purposes is becoming increasingly difficult due to limited availability (Bezu & Holden, 2014).Therefore, analyzing non-land-based wealth for agricultural households and its impact on child outcomes during an agricultural shock provides alternative evidence on wealth as a buffer, and acts as a test for the land-based "wealth paradox".In addition, understanding whether household wealth acts as a cushion against agricultural shocks in a dynamic context is crucial given the improvements in living standards over time and relatively low levels of wealth inequality and a high degree of imperfection in credit and labor markets in Ethiopia (Ali, Deininger, & Harris, 2016;Thiede, 2014).
This study provides a longitudinal time-use analysis of the costs of agricultural shocks on education, health and labor force participation for young people, and the role of wealth in mitigating the adverse impacts of recurring agricultural shocks.The study fills several gaps in the literature on education, health and labor responses to agricultural shocks: (i) we use an exogenous proxy for recurring agricultural shocks by using a dynamic instrumental variables approach, (ii) we examine child labor and household investment in child education and health as agricultural uncertainty becomes predictable in a dynamic setting, (iii) we account for the role of non-land based household wealth in responding to agricultural shocks in a dynamic setup, and (iv) we analyze agricultural shocks and time-use dynamics of young people.
We leverage five rounds of the Young Lives Study (YLS) to conduct dynamic panel databased instrumental variable regressions to examine the effects of recurring agricultural shocks on human development outcomes for children (Blundell & Bond, 1998).With the YLS featuring repeated rounds of data on the same child over fifteen years, the dynamic panel data estimation uses generalized method of moment conditions with lagged differences of the dependent variables as instruments for the difference equation, and the lagged treatment variable as an instrument for the level equation.This approach eliminates potential serial correlation in the outcome variable in previous and current periods, and the error correlations between the treatment and outcome variables in the current period.We begin our estimation with linear ordinary least squares (OLS) regressions, then test model coefficients with two-way fixed effects (FE) regressions which control for time-invariant unobserved heterogeneity.Finally, we compare the OLS and FE estimates with the GMM-IV model to examine the nature of the selection bias in the coefficients of interest.
Results from OLS, FE and GMM-IV models show that agricultural shocks have a significant negative (positive) impact on the extensive margin of schooling (labor) for children aged 5 to 22. Specifically, agricultural shocks lead to children dropping out of school and joining the labor force.Agricultural shocks significantly reduce the reported health condition of children and increase the likelihood of child wasting.Concerning the intensive margin of time use, agricultural shocks significantly reduce the time children spend in school and disproportionately increase the time children allocate to agricultural labor.Agricultural shocks also increase leisure (play time) for children both at the extensive and intensive margins.Moreover, higher incidents of agricultural shocks imply higher increases in child labor and leisure and reductions in child education time.
We find that wealth acts as a buffer at the intensive margin by allowing wealthier households to mitigate the adverse impact of agricultural shocks on a child's time spent in schooling and labor.Moreover, wealth also serves as a buffer for schooling at the extensive margin (schooling participation of children differs significantly between wealthier and poorer households).However, children from wealthier households are more likely to start working in response to a shock.This positive and significant interactive effect of wealth and agricultural shocks on child labor, though initially surprising (as child labor is often portrayed as being negatively associated with household wealth), is not unexpected.In some agricultural settings, a positive association between agricultural landholding and child agricultural labor has been previously noted in Pakistan and Ghana, underscoring the failures in credit and labor markets in developing economies (Beegle, Dehejia, & Gatti, 2006;Bhalotra & Heady, 2003).We do not find a significant difference in health outcomes based on wealth.
When examining results by child age, wealth appears to buffer the intensive margin for schooling more for younger children (age <15) than for older children.In contrast, wealth buffers the intensive margin for labor more for older children than for younger children.We also find that younger wealthy children are induced to join the labor force by a shock, driven mostly by farm labor.This result highlights the substitution from hired labor to own household labor for wealthy households.However, there is a decline in the number of hours spent in labor activities with agricultural shocks for children from wealthy households relative to poorer households highlighting the wealth cushion effect at the intensive margin.

