THE VULNERABILITY OF HOUSEHOLDS TO POVERTY IN PERU, 2004-2014

Javier Herrera Zúñiga1 Angelo Cozzubo Chaparro2

Abstract In the last decade the Peru has gone through a period of rapid economic growth, accompanied by a dramatic reduction of the poverty by 37 percentage points. However, the observed economic slowdown in the past four years have led to wonder if that growth has allowed to consolidate a middle class finally freed from the risk of falling back into poverty or if, on the contrary, those households who left poverty would return to it in an even less favorable macroeconomic context. Our study aims to fill a gap in the empirical literature concerning the household’s vulnerability to poverty in Peru using a dynamic approach. The concept of vulnerability is defined as the risk or ex ante probability for a household to move from a situation of not poor in the initial period to poor in the next period. Following the methodological proposal of Dang & Lanjouw (2014), we estimate a vulnerability index, analyze its determinants and build vulnerability lines which, in complement with the poverty lines, allow us to break down the population into three groups and analyze its evolution in time: poor, vulnerable and non-vulnerable households. The ENAHO household panel data was used considering jointly all 9 biannual panels over the 2004-2014 period, which allowed us to rely on more than 50,000 observations on poverty transitions. Our results are robust to various specifications and cut points, and show that, the vulnerable population have been increasing in this period of rapid growth and poverty reduction reaching levels greater than 30% of total households in 2014. Contrarily to other studies which stress the role of adverse shocks, we find that the households’ vulnerability is mainly due to structural factors, linked to household’s labor characteristics but also to specific exogenous shocks, as well as the macroeconomic growth fluctuations. Individual dissaving strategies and having health insurance contribute to reduce vulnerability to poverty. JEL classification: I32, D31 Keywords: vulnerability, poverty dynamics, adverse shocks, panel data, Peru

1 Institut de Recherche pour le Développement, Pontificia Universidad Católica del Perú (PUCP), Departamento de Economía. [email protected] 2 Pontificia Universidad Católica del Perú (PUCP), Departamento de Economía. Instituto de Estudios Peruanos (IEP). [email protected] 1

THE VULNERABILITY OF HOUSEHOLDS TO POVERTY IN PERU, 2004-2014

1. The Importance of Considering the Vulnerability of Households Poverty rates in Peru has declined significantly during the last decade of strong macroeconomic growth, resulting in more than half of households ceasing to be poor. However, with the slowdown in the growth rates, from 9.1% in the peak of the cycle in 2008 to 2.3% in 2014, the decline in poverty has been much lower and the outlook for 2015 is not encouraging. The deterioration of the general economic context, particularly in activities geared towards the internal market, could imply an increase in the vulnerable population besides a larger number of poor households that were previously out of poverty but belonging to the group of vulnerable households. Several reasons justify the interest in studying the vulnerability to poverty. In the first place, we can consider it as an intrinsic dimension of well-being, as manifested by the very poor households that were interviewed in the study of the Voices of the Poor (Narayan 2000). Uncertainty about having the minimum resources to eat, heal, dress, etc. is considered by the population as an essential component of their poverty condition. Vulnerability induces many households to behave in ways that can reduce risks and mitigate their consequences. These strategies often have an immediate cost in terms of missed opportunities as well as long-term costs, trapping households in a situation of persistent poverty that could be transmitted between generations. In second place, it is fundamental to distinguish the poor and the households that, while being out of poverty, are vulnerable to falling since it makes it possible to observe a population group that escapes the targeting from a perspective of static monetary poverty. It is so that vulnerable population may even present different characteristics than the poor and the non-vulnerable population, making it necessary to implement specific policies. In third place, paying attention to the vulnerability of households highlights the difference between ex ante poverty interventions and ex post poverty alleviation policies. In this sense, it will not be enough to attend households that already suffer a poverty condition, since the design of preventive interventions for those households at greater risk of experiencing a future deprivation becomes crucial in a context of economic deceleration, such as the one through which Peru is going through nowadays. In this way, the study of vulnerability in the Peruvian context raises a series of crucial questions for the design and implementation of policies in the following years: To what extent has poverty reduction enabled households to have a sustainable exit out of poverty? Were the households that emerged from poverty are part now of the vulnerable population or did they exit permanently without risk of falling back into poverty? To what extent are non-poor households susceptible to fall into poverty as a result of an adverse shock? To what extent do social protection mechanisms have a role in reducing the risk of poverty? Does health coverage reduces vulnerability? In this paper, we will present an estimation of vulnerability lines, analogous to the poverty ones, which enable us to decompose the non-poor population into households that are

