The Link between Vulnerability to Poverty and Depression: Evidence from Vietnam

Abstract Vulnerability to poverty is a measure of the downside risk of falling into poverty. We examine the relationship between vulnerability to poverty and self-reported mental health, using a four-wave longitudinal panel from rural Vietnam. Our findings indicate that vulnerability to poverty has a significant and adverse connection with an index of depression. The impacts are not only statistically significant but also large. An increase in vulnerability to poverty from zero to one is associated with a 3.3 unit increase in the depression (CES-D) score (or 47.1 percent increase over the sample mean). Moreover, vulnerability to poverty also increases the probability of being severely mentally distressed. An increase in vulnerability to poverty of one standard deviation is associated on average with a 10.5 percentage point increase in the probability of severe mental distress. Risks of poverty that come from idiosyncratic shocks have stronger links to mental health than risks from covariate shocks. Vulnerability to poverty increases the likelihood of depressive symptoms more for men than for women and for the major ethnicity group compared to other ethnicities. Overall, we provide clear evidence that lives marked by greater downside risk are also blemished with higher rates of depressive symptoms.


Introduction
Mental health is an essential part of human general health and well-being and the causal impact of poverty on mental well-being is long established (Das, Do, Friedman, McKenzie, & Scott, 2007;Ridley, Rao, Schilbach, & Patel, 2020).More widely, income has been linked to happiness and measures of life-satisfaction (Clark, Frijters, & Shields, 2008;Cuong, 2021).In developing countries, it is particularly the poor and the vulnerable in society who are at greater risk of mental ill-health and who are also the least likely to receive adequate services.In such an environment, where households are regularly exposed to shocks to livelihoods or to physical health, or where natural disasters are prominent it might be expected that the risk of poverty -or vulnerability to poverty -and not just poverty itself can affect mental health.The possibility of a slide into poverty might, for example, create anxiety and depression in some individuals.And there is certainly evidence that shocks can affect subjective well-being (Charles, Wu, & Wu, 2019).Yet, while there is also data linking poverty risk to life satisfaction (Caria & Falco, 2018;Dang, Abanokova, & Lokshin, 2020), what is lacking is an empirical investigation of the link between the risk of poverty and mental health.
In Vietnam, one of the greatest threats to people in rural area is natural disasters and weather-related extreme events that may become more unpredictable due to climate change and which already have negative effects on household income and expenditure (Arouri, Nguyen, & Youssef, 2015;Imai et al., 2011).The country is ranked one of the five countries in the world most vulnerable to natural hazards (The World Bank Group & Asian Development Bank, 2020).Among those hazards, storms, floods, and droughts take place more frequently and they are typical threats for agriculture -the major source of earnings for nearly two thirds of the rural population.
Our main goal is to investigate the relationship between vulnerability to poverty and mental health.To achieve these objectives, we use the four-wave longitudinal the Vietnam Access to Resources Household Survey (VARHS) data from 2012 to 2018.The survey includes a set of questions that is a comprehensive measure of mental health, namely the well-established Center for Epidemiologic Studies Depression (CES-D) Scale.To identify households who are at risk of falling into poverty, we employ the extended Vulnerability as Expected Poverty (VEP) approach from Mina and Imai (2017).This framework is notable for being based on a threelevel model of shocks, with the three levels being time of measurement, household, and commune.Using this we can link measures of vulnerability to self-reported mental health The main finding of our study is that vulnerability to poverty has a significant and adverse connection with mental health.An increase in vulnerability to poverty from zero to one is positively associated with a 3.3 unit increase in the CES-D score (or 47.1 percent increase over the sample mean).Moreover, vulnerability to poverty also increases the probability of being severely distressed.The impact is not only statistically significant but is also large.A one standard deviation increase in vulnerability to poverty is associated on average with a 10.5 percentage point increase in the probability of severe mental distress.
Along the way, we learn much about the nature of shocks that drive vulnerability to poverty in Vietnam.Defining households as vulnerable if they face a probability of at least 50% of falling below the poverty line in the next two years, our results show that 31.4 percent of households are classified as vulnerable.Shocks are closely and significantly related to households' poverty situation and drive this vulnerability.We can be more specific: idiosyncratic shocks refer to shocks that affect only members of a household such as illness or death of a household members while covariate shocks refer to shocks that affect entire villages or communities such as epidemics or flooding.Around 28.5 percent are vulnerable to unobservable idiosyncratic shocks while around 16.4 percent of panel households are classified as vulnerable to unobservable covariate shocks.Households, who are exposed to natural disasters are at higher risk of falling into poverty by 4.6 percentage point while living in a commune that faces natural disasters can increase the risk of being poor by 2 percentage points.Putting the evidence on vulnerability and depression together, we find that direct experience of natural disasters is positively associated with a 10 unit increase in the CES-D score while living in a commune affected by natural disasters are positively associated with a 6.6 unit increase in the CES-D score.
Overall, we find that the risks of poverty that arise from idiosyncratic shocks have a greater link with mental health than those created by covariate shocks.Meanwhile we do not find a significant relationship between poverty risk from covariate shocks and mental health.We speculate about the reasons for this distinction in the conclusion but note that it might arise through the shared experience of covariant shocks compared to the isolating experience of idiosyncratic shocks.
In setting out this evidence, we make several contributions to the existing sparse literature on the relationship between vulnerability to poverty and mental health.First, to the best of our knowledge, this is the first paper that focuses on vulnerability to poverty and measures of mental distress (rather than general life satisfaction).Our data contains a comprehensive set of questions on mental states that enables us to construct a standard index of depression that captures many aspects of mental health, including an severe distress.Second, we estimate vulnerability by applying a three-level model, which allows us to construct not only total vulnerability, but also the two components: vulnerability to unobservable idiosyncratic shocks and vulnerability to unobservable covariate shocks.Our results are informative for understanding and learning the source of poverty risk as well as its relationship with mental state, which are important for policymakers to consider then designing social welfare policies.Third, our study is implemented in the context of rural areas; specifically, the subjects of the study are mostly farmers and living in area affected by disasters while most previous studies on life satisfaction and poverty risk focus on workers in urban areas or in middle-income country (Caria & Falco, 2018;Dang et al., 2020).
Findings from our study may provide useful information for policy makers as well as development agencies.Social policies, that support household to be more resilient with shocks, can help reduce severe distress considerably.In addition, our results may also provide important implications for other emerging and transition economies similar to Vietnam.
The paper proceeds as follows.Section 2 provides a brief review of the related literature on estimation of vulnerability and the impact of vulnerability on mental health.In Section 3, we describe the household survey data used in more detail focusing on the measurement of mental health.Our empirical methodology is explained in Section 4. Section 5 presents our main estimation results.Section 6 contains heterogeneity analysis and robustness checks, while section 7 concludes the paper.