Data
We use data from the YLS, which conducts surveys of 2,000 children born between May 2001 and May 2002, and 1,000 children born between January 1994 and January 1995 in twenty sites (clusters) across Ethiopia (Outes-Leon & Sanchez, 2008).The surveys are sponsored by the Department for International Development (DfID) and collected by the Young Lives team at Oxford University.Data is available from five rounds of surveys conducted in 2002,2007,2009,2013 and 2016when children were approximately one, five, eight, twelve and fifteenyears old (younger cohort), and eight, twelve, fifteen, eighteen and twenty-two years old (older cohort).As young children are unlikely to attend school or participate in the labor force, observations collected when children were under five years old are excluded from all analyses.Each survey round occurred every three or four years, ensuring that children from both cohorts were interviewed at similar ages.The study collected information on child time use in education, labor and leisure, as well as categorical variables for general health.From an empirical standpoint, as people often engage in multiple activities concurrently, simultaneity could lead to biased estimates related to time-use.However, the issue of multiple activities being done simultaneously is less of a concern in this study as activities are aggregated at the broad level, and it is unlikely that studying and labor activities would be performed at the same time.
Our outcome variables, derived from the relevant child-level data, are the extensive margins of education and labor, general health, and the intensive margins of education, labor and leisure.The extensive margin outcome for education is set to 1 if any time is spent on school and study activities on a typical day, and 0 otherwise.The extensive margin for labor is 1 if any time is spent on farming, farm business or paid market activity, and 0 otherwise.The intensive margin outcomes (hours on a typical day) are contingent on at least some time spent on education and labor.Our categorical measure of child health is based on five relative states of health: very good, good, average, poor and very poor.We assign a code of 1 to the first three states (very good, good, and average) and a code of 0 to the last two states (poor and very poor). 3 While the study over-sampled poor and food-poor areas, the communities span the geographical regions where almost 97% of the Ethiopian population resides (Miller, 2017). 4 Our exposure measure is constructed both at the extensive and intensive margins based on the household survey questionnaire, which asks whether the household experienced any of the following agricultural shocks: droughts, flooding, soil erosion, frost, pests on crops, crop failure, pests on storage, pests on livestock and natural disasters since the previous survey. 5We first create a categorical variable that takes the value 1 if any of the above shocks affected the household since the last survey round, and 0 otherwise.At the intensive margin, we aggregate all the shocks and create a variable that takes values from zero to eight depending on the number of shocks experienced by the household.
Control variables in our analysis include the child's age in months, household size, educational background of the household head and the child's Body Mass Index (BMI).Our analysis predominately focuses on the child's education and labor force participation, as well as Persistent agricultural shocks and child poverty 33 examining household wealth as a buffer against agricultural shocks.To examine the effect of agricultural shocks, we include the child's BMI as a covariate and control for the health channel. 6Lastly, the YLS provided a constructed wealth index based on sub-indices of housing quality, access to services, and consumer durables (for details, see Briones (2018), also discussed in the empirical model section).We use the constructed household wealth index for each of the survey rounds to examine if wealth acts as a cushion for the impact of agricultural shocks.
Table 1 presents the descriptive statistics for households with and without agricultural shocks across the five survey rounds combined (2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016).The average number of shocks experienced by households with agricultural shocks is approximately two (specifically, 1.845) since the previous survey.Furthermore, households without agricultural shocks exhibit a lower cluster-level average of shocks since the last survey.At the extensive margin, there is a 10.2 percentage point (pp) higher likelihood of a child attending school in households without agricultural shocks compared to households without such shocks.Similarly, this is a 25 pp lower likelihood that a child from a household without agricultural shocks will enter the labor force.Children from households experiencing agricultural shocks report poorer general health levels compared to their counterparts, with a 6.9 pp decrease in the likelihood of reporting good health.Children from households without agricultural shocks are 5 pp less likely to experience wasting.On the intensive (time-use) margin, children in households without agricultural shocks spent about 1.5 more hours on education and 0.91 fewer hours in farm and paid work compared to households with agricultural shocks.As expected, households without agricultural shocks are wealthier, and their children have higher BMI than households with agricultural shocks.Figure A1 shows the kernel densities (with epenechnikov bandwidth ¼ 1) of the time spent in various activities by children with and without agricultural shocks.As evident from the figure, children from households with agricultural shocks allocate significantly less time to school and study activities (Panel A).Additionally, there is notably more time devoted to farming and other labor activities for children from households impacted by agricultural shocks (Panels B & C).