1 vulnerable to poverty and those that are not vulnerable. Our approach adopts a definition of vulnerability as the ex-ante risk of an unfavorable transition to poverty. We estimate vulnerability indexes and vulnerability lines based on panel data from the National Household Survey (ENAHO) produced by the National Institute of Statistics (INEI). We present the evolution of the vulnerable households and their corresponding share within the total population and discuss the determinants of the vulnerability of households. Our results show that the vulnerability of households to poverty is "structural", related mainly to the characteristics of the labor market insertion, which produces high income instability; the demographic structure of the household; the level of education and ethnicity condition; and the geographical environment that defines productive opportunities and reflects the State density. Adverse shocks, which have traditionally been considered as the distinguishing factor of vulnerable households, only have a significant impact in the case of major shocks, such as natural catastrophes, or when they occur cumulatively. The document is divided into five additional sections to this introduction. The second section discusses the concept and measurement of vulnerability and reviews the empirical literature on the subject. In the third section, we present the data and the empirical strategy used for the vulnerability lines estimation. The fourth section show and discuss the results obtained; while in the fifth, several analyzes of sensitivity and robustness of our results are presented. Finally, the sixth section concludes. 2. Concept and Measurement of Vulnerability The conceptualization of vulnerability has had a long discussion coming from several approaches and with different nuances both in its definition and in the way in which it should be measured and operationalized in the empirical research. First, it is crucial to differentiate the concept of vulnerability, in its broadest sense, from the poverty one. As Chaudhuri (2003) mentions, vulnerability and poverty can be understood as two sides of the same coin. Poverty is an ex post welfare situation or realization which reflects a state of deprivation, lack of resources or lack of capabilities. Vulnerability, on the other hand, is an ex ante measure of well-being whose main intention is not to reflect the present welfare but rather an uncertain state in the future. In this sense, we can argue that poverty is the ex post realization of a stochastic variable; while vulnerability will be the expectation on that variable in relation to a threshold previously defined (Christiaensen & Subbarao, 2005). Thus, poverty can be observed for a specific moment while vulnerability is shown as a dynamic and uncertain phenomenon because it represents the possibility (risk) of ending in a state of deprivation in a future period of the life cycle of the household. In this way, a household will become vulnerable when it is especially susceptible to losses due to negative shocks, caused by a long exposure to risks, weak internal conditions or by bad risks management (World Bank 2013). Hoddinott & Quisumbing (2003) point out that the first step in understanding vulnerability consists in relating the source of the risks faced by economic unity, resource holding, and risk management techniques applied. In the core of their conceptual framework, the authors present three key components of the household: adjustments, which describe the

2 context in which a household reproduces and develops; assets, understood in its broadest sense as income-generating resources; and activities, which refers to the distribution of these resources in income-generating activities conditioned by the adjustment or context. On the basis of this conception we find shocks, ex ante management tools and ex post responses within the adjustments. Immersed in this context, households have labor and capital endowments - physical, natural, human, financial and social - that will decide to invest in activities. Such allocation of resources to activities, in addition of being conditioned by the context, carries certain probability of a shocks occurrence. For example, investing time and money in being a dryland farmer or using social capital to found a risk pooling society. The presence of adverse shocks can reduce the welfare of the household below the socially accepted thresholds, that is, what we call a negative transition or fall into poverty, which will put the household in a situation of deprivation. Thinking about vulnerability and exposure to exogenous shocks removes the assumption about the perfect certainty of households to the future; for which vulnerability can also be understood as the burden caused by the threat of poverty (Calvo & Dercon 2013). The analysis of welfare under the conceptualization of vulnerability leads us to think of potential realizations of deprivation states and the identification of the poor in beforehand which allows us to design ad hoc prevention programs targeting the households with these risk conditions (Celidoni & Procidano 2015; Chaudhuri 2003). To this end, the operationalization of the concept of vulnerability in an ex ante measure must be able to inform about potential future deprivations, must be associated with a negative potential result, and must be a measure at each economic agent level (Foster, Dutta & Mishra 2010). The operationalization of the concept of vulnerability discussed previously has generated different lines of investigation, having different operationalization. The first moment of divergence appears when choosing the outcome variable on which to measure the uncertainty or vulnerability. The mainstream favors the expenditure as a summary indicator of household well-being, as it is used for national poverty measures which allows comparison between both indicators, since expenditure presents a less volatile behavior over time, and because it also reflects different strategies for mitigating shocks due to the loss of short-term or long term income generation capacities such as the sale of assets, the use of savings or recurring to loans. At the empirical level the discussion of vulnerability has taken a "triple bifurcation". First, we have the vulnerability understood as the ex-ante risk of falling into poverty (VEP), in which vulnerability is defined as the ex-ante probability that a household will suffer a future poverty episode. That is to say, their level of expenditure or income (푦ℎ,푡+1) falls below the defined poverty line (푧). Under this approach, a household will be vulnerable if its probability of falling into poverty (푉ℎ푡) passes a certain threshold for a future period:

푉ℎ푡 = Pr⁡(푦ℎ,푡+1 < 푧). This empirical treatment has been widely used for its ease of interpretation and its direct relationship with national poverty lines and calculations. However, the VEP approach is

3 often data-intensive, as well as setting an "arbitrary"3 threshold for the probability of falling into poverty from which if a household surpasses it will be considered as vulnerable. The second conceptual approach defines vulnerability as the ex-ante risk of presenting low utility in the future, understood as a condition of low expected utility (VEU) which is derived from a utility level determined by a certainty equivalent (푧) (Ligon & Schechter 2002; Pritchett, Sumarto & Suryahadi 2000). In this sense, we will have that the level of vulnerability is calculated as the difference between the household's ℎ utility at a level of consumption determined by the certainty equivalent and the expected utility of consumption, 푐ℎ,

푉ℎ = 푈ℎ(푧퐶퐸) − 퐸푈ℎ(푐ℎ) This approach, that may be attractive in terms of the consideration of risks, requires several assumptions such as choosing a functional form for the utility, the risk aversion parameter of the household; while has the disadvantage of measuring vulnerability in term of utility units; which can be unhelpful for public policies (Celidoni & Procidano 2015). The third approach considered in the measurement of vulnerability is known as uninsured risk exposure (VER), related to the literature on consumption smoothing and considered backward-looking by being an ex-post evaluation of whether the household had the capacity to smooth consumption and mitigate the effects of income shocks (Celidoni & Procidano 2015). In its objective of understanding the mitigation capacity of households, this approach will conclude that if consumption and income are highly correlated then households will have employed inefficient risk management tools increasing their vulnerability to shocks. Following the next equation:

△ 푐ℎ,푡,푣 = 훽 △ 푙푛푦ℎ,푡,푣 + 훿푿ℎ,푡,푣 + ∑ 훿푡,푣푫푡,푣 +△ 휀ℎ,푡,푣⁡ 푡,푣

Where △ 푐ℎ,푡,푣⁡denotes the consumption growth rate of household ℎ in the locality 푣 between two time periods, △ 푙푛푦ℎ,푡,푣 is the income growth rate, 푿 is the household’s observable characteristics and 푫 are community or time level controls,; the parameter of interest 훽 will represent the level of correlation between the growth of income and expenditures between two periods. This will reflect the ability to mitigate income shocks that would have an impact on household expenditure levels. This empirical methodology does not attempt to construct an aggregate indicator of vulnerability, but only quantify the impact of shocks over consumption (Tesliuc & Lindert 2002). In reference to this, Morduch (1994) emphasizes that the operationalization of vulnerability as the measure of the inability to smooth income over a temporal period may not be an adequate indicator.