Background and related literature
Vulnerability to poverty is defined as the risk of households or individuals falling below the poverty line in the future (World Bank, 2001).This definition implies a forward-looking (ex-ante) and comprehensive approach: evaluating the current situation and characteristics of households and where they live, and the shocks they experience; and then predicting their probability of falling into poverty.In empirical analysis, ex-ante vulnerability measures are often specified as a function of the expected mean and variance of household welfare (Dercon & Krishnan, 2000;Gallardo, 2018).The mean of the expected welfare (in terms of either consumption or income) is determined by household and community characteristics, whereas the variance in household welfare is determined by (a) the severity and frequency of idiosyncratic and covariate shocks; and (b) the strength of a household's coping mechanisms for insuring welfare against these shocks (G€ unther & Harttgen, 2009).
The literature on ex-ante quantitative measures of vulnerability to poverty has been growing rapidly, but broadly speaking, measures are classified into three main approaches (Ceriani, 2018;Hoddinott & Quisumbing, 2003).First, vulnerability as low expected utility (VEU), that is expressed in terms of utility gaps (Ligon & Schechter, 2003;Calvo & Dercon, 2013;G€ unther & Maier, 2014;and Chen, Rong, & Song, 2021).The second approach to vulnerability measurement is vulnerability by mean risk (VMR), which is based on the mean-risk ordering of uncertain welfare outcomes (Chiwaula, Witt, & Waibel, 2011;Gallardo, 2013); or sometimes vulnerability is defined in terms of a low-mean outcome and the risk of divergence from that mean (Gallardo, 2013); or vulnerability threshold (Dang et al., 2020).The third approach is Vulnerability to Expected Poverty (VEP) which identifies expected poverty as the essential feature of vulnerability to poverty (Pritchett, Suryahadi, and Sumarto (2000), Chaudhuri, Jalan, and Suryahadi (2002)).This is the method we use.The first method is not suitable in this study, because one of components in mental healthlife satisfactionis widely considered as a proxy for utility (Benjamin, Heffetz, Kimball, & Rees-Jones, 2012;Kimball & Willis, 2006).The VMR method is potentially suited, but it usually requires a long panel and very rich information about the nature of the shocks.Given our four-wave panel data with limited information about all types of shocks to households, this approach is not reliable.