Empirical framework
Our estimation of the exposure effects of agricultural shocks on child education, labor and health begins with the two-way individual fixed effects specification, as proposed by Imai and Kim (2021) and contended by de Chaisemartin and D'Haultfoeuille (2020). 7The benchmark fixed effect estimator is as follows: Here Y it represents the outcome of interest: education, labor force participation, health or leisure for child i in survey year t.Shock it indicates agricultural shocks occurring between survey waves.X 0 it is a vector of household and child-level observable socioeconomic and demographic characteristics that could potentially affect the outcome variables of interest.These include the child's age, household size, child's BMI and household head's education.g i and d t are child and time-fixed effects (accounting for major macroeconomic trends).Finally, it is the error term.Our primary interest is on the effect of agricultural shocks on child welfare measures, namely b 1 .
In practice, agricultural shocks in the Ethiopian context are recurrent and not random (Debebe, 2010;Gebremariam & Tesfaye, 2018;Porter, 2012).Self-selection and sorting take place based on income, ethnicity, household size and knowledge of the neighborhood.Households might adapt their consumption, savings patterns, labor supply, and time allocation in anticipation of recurring agricultural shocks, as emphasized by Giles and Yoo (2007).These complexities prove challenging to address using observational data, leading to potential endogeneity issues and biases in estimated effects.
Our regressions control for an extensive set of household and individual observable characteristics that could also impact child outcomes.Important time-invariant characteristics such as ethnicity, religion, and geography are also controlled for by the child fixed effect g i , which also accounts for any time-invariant unobserved characteristics of the child.The time fixed effect d t further controls for aggregate time trends.However, the model's assumption of common trends (de Chaisemartin & D'Haultfoeuille, 2020) does not strictly apply in the case of agricultural shocks in Ethiopia.For example, the impact of agricultural shocks on schooling, health and labor might exhibit geographical variations across survey areas and shift over time.Moreover, fixed effect estimates are weighted sums of the average treatment effects (ATE) in each group (child in our study) and period, with potentially negative weights that could introduce an underestimation bias in the coefficients (de Chaisemartin & D'Haultfoeuille, 2020).To account for the endogeneity in time use with adaptive behavior, we use a dynamic panel data instrumental variables approach and compare it with the benchmark fixed effects model.The GMM-IV model is as follows: Where a 1 , ::::a q are the q parameters of the lagged values for the dependent variable.Y i, tÀj is the lag of the dependent variable, the order of which is determined by j.X it is a 1 Ã k 2 vector of covariates in the model: age of the child in months, survey waves, BMI of the child, household size and head's education.Shock it is a 1 Ã k 1 vector of the strictly endogenous agricultural shocks.As in our benchmark model, b 1 is the parameter of interest in our analysis.The above linear dynamic panel data model uses the generalized method of moment conditions (GMM) in which lagged differences of dependent variables are used as instruments for the differenced equation (GMM type: two lags of dependent variables) along with standard differences of the covariates.Furthermore, the model also uses moment conditions in which lagged levels of the Persistent agricultural shocks and child poverty 35 agricultural shocks are instruments for the level equation (GMM type: LD.agricultural shocks) (Blundell & Bond, 1998).g i are the panel-level effects (which may be correlated with Shock it or X it ), and it are independent and identically distributed over the whole sample with variance r 2 : In all our specifications, we use the standard two lags of the dependent variables as instruments for the difference equation. 8Additionally, since one individual is surveyed from each household, the standard errors are clustered at the household level.

Model for wealth effects
In order to analyze household wealth as a potential mechanism for heterogeneous effects of agricultural shocks on child welfare, we extend our analysis by estimating an augmented version of Eq. (1).In this extended model, we introduce an interaction term between non-land-based household wealth and the dummy variable representing agricultural shocks within a two-way fixed effect framework.Household wealth is a simple average of three sub-indiceshousing quality (HQ), consumer durables (CD), and access to services (AS)that range in value from zero to one.The HQ index is based on crowding (rooms per person) and the quality of materials used for the walls, roof, and floor.The CD index is based on the number of non-productive assets owned by the household.The AS index is constructed by taking a simple average of the following four dummy variables: access to safe drinking water, clean cooking fuel, electricity, and sanitation facilities.All these measures are used to measure deprivation following the Multidimensional Poverty Index (Todaro & Smith, 2009).