3 Previous research in the VEP approach has given a series of recommendations for choosing this threshold. We will return to this point in the methodology section. 4

2.1 Empirical research in vulnerability Table 1 presents a synthesis of the empirical research published in the last years about the vulnerability of the households, emphasizing the investigations carried out for the Peruvian case. Table 1: Previous research on household vulnerability

Author Year Country Estimation Main Finding

Panel A. VEP Approach Share of vulnerable is between 1.5 Pritchett et al 2000 Indonesia Vulnerability Lines and 2.5 times the number of poor Christiaensen & Pseudo panels: average National average of 39% 2005 Kenia Subbarao consumption probability of falling into poverty OLS: per capita Heterogeneity of shocks respect Dercon et al. 2005 Ethiopia consumption to the magnitude and persistence Nicaragua Synthetic panels: Mincer Empirical validation of the Cruces et al. 2011 Chile and Peru equations synthetic panel methods Probability of having a Higher exposure of rural areas, Haq 2012 Pakistan shock informal insurance mechanisms Lopez-Calva & Mexico Probability of falling into Increase in non-vulnerable 2014 Ortiz-Juarez Chile and Peru poverty and Mincer households Dang & Vietnam, India Vietnam as a country with greater 2015 Insecurity Index Lanjouw and USA "shared prosperity"

Panel B. VEU Approach Morduch & High vulnerability of households 2002 Ivory Coast Monte Carlo methods Gamanou outside the country's capital Ligon & Deviation of Poverty and risk have similar 2002 Bulgaria Schechter consumption differntial incidence in welfare reduction Patterns of vulnerability different Calvo & Autorregresive 2013 Ethiopia from poverty patterns in the Dercon consumption model country

Panel C. VER Approach Bangladesh, Variability in consumption relative Skoufias & Mali, Ethiopia, Volatility of consumption 2004 to income related to negative Quisumbing Mexico and relative to income results Russia Source: Elaborated by authors. Under the VEP approach, we identified seven studies between 2000 and 2015. In the first place, we have the work of Pritchett et al. (2000) regarding vulnerability in Indonesia. In this paper the authors calculate vulnerability lines valued in terms of household expenditures using the Indonesian household panel and vulnerability is conceptualized as the probability of being poor in a near period, independently if the household is already poor or if it is out of poverty. Thus, considering as vulnerable a household with a probability

5 of 50% or more of falling into poverty in the next 3 years, the authors find that vulnerable households represent between 30% and 50% of the total population. On the other hand, Christiaensen & Subbarao (2005) employ pseudo panels derived from the cross-sectional databases to analyze the level of vulnerability in Kenya. The authors estimate the average and variance of future household consumption as a function of their characteristics and their localities’ attributes in the previous period. In this way, authors calculate that households in Kenya have an average probability of 39% of being poor in the future, being important risk factors the volatility of rains, for households located in arid zones, and the presence of malaria in non-arid areas. Dercon et al. (2005) analyze the situation of vulnerability and exogenous shocks in Ethiopia, for which they define different profiles and use an econometric strategy to evidence the impact and magnitude of shocks on the reduction of the level of future expenditure of the households. This research shows that the droughts and diseases are the most important shocks. In this line, a closely related analysis is done by Haq (2012) who studies the exogenous shocks in Pakistan and found that the most exposed households in Pakistan turn out to be the rural ones with many members and with elders as household head; while natural and agricultural shocks are those with the greatest share on the total impact of shocks. The article by Cruces et al. (2011), which analyzes three countries in the region including Peru, constructs synthetic panels to predict the future income of households and study the intergenerational mobility of entries and exits of poverty through transition matrices and growth incidence curves (GIC), concluding that the estimates of synthetic panels are close to those obtained from the observed panels and suggest that they are an interesting option to consider in the case of countries without longitudinal household data. Finally, under the VEP approach, we have the works of Dang & Lanjouw (2015) and López- Calva & Ortiz-Juárez (2014), who employ two different ways of estimating vulnerability lines. The first paper, constructs vulnerability lines that we call "unconditional". This line is calculated directly through an estimate of the probability of falling into poverty for the following period; whereby a probability threshold is chosen and the average value of the expenditure (or income) is calculated for those households with a probability of falling equal to this threshold. This monetary amount will be the value of the unconditional vulnerability line, and those households that are above it will be considered as non- vulnerable. In the case of López-Calva & Ortiz-Juárez, they also estimate a vulnerability line that we call "conditional", because it uses the predicted expenditure (or income) of a Mincer equation estimate based on household attributes instead of the observed values to calculate the line. One considerable advantage of building such lines is that it allows the decomposition of households into three categories: poor, vulnerable and non-vulnerable. Dang & Lanjow find that Vietnam presented the largest expansion in pro-poor growth for the last five years of the 2000 when compared to the cases of India and the USA. On the other hand, López- Calva & Ortiz-Juárez evidenced a significant growth of the middle class in the three countries analyzed along with the persistence of a considerable group of households that still face the possibility of falling into poverty.