Vulnerability to poverty and depression 1851
We apply the extended VEP, with multi-level modelling, proposed by G€ unther and Harttgen (2009) with two-levels, and further developed by Mina and Imai (2017) with a three-level random coefficient (RC) model.Unlike the fixed effect (FE) model (Kamanou & Morduch, 2002;Klasen, Lechtenfeld, & Povel, 2015) the multilevel models recognizes by design the existence of data hierarchies by allowing for residual components at each level in the hierarchy. 1And so information at both household and commune levels can be included simultaneously in the same model without violating the assumption of independent observations, thus providing correct standard errors and significance tests (Goldstein, 1999).In other words, unlike the FE model, the multilevel model allows the shocks that affect households to occur at both household level and commune level.Furthermore, the multi-level model helps us to know the source of vulnerability by decomposing the unexplained variance of household income into a lower level (household) and a higher level (commune).
A limitation of the three-level random coefficient model is potential correlation between covariates and an unobservable term at the household or commune level.To address the limitation, we use Mina and Imai (2017)'s technique of applying the 'within-between' formulation proposed by Bell and Jones (2015).'Within variation' refers to a vector of demeaned terms of time-varying covariates in level 1 and level 3 (time-varying covariates minus time-series mean of time-varying covariates).And 'between variation' refers to the use of a vector of time-series mean of time-varying covariates.
When we turn to the link between poverty and measures of life satisfaction, the literature is also extensive, with many studies examining the impacts of poverty on happiness, particularly life satisfaction (Das et al., 2007, Markussen, Fibaek, Tarp, & Tuan, 2018) or the effects of (income) shocks on mental health (Ridley et al., 2020).Nevertheless, research on the link from vulnerability to well-being is extremely limited.Two recent papers (Caria and Falco (2018) for Ghana and Dang et al. (2020) for Russia) stand outfor finding a negative link between vulnerability to poverty and a broad measure of life satisfaction.Neither study explicitly looks at mental health.
As prior studies focus on broad measures of life satisfaction, they leave open the question of whether vulnerability to poverty is explicitly linked to mental health, and to answer that question requires some clearer measure of mental health.The panel that we use includes in later waves a well-established depression index that captures many aspects of mental health, including indicators of severe distress.This is a module of the Center for Epidemiologic Studies Depression (CES-D) Scale (Radloff (1977)).The CES-D scale is based on a comprehensive set of questions that capture all aspects of mental condition of a person, including their emotional, psychological, and social well-being.Our panel features a 10-item short-form version of CES-D, developed by Andresen, Malmgren, Carter, and Patrick (1994), which has also been shown to have good psychometric properties in a variety of contexts (Boey, 1999;Bj€ orgvinsson, Kertz, Bigda-Peyton, McCoy, & Aderka, 2013;Blattman, Jamison, Koroknay-Palicz, Rodrigues, & Sheridan, 2016;Kilburn et al., 2018;Eyal & Burns, 2019).The validity and applicability of the CES-D have been demonstrated in low-and middle-income countries in Asia, Africa, the Caribbean and South America (Thanh, Quyen, & Tien, 2016;James et al., 2020).
In short by applying Mina and Imai (2017)'s three-level model, 2 we construct total vulnerability, vulnerability to idiosyncratic shocks and vulnerability to covariate shocks.From there, we use the CES-D measure to understand better the relationship between vulnerability and mental health.