Due to the absence of a continuous IV for wealth and insufficient lagged outcome and treatment variables to utilize as instruments, we refrain from employing a GMM-IV model to examine the wealth effect.Given the data limitations and the empirical design, we are unable to address the endogeneity of wealth with respect to agricultural shocks.Consequently, the estimates of the two-way fixed effects model reflect associations rather than causal effects.Nonetheless, incorporating a non-land wealth measure allows us to avoid potential endogeneity resulting from a presumably strong negative correlation between land-based wealth and agricultural shocks.
We use the following specification to evaluate our hypothesis that household wealth acts as a cushion against child poverty during agricultural shocks: b1 effectively gives us the estimated effect of agricultural shocks for households with a very low wealth index, while b3 is the additional effect of agricultural shocks on child welfare as household wealth increases.As the wealth index is scaled from zero to one, b1 þ b3 is effectively the total effect of agricultural shocks on child welfare for households that have a very high wealth index.

Results
Our empirical analysis is presented in two sections.Firstly, we examine the impact of agricultural shocks on the extensive margins of schooling, labor and general health, along with the intensive margins of time allocated to schooling, labor and leisure.The second section explores the role of wealth in mitigating the impact of agricultural shocks on child welfare, considering both the extensive and intensive margins.
It is important to acknowledge that agricultural shocks are self-reported and may exhibit endogeneity over time.This endogeneity could stem from two main factors: (1) households adopting mitigation behavior that limits the frequency or intensity of shocks, and (2) households updating beliefs about what constitutes a "shock" as more adverse events are realized over time.Therefore, our results likely understate the true cost of these shock-type events because they do not capture costs associated with mitigation efforts or belief changes.Due to data and empirical approach limitations, we can only capture the costs associated with adaptive behavior.
For additional insights, Table A1 in the Supplementary Materials provides the correlations between current agricultural shocks and lagged values of agricultural shocks, wealth, and other time-use variables such as hours of schooling, farm and business labor, unpaid care and chores.The table shows that lagged variables have significant predictive power for future agricultural shocks.However, as we move from the parsimonious specification to ones including more variables, the predictive power reduces, though it remains significant.Most importantly, we observe that the previous period's agricultural shock is a significant positive predictor of the current period's agricultural shock.This observation substantiates our argument that agricultural shocks are persistent.

Agricultural shocks: schooling, labor and health
Table 2 presents the analysis from OLS, FE, and the GMM-IV models across three panels.The treatment variable is the count of agricultural shocks.Panel (a) shows the estimates of the GMM-IV model.At the extensive margin, a one unit increase in agricultural shock (equivalent to one additional shock) significantly reduces the likelihood of school participation by 3.8 pp (column 1), increases the likelihood of engaging in farm labor and business by 1.3 pp (column 2) and reduces the status of general health by 2.8 pp (column 3).At the intensive margin, a oneunit increase in agricultural shock significantly reduces school and study time by 0.33 hours  2002,2007,2009,2012, and 2016, for children above 4 years of age.In columns 1, 2, and 3, the b coefficient is interpreted as percentage point changes.In columns 4 and 5, the b coefficient is interpreted as hours.Dynamic panel data based over-identification criteria are met in all specifications in panel (a) (Andrews & Lu, 2001).Robust standard errors are clustered at the household level.Significance ÃÃÃ p < 0.01, ÃÃ p < 0.05, Ã p < 0.1.The instruments used in the GMM-IV analysis are: (i) two lags of the dependent variable for the differenced equation and (ii) one lag of the agricultural shock for the level equation.
Persistent agricultural shocks and child poverty 37 (column 4) and increases farm labor and business activity by 0.16 hours (column 5).Overall, supporting the previous findings in Africa, these results show that agricultural shocks are detrimental to child welfare and a deterrent to human capital formation in Ethiopia.