6

For the VEU approach, the first research applied to Ivory Coast by Morduch & Gamanou (2002) uses Monte Carlo simulations to estimate the expected distribution of future household expenditure and thus to calculate the vulnerability as a function of these distributions. The authors find a situation of high vulnerability in households outside the capital, which until that moment was not revealed by previous methods of analysis. Ligon and Schechter (2002) apply a measure of vulnerability that quantifies the loss of welfare (utility) associated with poverty as well as uncertainty in Bulgaria. The measurement of the proposed vulnerability is calculated from the difference between the utility of the household when consuming a basket with certainty and the expected utility of consumption. They concluded that poverty and risk play similar roles in the reduction of well-being, while the covariant shocks are of greater importance than the idiosyncratic ones. As a third application of the VEU approach, we have the research published by Calvo & Dercon (2013) where a set of vulnerability indicators is proposed, understanding the vulnerability as the burden caused by the threat of poverty before and after it occurs or not. Although the main purpose of the work is to derive these measurements under a framework of expected utility and uncertainty, an empirical application is also made for the Ethiopian case using longitudinal data. The authors find that the vulnerability profile shows different correlates compared to the standard ex-post poverty profile. Finally, in the VER approach, the research of Skoufias & Quisumbing (2004) synthesizes the evidence for five countries, examining in particular the ability of households to secure their consumption through formal or informal arrangements against economic shocks and income fluctuations. The results show a strong heterogeneity in the ability of households to protect themselves against shocks of various types. The authors suggest that the creation of social protection schemes, not only targeted on the poor but also assisting households and communities in better risk management, would bring significant improvements in the level of well-being. 3. Estimation of the Vulnerability of Households to Poverty in Peru While in the case of the poor households, their identification is based on an ex-post finding, the identification in the case of the vulnerable population is rather based in an unobserved situation. In this research we adopt the conceptualization of vulnerability as expected poverty (VEP), thus the vulnerable population will be those non-poor households in the initial period but with a high probability of falling into poverty. We adopt the approach proposed by Dang and Lanjouw (2014) regarding the construction of vulnerability lines obtained directly from a vulnerability index and a chosen threshold corresponding to the lower limit of the risk of becoming poor. Given these values then the value of the vulnerability line will be calculated as the value of the average income or expenditure of those households that has exactly the probability of falling chosen as threshold. In this way, the construction of such lines will allow us to differentiate the population into four large groups: (i) Extreme poor, composed by households that are below the extreme poverty line; (ii) Non-extreme poor, as those households that are above the extreme poverty line but below the non-extreme poverty line; (iii) Vulnerable non-poor, as those

7 households that surpass both poverty lines but are below the vulnerability line; and (iv) Non-vulnerable non-poor households which will have income or expenditures above the three lines. The households in this latter group have been described by various authors as belonging to the "middle class" (Dang & Lanjouw, 2014; López-Calva & Ortiz-Juarez, 2014).

3.1 Database In the absence of panel data, the identification of the vulnerable population has generally been based on the examination of the factors associated with the probability of poverty in a given year. However, the poor in the initial period are composed of chronic poor and transient poor, who have experienced episodes of poverty and non-poverty. We will construct a vulnerability index that indicates the probability of falling into poverty for the households along all the income or expenditure distribution by estimating the risk of poverty for those household that have become poor and those that did not fall. Alternative proposals have considered the use of synthetic panels for the vulnerability estimation. Taking as an example the Peruvian case for which an observed panel is available and comparing it with the results of the synthetic panel, Dang & Lanjouw (2013) obtain results very close to the observed values of the poverty transitions. However, the estimated profiles of poor and non-poor households in the synthetic panels differ significantly from the observed population values. As shown in Figures 10 and 11 of Annex A.1, the projections of the synthetic panels deviate from the population values in several central characteristics of the households and even end up being outside the confidence intervals, excepting the cases of those estimates with little precision and wide intervals, which are not very precise. This is due to the low incidence of households experiencing some kind of favorable or unfavorable mobility related to poverty. Furthermore, the use of synthetic projections implies a loss of precision in the cumulative estimate over time as the standard errors are increasing as the estimation period expands. To the extent that the estimation of the vulnerability lines and the identification of the vulnerable population according to the proposed approach must be able to observe the poverty transitions between two time periods, we decided to use the long series of panel data for Peru that correspond to an observed panel obtained from the National Household Survey (ENAHO) between the years 2004-2014, that allow us to obtain biannual panels that cover every couple of consecutive years4. The size of the panel sample obtained for each biannual period is considerably larger than the one used in the research by López-Calva and Juárez-Ortiz panel5 (2014), besides covering consecutive years, while in the case of the cited authors, poverty transitions between the initial year of 2002 and the final year of 2006 are considered which involves several potential biases. First, considering transitions between two distant years tends to underestimate the vulnerable population to the extent that households may have been

4 Given the changes in the sample design, there is a break in the panel series in 2006/2007. In 2011 a new panel is started. However, a part of the 2010 sample (7450 households) remained in the 2011 sample making it possible to construct a biannual panel. 5 In the ENAHO panel for 2002 and 2006, there are only 260 observations for entries into poverty and 1261 observations combining household that entered and “remained” poor out of a total of 3142 households present in both years.