Data description
Our data comes from the Vietnam Access to Resources Household Survey (VARHS) for four waves , 2012, 2014, 2016 and 2018.The datasets are collected every two years in rural areas of 12 provinces throughout Vietnam. 3 The specific data we use to estimate vulnerability is at the household level with 2,800 4 households from 466 communes in the VARHS in the period 2012-2018.The scale of attrition is small over the four waves, amounting to 3.53 per cent in total from 2012 to 2018.
In the VARHS survey, a module of a 10-item CES-D has been included since 2016.Respondents are asked to indicate how often they had certain feelings in the past week in terms of various emotional and physical condition, on a 0-3 scale of 'never (0 days in a week)'; 'sometimes (1-2 days of the week)'; 'often (3-4 days of the week)'; and 'all the time (5-7 days of the week).'Details of the questions are shown in the Appendix, Table A1.Answers to questions about negative feelings are scored progressively while answers for questions on positive feeling are scored on a decreasing scale.The resulting CES-D score is used to create an additive index.The possible range of scores is zero to 30, a higher indicating the presence of more symptoms of depression.A score of 10 and above is used as an indicator for the presence of significant depressive symptoms (Andresen et al., 1994).In 2018, there was no information on mental health for about 400 individuals (15% of the original sample).After those 400 individuals are dropped, we work with a balanced panel which tracks 2,312 individuals.

Descriptive statistics
We present key descriptive statistics from the main variables from the balanced sample for the years 2016 and 2018 (Table 1).Sixty-five percent of the sample come from minority ethnicity groups.83 percent of the sample is male.The average age is 53 years old.85 percent are married.Pest infection and crop disasters and natural disasters (all self-reported) are the main shocks that households are exposed to and occur frequently.The proportion of sample exposed to disasters is higher in 2016 (29%) than in 2018 (19%).
These statistics also show that the average score of the CES-D10 is 7.04, with a substantial gap between male and female.The mean score is 8.6 for women and 6.7 for men.Meanwhile, 40% of females show severe depressive symptoms, compared to 23% for males.Figure 1  Vulnerability to poverty and depression 1853 provides more information on the distribution of the CES-D index for our sample and also points to the significant proportion showing severe depressive symptoms.

Estimates of vulnerability to poverty
To estimate vulnerability, we proceed by first running a regression of income on household and commune characteristics using the three levels of household, commune, and time.The variables that we use for the regression include standard household demographic variables as well as commune-level measures of exposure to particular shocks. 5Particularly, household characteristics include profile of household head (gender, ethnicity, education, age), household size, dependency ration, proportion of household members finished secondary school/vocational school/as wage workers/work in agriculture, commune characteristics (roads, utilities, having irrigation system, crop land (ha)), shocks occurred at commune and household level (natural disasters, pest infection, crop diseases, serious illness of household member).Other variables include time, province dummies, and interaction terms such as those between household and commune characteristics, province dummies and shocks, time and shocks/commune and household characteristics.From this, we get predicted values for income as well as residuals at the household and commune level.Second, we regress the squared residuals to estimate the following variances: unobservable idiosyncratic variances, covariate variance and total variance of unobservable shocks.Third, using these three variances in turn, we compute three estimates of vulnerability (using total, idiosyncratic and covariate variances separately), as follows.
Vulnerability to poverty (V) of a household (i) in a commune (j) at time (t) ¼ Probability (income < official poverty line j characteristics of households and communes) ¼ Function (official poverty line (y), expected mean of log of income (c y ijt Þ and expected variance of log of income, c d 2 ijt ) This produces three continuous variables: total vulnerability, vulnerability to unobservable covariate shocks and vulnerability to unobservable idiosyncratic shocks.In this method, vulnerability is a scale from 0 to 1, but by convention, vulnerable households are usually referred to as the ones with estimated vulnerability greater than or equal to 0.5 over the relevant time period (Chaudhuri et al. (2002).