In panel (b), we show the two-way fixed effects model, which accounts for time-invariant heterogeneity and any macroeconomic trends in Ethiopia between 2002 and 2016.Results from the fixed effects analysis corroborate the findings from the GMM-IV regressions, displaying consistent signs and statistical significance for all outcome variables.However, the magnitude of impact is substantially lower for schooling (both at the extensive and intensive margins) in the fixed effect model.This suggests negative selection, possibly due to unobservable time-varying heterogeneity strongly correlated with more schooling hours and fewer agricultural shocks.Results suggest a bias towards zero in the fixed effects regressions compared to our GMM-IV estimate consistent with households updating their behavior in response to persistent agricultural shocks.The fixed effects models provide estimates based on the transition from one shock state to another in consecutive periods.However, these models fail to account for households adapting their behavior due to shocks occurring before the current period, which can lead to smaller estimates when comparing any two periods within the survey time frame (de Chaisemartin & D'Haultfoeuille, 2020).In addition to the issue of heterogeneous treatment effects (varying across units or time periods), the five waves of the YLS survey can lead to under-estimation with two-way fixed effects, as the model fails to compare newly treated units (at time t) to already treated units at time periods t -2, 3 & 4 (Callaway & Sant'Anna, 2021).Therefore, we argue that the GMM-IV model provides a robust estimation of the point estimates after taking into account the exposure to treatments in periods before t -1.The OLS model in panel (c) presents similar signs and significance as the GMM-IV model, albeit with anticipated larger coefficients due to a negative bias, except for schooling at the intensive margin.Overall, our results are robust to changes in econometric models.In all results in Table 2 (except the health result in column 3), we control for the child's BMI to account for the health effect of agricultural shocks on schooling and labor outcomes.
Table 3 shows the GMM-IV analysis of unconditional and conditional results for the outcome variables of interest: schooling, labor and leisure.Unconditional time spent in an activity includes children who spend zero minutes on that activity, whereas conditional time considers only children with positive time values (greater than zero minutes).The analysis shows non-linear results by examining any shock (dummy: 0/1) faced by the household, more than one shock faced by the household, and more than two shocks faced by the household.The first three columns (unconditional effects) in panel (a) show that any agricultural shock (0/1) in the household reduces time spent on schooling activities by 0.61 hours (37 minutes, from a comparable mean of 6.44 hours) 9 , increases time spent on labor activity by 0.32 hours (19 minutes, from a mean of 1.57 hours) and increases time spent on leisure activities by 0.21 hours (13 minutes, from a mean of 4.35 hours).The last three columns (conditional effects) show that any agricultural shock faced by the household reduces time spent on schooling activities by 0.32 hours (20 minutes, from a mean of 8.03 hours), increases labor time by 1.12 hours (72 minutes, from a mean of 4.54 hours) and increases leisure time by 0.22 hours (14 minutes, from a mean of 4.38 hours).Comparing the unconditional and conditional effects shows the impact of the switch from no time spent to at least some time spent in the activity.As anticipated, conditional effects on schooling are smaller than unconditional effects due to the notable categorical effect of shocks on school dropout.In contrast, conditional effects on labor are more pronounced than unconditional effects, highlighting an increased labor contribution among individuals already part of the labor force.There is no statistically significant difference between unconditional and conditional effects of shocks on leisure as we expect minimal switching effects from leisure to no leisure.
Results in panels (b) (more than 1 shock) and (c) (more than 2 shocks) reveal that the effects of agricultural shocks follow a non-linear pattern.The impact of higher-order shocks, both in conditional and unconditional margins, demonstrates that greater shock severity corresponds to more pronounced reductions in school participation and schooling time, as well as amplified labor force participation and labor time.The unconditional results for labor are mainly driven by entrance into the labor force (with 66% of the control mean having zero minutes of labor activity), highlighting the push factor for children to join the labor force during times of economic hardship created by a higher level of shocks.This can be understood through the lens of weak labor and credit markets in developing African countries, as exemplified by the case of Ghana (Bhalotra & Heady, 2003).Elevated agricultural shocks lead to reduced employment prospects due to decreased economic activity and heightened challenges in finding non-agricultural employment due to diminished labor demand and credit availability for alternative income-generating endeavors.Hence, households seek to employ children in farming activities to compensate for the lost labor.Overall, when household wealth is unaccounted for, persistent agricultural shocks in Ethiopia reduce children's school participation and schooling time while increasing labor force participation and labor time.