8 non-poor in the extreme observed years and yet have experienced a poverty episode between those years. Second, considering poverty transitions between distant years ignores the more complex trajectories and not distinguishing neither different degrees of vulnerability according to the number of episodes into poverty. For example, there will be no difference between non-poor households in the initial period that remain in poverty for the next 4 years of the panel and those household which are non-poor for all periods except the latter, where they finally experience a negative transition from poverty. Third, the authors, for reasons related to the small size of the sample examined, include as a dependent variable in the estimation of the vulnerability index both the households that fall into poverty and those that "remain poor". To the extent that the factors associated with vulnerability (transient poverty) may be different from those associated with chronic poverty, the vulnerability index is difficult to interpret besides that both the vulnerability lines and the decomposition of the population in poor, vulnerable and non-poor non- vulnerable will be somewhat distorted. Finally, the attrition of a non-rotating panel of 5 years is cumulative and by the end of the period the panel sample can show important biases correlated with the income variables. The biannual panels allow us to consider a more recent and extended period for the analysis (2004-2014), which makes it possible to analyze whether or not the vulnerability index is constant over time. Having panels that cover a decade in which the growth patterns have presented strong contrasts let us consider the potential effects of the macroeconomic cycle on the poverty transitions. Table 2: Size of the biannual panel simple ENAHO 2004-2014

Falls into Remains Total Panel poverty non-poor panel 2004/2005 349 1605 4229 2005/2006 372 1997 4606 2007/2008 455 3062 6137 2008/2009 497 3208 6066 2009/2010 436 3478 6136 2010/2011 504 3623 6007 2011/2012 593 4784 7450 2012/2013 398 4877 7068 2013/2014 602 5538 7924 Total 4206 32172 55623 Note: Transitions of poverty by expenditures. Source: ENAHO 2004-2014. Elaborated by authors. Studies on the dynamics of poverty and vulnerability recognize the importance of considering the macroeconomic dimension of vulnerability6 but due to lack of data this dimension is usually not taken into account in the empirical estimates. In fact, it is possible

6 See for instance Chaudhuri (2003) 9 that the economic slowdown or variations in the idiosyncratic or macroeconomic shocks have caused the changes in the levels of the vulnerability lines. In order to take account of the macroeconomic dimension, we build a pooled database using the 9 biannual panels and included fixed effects per year, which will allow us to capture the effect of macroeconomic factors. The set pooled biannual panels allows to obtain a sample of more than 55 thousand observations and 4206 observations of households falling into poverty (Table 2). Our results from the pooled panel will be compared with the results obtained from the estimates of each biannual panel as a measure of robustness (Annex A.6 and A.7). In particular, we would analyze whether the vulnerability lines are sensitive to fluctuations in the growth rates and whether the economic slowdown has lowered the vulnerability threshold by increasing the proportion of vulnerable households.

3.2 Identification Strategy: Vulnerability Lines With respect to the VEP approach, Dang & Lanjouw (2014) develop a vulnerability index which continues directly with the construction of vulnerability lines analogous to the poverty lines. These vulnerability lines represent a threshold of income or expenditure, higher than the poverty line, which identify the households that face a high risk of falling into poverty if they are below this threshold. Pritchett et al. (2000) classify households below the vulnerability line as those households with a high risk of future poverty; this group comprehends the current poor households with high probability of remaining in poverty and non-poor households in the first period are at high risk of falling. We have opted for the approach proposed by Dang & Lanjouw (2014) to use the vulnerability lines to identify non-poor households at high risk of becoming deprived. In contrast to other methodologies that choose ad hoc vulnerability lines in which the poverty line is multiplied by an arbitrary factor, the proposed methodology estimates an upper bound of expenditure or income, above which non-poor households have a low risk of falling into poverty; while non-poor households with incomes or expenditures below such bound face a considerable risk of falling into poverty. In this case, the bound’s value corresponds to the vulnerability line, analogous to a poverty line. The advantages of this non-parametric estimation of the vulnerability line lies in the possibility of comparing the incidence of vulnerability for a country at different time periods, achieving a direct estimation through panel data or repeated cross-sections and the possibility of the exact decomposition of the population in the three groups mentioned. Formally, the definition of the vulnerability line is obtained from the equation referring to the vulnerability index. Given a vulnerability line (푉0) such that a specific proportion of households with consumption levels below it in the first period (푦0) will fall below the poverty line (푧) in the following period, the population above this vulnerability line will be considered as non-vulnerable. The probability that a household falls below the poverty line for the second period is what the authors call the insecurity index denoted by Ρ1. 1 Ρ = Pr⁡(푦1 ≤ 푍1|푦0 > 푉0)

10

In the same way, by specifying a minimum level of insecurity that differentiates vulnerable and non-vulnerable households, the vulnerability line of the previous equation 푉0 must satisfy the equality, whereby we obtain the exact value of the line in terms of expenditure or income. At this point, it is necessary to choose the vulnerability threshold of households that fall into poverty Ρ1, because higher levels of the insecurity index will generate lower vulnerability lines and vice versa. For this, the authors point out that most objective ways of setting the level of Ρ1 is to use as benchmark the observed value of the proportion of households that fall into poverty through the estimation of transition matrices in case there is panel data available as applied by Cruces et al. (2011). An approach similar to the one by Dang & Lanjouw is the parametric estimation of vulnerability lines, or what we call the "conditional" vulnerability lines developed by López- Calva & Ortiz-Juarez (2014). In this research, the authors apply the same concept of vulnerability developed in Dang & Lanjouw (2014) but they condition the estimation to a set of household variables in a three-stage process that ends with the modeling and estimation of a Mincer equation that would, according to the authors, have the advantage of reducing the volatility of income by using the predicted income instead of the observed one, thus become an index related to the stock of assets as the income generation capacity of the households (López-Calva & Ortiz-Juarez 2014). The Dang & Lanjouw's (2014) "unconditional" strategy consists of a two-step estimation in which we first estimate a vulnerability equation using a limited dependent variable model to obtain the estimated risk for each household to become poor in the following year. Then, using the estimated probabilities, we apply a probability threshold and the average value of the result variable (income or expenditure) is calculated for households that are in that threshold value. Following the proposal of Dang & Lanjouw (2014) and Cruces et al. (2011) we set the threshold of vulnerability according to the value observed in the transition matrices of the proportion of non-poor households that falls each year in poverty. As we can see in Table 4, the average proportion of non-poor households that becomes poor for the decade under study is 10.5%, therefore we decided to set our threshold at 10%7. At this level, the choice of the threshold seeks to be conservative, in the sense of considering non-vulnerable households those that face a risk of less than 10% of falling into poverty, which assures us that we are considering households with relatively sustainable escapes from the deprivation situation. Moreover, according to Ravallion (2016), this choice is based on considering non-poor non-vulnerable households as a measure of the middle class; that is, households that have such a level of income or expenditure that makes it very unlikely to fall into poverty. Such households would be safe from unfavorable transitions of poverty by having a greater capacity and resources to cope with adverse shocks and crises. Once these households were identified, we proceed to estimate the value of the result variable, for which we have considered both the level of real per capita expenditure and income at Lima prices. In this second stage, we obtain the average level of income and