Identifying vulnerable households and the decomposition of vulnerability
Overall, 31.4 percent of households are classified as vulnerable at least once in the next two years (i.e.2012-2014).Around 28.5 percent are vulnerable to unobservable idiosyncratic shocks (or shocks at household level) while around 16.4 percent of panel households are classified as vulnerable to unobservable covariate (or shocks at commune level). 6The stronger role of idiosyncratic shocks is similar to that found by Pham, Mukhopadhaya, and Vu (2021) who applied the extended VEP approach in Vietnam for an earlier period and also to Mina and Imai (2017) and Gaiha and Imai (2008) for the Philippines and India respectively.Our results provide further support for the claim that idiosyncratic shocks might not be insured perfectly by villagelevel mutual help or in general by social network as informal mutual assistance.Natural disasters (such as floods, droughts, and storms) reported either at commune level or household level have a negative connection with household vulnerability.Natural disasters, reported at commune level, seem to have smaller impact than natural disasters reported at household level.According to Table 2, households, that are exposed to natural disasters, are at a higher risk of falling into poverty by 3.1 percentage points, while living in a commune with natural disasters can increase the risk of being poor by nearly 2 percentage points.This finding is similar to Arouri et al. (2015) who find that living in a commune with floods increases the probability of being poor by 1.8 percentage point.

Mental health: empirical methodology
We then estimate a model of the following form to examine the effect of poverty risk on mental health with individual fixed effects: where M i,t is the CES-D score or indicator for severe depressive symptoms of individual i in the year t.
V i,t is estimated vulnerability to poverty that indicates the probability of household income falling below the income poverty threshold once in the next two years.
X i,t is a vector of individual and household characteristics.g i and E i,t are respectively individual fixed effects and unobserved variables; The coefficients of interest are a 1 .Our main hypothesis is that a 1 is positive.In other words, increasing vulnerability to poverty is associated with increase in the CES-D score.

Vulnerability to poverty and depression 1855
To interpret the results as the causal relationship between vulnerability to poverty and individual mental health, we acknowledge two identification challenges.The first challenge is omitted variable bias, particularly the possibility of unobserved characteristics of individuals.The use of individual fixed-effect regression eliminates the effect of those time-invariant individual characteristics, but at least in theory there may be time varying individual effects that are not captured in the model.A second challenge is reverse causality.Mental distress may affect the motivation to work.Happy individuals may be more creative and find ways to earn more income and hence lessen their vulnerable situation (Krekel, Ward, & Neve, 2019).
One way to deal with these challenges is to adopt an instrument variable (IV) approach.However, even if the instrument approach performs well, the causality from vulnerability to happiness will not be identified because of the nature of the vulnerability model.Recall that vulnerability is estimated from the prediction based on household and community characteristics.A key assumption to justify the instrument approach is that the explanatory variables which are used for the vulnerability model do not affect the subjective well-being measures.Yet, since we include household shocks in the vulnerability model, a claim that the exclusion restriction hold is unconvincing.We therefore rely on the individual fixed effects for identification. 7

Results
Table 3 summarizes the main results of this paper.It shows our estimates of the impact of vulnerability to poverty on mental health outcomes.The dependent variable in the first and third column is the CES-D score.A higher score reflects poorer mental health.In the second and fourth column, the dependent variable is an indicator for severe mental stress that takes the value 1 if the score on the CES-D index is greater than or equal to 10.Since this variable is dichotomous we use a fixed effects logit model.For logit models, the meaning of coefficients can be difficult to interpret and thus it is common to present the marginal effects.For fixed effect models, the marginal effects depend on the unidentified fixed effects.For that reason, following Kitazawa (2012) and Kemp and Santos-Silva (2016), for the equations involving dichotomous variables we report the average semi-elasticities, which can be consistently estimated.
As shown in Table 3, being vulnerable to poverty has a significant and adverse connection with mental health, both before and after including controls.The results in column 3 show that an increase in vulnerability to poverty from zero to one is positively associated with a 3.3 unit increase in the CES-D score (or 47.1 percent increase over the sample mean).Moreover, the result in column 4 of the table indicates that vulnerability to poverty also increases the probability of being severely distressed.The impact is not only statistically significant but also large in terms of the underlying measure.An increase in vulnerability to poverty of one standard deviation is associated on average with a 10.51 percentage point increase in the probability of severe mental distress. 8.And, as can be seen in the table, the impact of vulnerability to poverty on mental health is comparable to that of major life shocks, such as divorce, unemployment, or major illness.For people on the poverty line in 2018, one more natural disaster per year (the mean is 0.19) has an effect on the depression score equal to a two per cent drop in income.Figure 2 shows the role of vulnerability.It shows predicted depression by income decile for the full model ('with vulnerability') and with the vulnerability variable set to zero ('without vulnerability').The impact is largest at low-income levels where vulnerability is highest.For the lowest decile, removing vulnerability lowers the CES-D score by 26%.For the top decile the impact is only 0.5%.