Wealth: buffer for agricultural shocks
As the state owns all land in Ethiopia, rural residents have been guaranteed land access through a law granting them the right to obtain agricultural land for free.However, it has become increasingly difficult to fulfill this right for the younger generation.In parts of the highlands where population densities are exceptionally high and farm sizes have shrunk, Ethiopia is grappling with significant land scarcity.Consequently, the safety net that land provided is diminishing, leading to the emergence of landlessness among the youth who are unable to access their parents' land.This trend is particularly pronounced in Southern Ethiopia, where a significant portion of farmers cultivates less than one hectare of land (Bezu & Holden, 2014).In this Persistent agricultural shocks and child poverty 39 context, examining the non-land-based environmental resilience of household wealth is a critical determinant of child poverty in Ethiopia.Table 4 presents an analysis exploring whether household wealth has a significant dynamic impact on the observed relationship between environmental shocks and child poverty.The table shows the results of two-way fixed effects regressions, focusing on the interactive impact of agricultural shocks and household wealth on children's allocation of time to schooling and labor.The analysis shows schooling, labor, and health results at the extensive margins, while the intensive margin results are shown for schooling and labor.The analysis portrays the impact of an agricultural shock as household wealth increases.The linear combination (i.e., â1 þ â3 ) shows the total effect of agricultural shocks for the wealthiest possible households (i.e., wealth index equal to one).
At the extensive margin, for the poorest households (with a wealth index of zero), an agricultural shock significantly reduces children's schooling by 7.4 pp (column 1).However, it insignificantly reduces child labor and general health by 2.7 pp (column 2) and 2.4 pp (column 3), respectively.At the intensive margin, an agriculture shock significantly reduces schooling by 23 minutes (0.37 hours) (column 4) and increases child labor by 45 minutes (0.75 hours) (column 5) for the poorest households.For the wealthiest households (with a wealth index of one), an agricultural shock significantly increases schooling by 12.6 pp (column 1) and significantly increases the switch to child labor by 11.6 pp (column 2).However, it has an insignificant impact on general health, reducing it by 7.2 pp (column 3).At the intensive margin, an agricultural shock has an insignificant impact on schooling for the wealthiest households, but it significantly decreases child labor by 58 minutes (0.96 hours) (column 5).
The analysis suggests that households in the sample utilize non-land-based wealth as a buffer against the adverse effects of agricultural shocks on child education and health.However, in terms of joining the labor force at the extensive margin, wealthier households are more inclined to have their children engage in agricultural work.These results are consistent with the idea that persistent agricultural shocks can distort land, labor, and credit markets.Drawing parallels with Ghana, where these markets are imperfect, Bhalotra and Heady (2003) hypothesize and demonstrate empirically that households with available farmland have an incentive to employ child labor, a phenomenon they refer to as the "wealth paradox".A similar analysis of agricultural shocks and household wealth is presented in Table 5, except in this case, the analysis focuses on children below 15 years of age to align with the minimum employment age stipulated in the 2019 Labour Proclamation. 10At the extensive margin, the outcomes include schooling, labor, farming and paid work 11 while at the intensive margin, the outcomes are schooling and labor.For the poorest households, the extensive margin analysis reveals that an agricultural shock significantly diminishes schooling by 4.9 pp (column 1) but does not significantly impact child labor, farming, or paid work.At the intensive margin, the effect of an agricultural shock for poorer households is a significant reduction in schooling by approximately 26 minutes (column 5), while its effect on child labor is not statistically significant.These findings highlight the constrained labor market effects of agricultural shocks, potentially due to a lack of employment opportunities in farming, despite the insignificance of this impact.
Conversely, for the wealthiest households, an agricultural shock significantly increases schooling, child labor, and farming by 9 pp (column 1), 16.4 pp (column 2) and 13.4 pp (column 3), respectively. 12However, when considering the total effect at the intensive margin, agricultural shocks have an insignificant impact on schooling and labor for wealthier households.Thus, wealth plays a critical role as a buffer against agricultural shocks.Overall, the extensive (switching) and intensive (time-use) analyses show the differential impact of wealth on child labor and calls for more detailed surveys encompassing elements of time-use (intensity of labor) for children in developing economies.

Robustness
We performed a number of robustness checks to validate our main results, which are presented in Table 2.We summarize the findings here with all results reported in the Supplementary Materials section (hereafter called Appendix).As different types of agricultural shocks could affect child poverty differently, we begin by examining how each reported shock in the survey (e.g., drought, flood, etc.) affects outcomes (Appendix Table A2).Results are noisier when examining each shock type in isolation, but the majority of shocks show similar patterns as our overall results.Specifically, almost all shock types show at least weakly negative effects on schooling and health outcomes and positive effects on labor supply.