7 As a measure of robustness, other thresholds (5, 15, 20 and 25%) were tested to verify the change in the value of the lines. The detail is presented in Annex A.4. 11 expenditure for all households that are at the threshold of 10% probability of falling in poverty, given our estimate of the vulnerability equation. We consider a caliper of one percentage point; that is, households with an estimated probability of fall greater than 9% and less than 11%. We employ this caliper since it is unlikely to find a considerable number of households with an exact predicted probability of 10%, with which a neighborhood of one percentage point turns as a conservative choice8. We estimate a logit model with four groups of explanatory variables of the probability of falling into poverty for the next period: ′ ′ ′ ′ 푃푖푡 = 퐸(푝표표푟푖푡+1|푿푖푡, 푯푖푡, 푮푖푡, 흃푖푡) = 훼 + 푿푖푡휷 + 푯푖푡휸 + 푮푖푡휽 + 흃푖푡휹 + 휆푡 + 휀푖푡

Where the dichotomous dependent variable 푃푖푡 takes the value of one if the household falls into poverty in the following period while the base category - the zero value – are the households that remain non-poor. It is important to note that we are only considering households that fall into poverty compared to those who are never poor in the biannual transition, thus discarding the ever-poor and the households that exit poverty.

The vectors 푿푖푡, 푯푖푡, 푮푖푡⁡푦⁡흃푖푡 denotes the characteristics of the living place, the household’s head, the geographic context and the shocks experienced by the household’s members respectively. All the independent variables of the regressions are taken in the initial year of the transition9 . The definitions as well as summary statistics are presented in Table 5 and Table 6. The logit regression was carried out considering 9 biannual panels available from ENAHO in a "pooled" database. We included heteroskedasticity robust errors as well as year fixed effects (휆푡) since the observations came from different biannual panels and for capturing the effect of macroeconomic fluctuation on the vulnerability of households. Likewise, observations with extreme values for both income and expenditure were eliminated before making the estimation10. Given the potential improvement in the adjustment of the vulnerability equation given the model, we also estimate the same specification using a multinomial logit, which was discarded because it had a lower predictive capacity than the binomial case. We return to this point in the final section of sensitivity and robustness. With this, the estimated vulnerability line is calculated through the average given by:

0.11 1 푉푢푙푛푒푟푎푏𝑖푙𝑖푡푦⁡퐿𝑖푛푒⁡10% = ∑ 푌 푁 푝̂ 푝̂=0.09 Where 푌 denotes the income or expenditure variable, 푝̂ represents the estimated probability of falling into poverty (or vulnerability index) for each household, while N is the number of households within the range 푝̂⁡휖⁡]0.09⁡; 0.11[. In other words, we take the

8 We also analyzed the sensitivity of the results to different amplitudes of the interval or caliper in section 5. Given the sample size we were able to obtain a robust estimate with a small caliper. 9 Except for the abandonment of the household’s head that could result endogenous. The other variables are measured in the initial period (i.e. before the transition) 10 We employ the criterion of eliminating the values of the upper 1% of the distribution. 12 average of the value of the income or expenditure for all the households in the one percentage point caliper. On the other hand, the conditional methodology of the lines of vulnerability includes, like the previous strategy two stages of estimation. In the first one, the same vulnerability equation is estimated. The difference between these two methodologies is the second stage of estimation where, for the conditional methodology, we calculate the averages of the variables included in the regression for each percentile according to the distribution of the estimated probability. In this way, an income or expenditure equation is estimated and these estimators are multiplied by the average values per percentile of the first stage to obtain the predicted income or expenditure. Finally, the estimated value of these variables is considered for the tenth percentile, which corresponds to the value of the 10% vulnerability threshold. Following this, the first stage of the conditional methodology will be equal to the regression equation previously shown, while in the second stage the income or expenditure equation will be estimated through ′ ′ ′ ′ ln(푌푖푡) = 휑 + 푿푖푡흎 + 푯푖푡흁 + 푮푖푡흉 + 흃푖푡흅 + 휆푡 + 휂푖푡 In order to obtain the estimated income or expenditure through ̂ ̅′ ̅ ′ ̅′ ̅′ ̂ ln(푌푖푡) = 휑̂ + 푿푖푡흎̂ + 푯푖푡흁̂ + 푮푖푡흉̂ + 흃푖푡흅̂ + 휆푡 In which the variables denoted with a bar represent the averages of the independent variables for the threshold percentile of 10% of the estimated probability of falling into the poverty of the first stage, taking the same caliper of one percentage point. That is, the estimated income is obtained by multiplying the estimated returns to the factors in the Mincer equation by the average value of the attributes of those households with a probability of falling in poverty equal to the defined threshold. Table 3: Coefficient of Variation – Income and Expenditures

Average Using sample weights Coefficient of Observed Estimated T-test (p-value) Variation Income 0.312 0.310 0.972 Expenditures 0.227 0.219 0.704 Note: To find the estimated values an OLS model was used with the same specification as the vulnerability equation except for the variable of educational attainment that was replaced by years of education and the labor sector was suppressed by missing values and small amount of sample (only 912 households observed for the 5 consecutive years). Source: ENAHO 2004-2014. Elaborated by authors.