Vulnerability to poverty and depression 1857
The relationship of income with poor mental health (i.e. a higher depression score) is negatively strong and statistically significant across all regressions.While this general finding is consistent with other studies (e.g.Clark et al., 2008;Cuong, 2021), making a specific comparison is more difficult given the absence of similar research.Some comparative evidence is available from evaluations of cash transfer programs in other low-income countries.Studies of a one-off intervention in Kenya conducted by Haushofer and Shapiro (2016) find that cash transfers amounting to approximately 24.7% of mean annual per capita GDP, reduce depression by 0.16 SD (standard deviations) after almost three years.In our case, a rise in income of 24.7% of the mean income in 2018, for example, would imply a total reduction in the depression score of 0.13 SD for someone on the poverty line, which is smaller than the Kenyan figure, but in keeping with results from other interventions, according to Ridley et al. (2020) in their review.Some 84% of the change is attributable to the direct effect of income and 16% to the impact via vulnerability (see Supplementary materials for calculation details).
For other specific shocks, there is also both a direct and indirect effect.Recall from, Table 2, shocks are closely and significantly related to household poverty.Meanwhile, in Table 3, we find that an increase in vulnerability to poverty from zero to one is positively associated with a 3.3 unit increase in the CES-D score.Putting the findings from Tables 2 and 3 together, we estimate that the direct experience of natural disasters is positively associated with a 10 unit (¼3.1 Â 3.3) increase in the CES-D while living in a commune affected by natural disasters are positively associated with a 6.6 unit (¼2 Â 3.3) increase in the CES-D score.Similarly, in terms of severe distress, direct experience and living in a commune affected by natural disasters are linked with a 6.4 percentage point and about 4 percentage points increase in severe distress, respectively.Given a sample mean of 33 percent, this translates into 19 percent and 12 percent increase.This is a large impact.
In Table 4, we break down the relationship of vulnerability with mental health using its two components: idiosyncratic vulnerability (Panel A) and covariate vulnerability (Panel B).
From the final column, an one SD increase in poverty risk from idiosyncratic shocks is significantly associated with a 0.09 probability increase in severe depressive symptom.Meanwhile there is no significant relationship between poverty risk from covariate shocks and mental health.This is possibly because the impacts of idiosyncratic shocks are more direct and more specific.Overall, the results suggest that being vulnerable to poverty significantly increases the  3. risk of depressive symptoms, but the impact depends at least partially on the source, with the role of idiosyncratic shocks being stronger. 9

Heterogeneity analysis and robustness checks
Panel A of Table 5 presents the relationship between risk of poverty and mental health estimated separately by gender.Women score more highly than men on the depression score, but the male depression score is more sensitive to changes in vulnerability to poverty compared to the women.In fact, for the women in our sample, there is no significant link between vulnerability and mental health measures.For men, on average, an increase in the vulnerability to poverty from zero to one is associated with 3.9 points increase in the CES-D score.A one SD rise in vulnerability is associated with a 10.3 percentage point rise in in the likelihood of severe mental distress.A possible explanation for the gender difference is that the man is usually the head and breadwinner in the family.Especially in the context of countries like Vietnam, men typically take more responsibility in providing sufficient food and money for the family and therefore might be more psychologically stressed by the threat of poverty.Another possible explanation is that women tend to be more resilient to shocks than men (Andersen, Verner, & Wiebelt, 2017).
Panel B of Table 5 provides results separately for Kinh (the ethnic majority in Vietnam) and non-Kinh (non-ethnically Vietnamese).Sub-sample sizes do not allow a more detailed breakdown of the model by ethnicity.In general, there is no large difference in mean depression score.Vulnerability to poverty has a significant relation with mental health for both groups but it seems that the connection is stronger for the Kinh than non-Kinh.Column 1 shows that on average, an increase in the vulnerability to poverty from zero to one is related to 3.8 units increase in depressive symptoms among non-Kinh group.Turning to column 2 we find that the magnitude is about two times higher for the Kinh group with a 6.9 unit increase in depressive Vulnerability to poverty and depression 1859 symptoms.Columns 3 and 4 show the corresponding analysis for severe mental distress.A one SD higher value for vulnerability is linked to severe mental distress higher by 9.3 percentage points among Non-Kinh and 20.34 percentage points among Kinh.
We examine the robustness of our findings in various ways.Detailed results can be found in the Supplementary Materials.We examine separately the relationship between being vulnerable and each individual component analysis of the CES-D index.Results show that vulnerability to poverty is significantly related to seven of the ten components of the CES-D score. 10 Second, we check if the results are sensitive to the definition of the poverty lines.We estimate the vulnerability index with poverty lines that range from about 1.2 to about two times the official poverty line and find no major changes in the relationship with the CES-D score.
We also estimate vulnerability using a fixed effect model, and then examine its relationship with mental health.The results still show a significant and positive relationship between vulnerability and the CES-D index.In fact, the magnitude of the coefficients is higher than using the multilevel model.
Lastly, we test if the findings change across the income distribution.Generally, we find our results stay significant across different income groups.