Persistent agricultural shocks and child poverty 41
As previously mentioned, shocks are idiosyncratic, self-reported, and therefore subject to endogeneity.A complementary approach could therefore be to examine the impact of covariate shocks, which arguably suffer from less endogeneity bias.Specifically, we calculate the average reported shocks in the community (cluster level) and include this as an additional variable in our regressions (Appendix Table A3).There is some evidence that covariate shocks reduce schooling and increase labor at both the extensive and intensive margins.However, these results are less consistent across models than idiosyncratic shocks (e.g., TWFE results in panel (b)).Moreover, the estimated impacts of idiosyncratic shocks are qualitatively insensitive to the inclusion of the covariate shock measure (with the exception of some attenuation in the simple OLS model).Nonetheless, these results tentatively suggest that community-level agricultural shocks could have important implications for time use above and beyond idiosyncratic shocks.
We used self-reported health as our primary health outcome, which could also potentially suffer from endogeneity.To analyze if the impacts of agricultural shocks are significantly different between self-reported health and objective health, we create a child wasting measure based on WHO BMI for age z-scores, which are standardized against an international reference population sample.Specifically, the wasting variable takes on the value of 1 if a child's BMI z-score is less than -2, and 0 otherwise.Using the wasting indicator as the health outcome produces qualitatively similar results as using self-reported health (Appendix Table A4).Specifically, agricultural shocks (both at the extensive and intensive margin) significantly increase the probability of wasting across all econometric model specifications.
Finally, in a longitudinal survey, there is bound to be some level of attrition of participants either because of death, 13 relocation, lack of formal addresses, or a waning of interest from respondents.This loss of participants can lead to bias in the inferential results because of possible correlations with observable characteristics and sample selection (Outes-Leon & Dercon, 2009).However, due to maintaining the same field supervisors over all rounds, 14 the attrition rates were kept very low in Ethiopia, at 5.3% for the younger cohort and 17.7% for the older cohort between rounds one and five.The main reason for attrition was international migration (for 22-year-olds) and tracking challenges for some children (Barnett et al., 2013;Young Lives, 2018).Nonetheless, despite the low attrition rates, Appendix Table A5 provides summary statistics comparing the sample of children present in all survey rounds to those missing in any round.There are no large differences between groups for most observable characteristics, though the attrited sample did report somewhat fewer shocks, more time in school, and less time in labor.However, when including only children that were present in all survey rounds, results did not change significantly from our benchmark results (Appendix Table A6).This suggests that attrition bias is likely limited.

Discussion and conclusion
The persistence of agricultural shocks, despite being widespread, has received insufficient attention in the literature, particularly concerning how these shocks influence child poverty outcomes in developing economies.In the presence of persistent agricultural shocks, household members may adjust their consumption, savings, and investments as a coping mechanism.In such contexts, household wealth assumes a critical role as a buffer against the adverse effects on children's education, labor, and health.Through the utilization of two-way individual fixed effects and dynamic panel instrumental variable regressions, we investigated how agricultural shocks impact education, child labor, and general health in a longitudinal study involving individuals aged between 5 and 22 years old.This study contributes to the existing literature in several key ways.Firstly, we employ an exogenous proxy for recurring and persistent agricultural shocks by utilizing a GMM-IV estimation model.Secondly, we analyze household decisions within a dynamic setting as agricultural uncertainty becomes more predictable.Thirdly, we assess the role of non-land-based household wealth in shaping investments in child education, general health, and labor within this dynamic environment.Lastly, in a distinctive contribution, we explore the effects of persistent agricultural shocks on the intensity of child education and labor through a time-use study.
Using data from five survey waves encompassing two age cohorts of children in Ethiopia, we find that agricultural shocks lead to children dropping out of school and joining the labor force.These shocks significantly reduce the general health condition of children, even when we control for their body mass index.Moreover, agricultural shocks significantly reduce the time children spend in school while simultaneously increasing time spent in labor and leisure activities (playtime).Interestingly, there is a significant increase in the likelihood that children from wealthy households join the household tasks of farming and farm business when there is an agricultural shock, corroborating the concept of a "wealth paradox" (Bhalotra & Heady, 2003).However, this positive child labor effect is reversed at the intensive margin, highlighting the limitations of dichotomous labor measures for children.Overall, wealthy families are better equipped to mitigate the impact of agricultural shocks on the welfare of their children.Our findings point to the resilience of an economy and underscore how wealth makes a difference in cushioning the anticipated adverse effects of shocks.