Finally, the desired vulnerability line will be obtained by correctly11 transforming the value

11 Following the proposal of Cameron & Trivedi (2009) about the correct transformation in level from logarithms, we cannot simply use the antilogarithm since

퐸[푌̃푖푡|푿푖푡, 푯푖푡, 푮푖푡, 흃푖푡] = exp⁡{ln(푌̂푖푡)} 퐸[exp(휂푖푡)]. With this we consider the multiplicative term of the error which, assuming that it is independent and identically distributed, is approximated 1 1 through 퐸[exp(휂 )] = ∑푁 ∑푇 exp⁡(휂̂ ) 푖푡 푁 푇 푗=1 푘=1 푗푘

13 of the logarithm of the estimated income or expenditure to its value in levels, since it will represent the prediction of those variables for households that are exactly at the threshold (or in the defined neighborhood). Our estimation employs the methodology of "unconditional" lines based on the proposal of Dang & Lanjouw because the second strategy of identification followed by López-Calva and Ortiz-Juárez (2014) of "conditioned" lines is used under the argument of avoiding the volatility of incomes or expenditures by using predictions to replace the observed values, without being able to present an empirical test about this point since they only consider two periods (op. cit. pg. 33). To test the instability argument of López-Calva and Ortiz-Juárez, we have calculated the average coefficients of variation for 5 years panel (2007-2011) in which we compare these coefficients between the observed and estimated (through an ordinary least squares model) income and expenditure. We obtain very similar levels of variation for both the estimated and observed values in the case of income and expenditures. Also, we see that the difference between the coefficients of variation is not statistically significant for any of the outcome variables, which was argued in favor of using the conditional methodology to obtain the vulnerability lines instead of an estimation using directly the observed average value around the vulnerability threshold. These results incline us to favor the estimation of unconditional lines given also our sample sizes and data quality. In our case, we prefer the unconditional formulation because our data refer to biannual panels, which means that the volatility is much lower than in the case of López-Calva and Ortiz-Juárez, who take a single panel considering two non- consecutive periods for the Peruvian case in a four years of interval. We present the results of the conditional lines in Table 11. In the estimation of the vulnerability lines with the pooled panel regression, we considered the expenditure or income variables at constant prices of 2014 and spatially deflated at Lima prices. For the purposes of the decomposition of households in extreme poor, non- extreme poor, vulnerable and non-poor, thresholds were calculated at spatially deflated current prices. The biannual expansion factors used were provided by the INEI sampling unit. The panels used were elaborated by the INEI from the matching of the households. A panel household is defined as one that has at least one member in common in two successive periods. It is important to note that in both estimation stages the values of the result variable - either this income or expenditure- were considered in monthly values, per capita12, deflated to Lima’s price level and in real value at 2014 prices. Therefore our estimated lines should be interpreted as the per capita expenditure or income level of the household at Lima 2014 prices below which a household will be considered vulnerable.

12 That is, divided between the number of household’s members 14

4. Results

4.1 Characterization of the dynamics of poverty in Peru Before addressing the operational definition, identification and measurement of the vulnerable population under the dynamic approach, we will examine the mobility of households year by year with respect to the poverty situation through transitional matrices from the ENAHO (2004-2014) 13 longitudinal data. Table 4 presents the transition matrices for each consecutive pair of years of the panel. The proportion of households that remain poor for two consecutive years declines sharply over the decade, from 42% to 11%; while the opposite happens for households that remain non-poor, whose relative share doubles in the period under analysis. However, in contrast to the large variations of households that remain in the same condition, households that escape or fall into poverty vary relatively little in percentage terms through the years. The first years of the decade show a strong pattern of exit from poverty with a peak of 13% in 2006/2005 that will gradually reduce to its lowest point with only 7.4% of households leaving poverty for the last couple of years. We see that the escapes from poverty follows a pro-cyclical behavior, which is in line with the last years of macroeconomic slowdown and the greater difficulty of continuing to reduce the national poverty rates. On the other hand, when we consider the percentage of households that fall into poverty, we find that the observed values fluctuate around 7.2%, presenting their lowest levels for the last two pairs of years. Table 4: Poverty transitions matrices (expenditures) 2004-2014

Escape % Poor % Non- Fall into Remain Remain Period from households poor poverty poor non-poor poverty that escaped that fall 2005/2004 7.9% 42.5% 36.9% 12.6% 22.8% 17.7% 2006/2005 8.1% 36.9% 42.0% 13.0% 26.1% 16.2% 2008/2007 7.1% 25.6% 56.0% 11.2% 30.4% 11.3% 2009/2008 7.3% 21.7% 61.4% 9.7% 30.8% 10.6% 2010/2009 7.4% 19.0% 64.4% 9.1% 32.2% 10.3% 2011/2010 8.0% 16.9% 66.0% 9.0% 34.7% 10.8% 2012/2011 7.5% 14.5% 70.0% 7.9% 35.4% 9.7% 2013/2012 4.9% 12.1% 73.7% 9.3% 43.7% 6.2% 2014/2013 6.5% 11.0% 75.1% 7.4% 40.4% 8.0% Average 7.2% 21.7% 61.3% 9.8% 31.1% 10.5% 2004-2014 Source: ENAHO 2004-2014. Elaborated by authors.