Concluding remarks
We examine the relationship between vulnerability to poverty and mental health using a panel data for rural Vietnam.Our results show first that 31.4 percent of households are classified as vulnerable at least once over a window of two years.In turn, households are more vulnerable to poverty from idiosyncratic shocks than from covariate shocks.These results once again lend support to the observation that idiosyncratic shocks might not be insured perfectly by villagelevel mutual help as informal mutual assistance.Factors associated with higher vulnerability include having a less educated household head, non-Vietnamese ethnicity, lack of access to infrastructure, and living in disasters-affected area.Our main findings indicate a significant and strong relationship between vulnerability to poverty and mental health.This association is more strongly related to idiosyncratic shocks than to covariate shocks, stronger for male than female, and stronger for the Kinh group than minority non-Kinh group.These conclusions both mean levels of mental health and to severe mental distress.Above all, the findings support a conclusion that the mental effects of poverty are not confined to households who are currently poor.Members of families at a high risk of poverty in the near future are also more likely to show symptoms of depression.Thus, the targets for policies aimed at well-being amongst low-income groups could profitably be broadened to include non-poor households with a significant probability of falling into poverty.
More specifically, the links between education, natural disaster frequency, infrastructure and vulnerability suggest that standard mechanisms for raising economic development can have important mental health benefits, particularly when they are targeted at reducing vulnerability to poverty.This gives a further argument in favor of 'pro-poor' development.Eliminating shocks (such as natural disasters) at commune-and household-level would, on the basis of our data, reduce the depressive symptom by 10 units for a household that has a direct experience of natural disasters and about 6.6 units for household living in a commune affected by natural disasters.
Second, the familiar finding that idiosyncratic shocks might not be insured perfectly by village-level informal mutual assistance suggests policies to encourage the engagement of the community in addition to the governments and markets.If feasible, greater mutual assistance towards at-risk groups would not just lower the probability of poverty, but also improve mental health.Third, the findings that non-Vietnamese ethnicity and households who lives in the natural disaster-affected area are more vulnerable to poverty than other households, can be used to develop targeted interventions aimed at protecting these households.For instance, ensuring appropriate and supportive systems of social protection and adequate safety nets as a cushion against the risk of falling into poverty could be tried as a priority for farmers in the East Northern Mountains and Central Coast of Vietnam.
It remains to be seen whether the strong links between vulnerability to poverty and depression found in Vietnam are also a feature of other societies.Notes 1.For example, a two-level model which allows for grouping of households within communes would include residuals at the household and commune level.Therefore, the residual variance is partitioned into a betweencommune component (the variance of the commune-level residuals) and a within-commune component (the variance of the household-level residuals).The commune effects, represent unobserved commune characteristics that affect household outcomes.It is these unobserved variables which lead to correlation between outcomes for households from the same commune.As a comparison, we do also estimate the fixed effect model for VEP. 2.More details of the method are presented in the Appendix.3. See Tarp (2017).Details and a map of the survey area are presented in the Supplementary material.4. Note that means we use the full sample for estimating the vulnerability measures but the balanced panel for the mental health model in Section 5. 5.More details of the method are presented in the Appendix.Summary statistics of the variables used to construct the vulnerability to poverty and regression to estimate vulnerability are presented in the Supplementary material.6.Some further detail on the results can be found in the Supplementary material.7.In Vu (2022), we do explore an IV model as a robustness check.It does not materially alter the core results, and since the instruments are weak, we do not report the results here.8. The coefficient in the table is the average semi-elasticity (i.e. the derivative of log probability with respect to the relevant variable).Hence, multiplying by the probability, p, (which is also the derivative of p with respect to the log probability) gives the approximation.9.It is possible for example that the shared experience of covariate risks reduces their impact on mental health compared to shocks that are faced alone.10.Vu (2022) also includes an analysis of the link between vulnerability to poverty and a standard measure of lifesatisfaction.As that analysis shows a similar negative relationship (higher vulnerability is associated with lower life satisfaction) it suggests that our results are not driven by any idiosyncrasies of the CES-D scale.
Vulnerability to poverty and depression 1861