This study is not without limitations.The YLS surveys featured an oversampling of poor areas (Barnett et al., 2013), which raises internal and external validity concerns.Nonetheless, despite the sampling design, the survey is representative of regions, policy contexts, living conditions, and rural and urban areas in Ethiopia.The Demographic and Health Survey of 2000 and the Welfare Monitoring Survey of 2000 also corroborated the representativeness of the YLS sample.Furthermore, the comparisons of several living standard indicators showed that the samples in YLS were similar to those nationally representative in Ethiopia (Barnett et al., 2013).
The implications of this study extend to future research in both academic and policy circles.Firstly, methodological approaches aimed at comprehending the impacts of agricultural shocks or similar persistent shocks must acknowledge that such shocks are often not exogenous and should thus be modeled accordingly.The persistent nature of agricultural shocks, including floods, droughts, pests, soil erosion, and frosts, underscores their endogenous nature and their potential to introduce bias into estimation results.Estimation strategies like the dynamic panel data model (GMM-IV method) employ lagged differences of dependent variables and lagged agricultural shocks as instruments to address possible endogeneity and serial correlations.Secondly, in the policy domain, it is imperative to recognize that persistent agricultural shocks foster adaptive behaviors, particularly when variations in wealth exist.Policymakers should develop informed expectations regarding potential child outcomes in contexts where agricultural shocks are likely to persist and consider the role of wealth, income, or wages in mitigating the impact of such shocks on child outcomes.Agricultural shocks can significantly impede lifelong prospects for children, particularly those from less affluent households.

Table 1 .
Descriptive statistics, Ethiopia, YLS, 2002-2016 Notes: General health is coded as 0 for very poor, poor and OK health, and 1 for good and very good health, and the data for the same is available in round 3, 4 and 5 only.Descriptive statistics presented are aggregated means for younger and older cohort above 4 years of age.The variable wasting is derived by first winsorizing the 'BMI for age z score' at 1 and 99 percentile and creating a dummy variable which takes the value of 1 if BMI for age z score is less than -2.

Table 2 .
Dynamic GMM-IV, fixed effects and OLS: impact of agricultural shocks on time use in school- Notes: The treatment variable is the agricultural shocks to the household.Control variables: age, BMI, household head's education, household size and wave intercept for the five rounds,

Table 3 .
(Andrews & Lu, 2001)g.shocks on children time use in schooling, labor and leisure at the intensive level.YLS, Ethiopia, 2002-2016Dynamic panel instrumental variable regressions with Generalized Methods of Moments in all specifications(Andrews & Lu, 2001).Control variables include wave intercept for the fiverounds, 2002,  2007, 2009, 2012, and 2016, age of the child, household head's education, household size, and BMI for children above 4 years of age.Unconditional regressions are irrespective of time spent in the said activity, while, conditional regressions are conditional on a positive time spent in the said activity.Robust standard errors are clustered at the household level.Significance ÃÃÃ p < 0.01, ÃÃ p < 0.05, Ã p < 0.1.

Table 4 .
Fixed effects: interactive impact of agricultural shock and household wealth on child's time use in schooling, labor and leisure.YLS, Ethiopia, 2002-2016Robust standard errors in parentheses, clustered at the household level.Significance ÃÃÃ p < 0.01, ÃÃ p < 0.05, Ã p < 0.1.â1þ â3 is the total effect of agricultural shocks for households that have a very high wealth index.Control variables are dummies for the five rounds of the survey, individual's age, household size, BMI, and household head's education.Sample consists of individuals above 4 years of age.40 R.Miller et al.

Table 5 .
Fixed effects: interactive impact of agricultural shock and household wealth on time use in schooling and labor for age less than 15, YLS, Ethiopia, 2002-2016 Robust standard errors in parentheses, clustered at the household level.Significance ÃÃÃ p < 0.01, ÃÃ p < 0.05, Ã p < 0.1.Here, labor is the sum of farming and paid work activity for a child below 15 years.â1 þ â3 is the total effect of agricultural shocks for households that have a very high wealth index.Control variables are dummies for the five rounds of the survey, individual's age, household size, BMI, and household head's education.The sample consists of individuals above 4 years of age.