Figure 1 .
Figure 1.Distribution of the CES-D Depression scores.Notes: The deep green line at a CES-D score of 10 represents the threshold for severe depressive symptoms.

Figure 2 .
Figure 2. Predicted mean depression by income decile, using model in Table3.

Table 1 .
below Summary statistics Notes:The range of possible scores on CES-D index is 0-30.Severe depressive symptoms is an indicator for CES-D score !10.Kinh (or Vietnamese) is an indicator variable for the ethnic majority group.Natural disasters shocks include storm, drought, and flood.Economic shocks include price change, investment failure and land loss.
Table 2 highlight the role played by ethnicity, gender and commune level factors such as being located in natural disasters-prone areas (e.g. in the East Northern Mountains or the Central Coast) and lack of access to roads.

Table 2 .
Determinants of vulnerability, 2012Determinants of vulnerability,  -2018 Notes: Robust standard errors are in parentheses; ÃÃÃ p < 0.01; ÃÃ p < 0.05; Ã p < 0.1.Vulnerability is the dependent variable.The set of controls are the same set of household and commune characteristics used to construct the VEP.Omitted controls in the table include: square of age, household size, squared of household size, education (high school, short-term vocational school, professional school and university), proportion of household members finished secondary school, vocational school, proportion of household members as wage workers/working in agriculture; utility index; commune with road index; pest infection, crop disease or avian flu reported at commune level; pest infection, crop disease or avian flu reported at household level.The complete table of results is presented in the Supplementary material.

Table 3 .
Vulnerability and mental health: fixed effect results Notes: A higher CES-D score reflects poorer mental health.Severe Depressive Symptoms is an indicator for CES-D score !10.For severe depressive symptoms, we use a fixed effect logit model and report the average semi-elasticities.The number of observations for the severe depressive symptom regressions are reduced because the outcome variable is a dummy and only groups with variations are kept.Omitted controls are age group dummies, economic shocks and pest infestation and crop diseases.Standard errors clustered at the commune level are reported in parentheses.ÃÃÃ Significant at 1%; ÃÃ significant at 5%; Ã significant at 10%.

Table 4 .
Vulnerability and mental health by source of vulnerability Severe depressive symptoms is an indicator for CESD index !10.For severe depressive symptoms, we use the fixed effect logit model and report the average semi-elasticities.Numbers of observations for severe depressive symptom regressions are reduced because the outcome variable is a dummy and only groups with variations are kept.Controls are age group dummies, marital status, days unable to work due to illness (log), unemployment, natural disasters shocks, economic shocks and pest infestation and crop diseases.Standard errors clustered at the commune level are reported in parentheses.ÃÃÃ Significant at 1%; ÃÃ significant at 5%; Ã significant at 10%.

Table 5 .
Heterogeneity findings Severe depressive symptoms is an indicator for CESD index !10.For severe depressive symptoms, we use the fixed effect logit model and report the average semi-elasticities.Numbers of observations for severe depressive symptom regressions are reduced because the outcome variable is a dummy and only groups with variations are kept.Controls are the same as for Table4.Standard errors clustered at the commune level are reported in parentheses.ÃÃÃ Significant at 1%; ÃÃ significant at 5%; Ã significant at 10%.