The "mobile" households (those that go from one situation of poverty to another) does not present strong patterns of change in the decade in comparison with those that maintain their condition. However, the downward mobility situation still represents a

13 The sampling design changed during the period so that three panel samples were available. The first one covers the period 2004-2006, the second goes from 2007 until 2011 and the third one covers the years 2012-2014 (last year available). The years 2002-2003 were not included because comparable data were only available for the fourth quarter of the year. 15 significant percentage of households that have not been able to sustain their escape or maintain their non-poor status. There is no evidence of a considerable reduction of the proportion of the downward mobility households over time, but on the contrary, in the last couple of years an incipient increase in households suffering from adverse mobility has been observed, a phenomenon certainly related to the deceleration of the macroeconomic growth. The last column of the table shows the proportion of households falling into poverty as a share of the total number of non-poor households. From the total non-poor households, we can observe a strong reduction of the share of households that fall into poverty in the first five years. However, in the following years a slow rate of reduction of the proportion of non-poor households falling into poverty is maintained, although for 2013/2012 there is a strong reduction compared to the previous couple of years. For the last couple of years the value presented shows an opposite trend of the previous years, with a rise in almost 2%, which could be aggravated by a lees favorable macroeconomic context. Another relevant element to be examined in relation to household mobility and complementing the information of the matrices is the measurement of how far from the poverty line are located those households that managed to escape the poverty situation, as well as the distance below the line at which households that suffered an adverse transition from poverty remains. This indicator will allow us to determine if the fall in poverty occurs mainly for households whose expenditures are close to the poverty line or if, on the contrary, the adverse transitions of poverty are due to significant falls in the spending capacity. We can also analyze whether households escaping from deprivation are sufficiently distant from the poverty line or remain in the "neighborhood" of this line, subject to a potential relapse.

16

Table 5: Variable definitions

Variable Definition Dichotomous variable that takes the value of one in case the household Vulnerable falls into poverty in the following period being non-poor in the initial period and zero in case it does not fall into poverty. Calculated for expenditures and income.

Living place characteristics Continuous variable calculated as the ratio of the number of income Dependency rate inverse perceptors among the total number of household members Nuclear household Dichotomous variable equal to one if the household is nuclear Single parent household Dichotomous variable equal to one for single parent households A series of dichotomous variables that identify if the household has Unmet basic needs (NBI) unmet basic needs Internet Access Dichotomous variable equal to one if the household has Internet access Dichotomous variable equal to one if the household’s member do not Associations belong to any association (religious, cultural, political or social). Number of assets owned by the household (car, refrigerator, TV, radio, Household’s assets etc.) A series of dichotomous variables that identify the strategy of Strategies against shocks households against shocks: savings spending and loan, only savings spending, only loan, no shock

Household’s head characteristics A series of dichotomous variables that identify whether the household Age cohort head is younger than 46, is between 46 and 65 years old, or is over 65. Gender: male Dichotomous variable equal to one if the household’s head is male Dichotomous variable equal to one if the mother tongue of the Indigenous mother tongue household’s head is native language A series of dichotomous variables that identify if the household’s head Education level does not have any educational level, has primary, secondary or tertiary education. A series of dichotomous variables that identify whether the household’s Insurance holding head has SIS insurance, EsSalud insurance, other health insurance or does not have any health insurance. A series of dichotomous variables that identify if the household’s head Income inestability perceives his income as very unstable, unstable or not at all unstable. A series of dichotomous variables that identify the household head’s Employment sector employment sector classified as primary, secondary or tertiary

Geographic context Living place area Dichotomous variable equal to one if the living place is in an urban area

A series of dichotomous variables that identify the life zone where the Life Zones (Pulgar Vidal) living place is located: Chala, Yunga, Quechua, Suni, Puna, Janca, Rupa Rupa y Omagua

17

Variable Definition

Exogenous shocks Dichotomous variable equal to one for those households in which the Household’s head household’s head of the initial period is not present in the household abandonment for the final period. Dichotomous variable equal to one for those households that suffered Member’s death the death of one of their members in the last 12 months. Dichotomous variable equal to one for households where members lost Employment shock their jobs Dichotomous variable equal to one for households whose family Business shock business went bankrupt Dichotomous variable equal to one for households that suffered a Robbery shock robbery Dichotomous variable equal to one for households that suffered natural Natural Disaster shock disasters Dichotomous variable equal to one for households whose proportion Health shock of health expenditure is more than twice the average proportion spent by households in the same quintile. Dichotomous variable equal to one for households that suffered other Other shocks adverse shocks Dichotomous variable equal to one for households that suffered more More than one shock than one adverse shocks Source: ENAHO 2004-2014. Elaborated by authors

18

Tabla 6: Summary Statistics

Variable N Mean Standard Deviation

Dependent variables Vulnerable by expenditures 54891 7% 0.26 Vulnerable by income 54891 8% 0.27

Living place characteristics Dependency rate inverse 54891 0.54 0.24 Nuclear household 54891 60% 0.49 Monoparental household 54891 26% 0.44 NBI2: Overcrowded household 54889 10% 0.31 NBI3: house without toilet 54891 14% 0.35 Internet Access 54891 12% 0.33 Associations 54891 54% 0.50 Household’s assets 54891 5.31 3.48 Strategies: savings spending and loan 54891 1% 0.07 Strategies: Only savings spending 54891 7% 0.26 Strategies: Only loan 54891 3% 0.18

Household’s head characteristics Age cohort: 46 years or fewer 54891 41% 0.49 Age cohort: Between 46 - 65 years 54891 44% 0.50 Gender: male 54891 80% 0.40 Indigenous mother tongue 54891 26% 0.44 Education level: Secondary 45050 45% 0.50 Education level: Tertiary 45050 25% 0.44 Health insurance: SIS 55197 25% 0.43 Health insurance: EsSalud 55197 24% 0.43 Health insurance: Others 55197 2% 0.15 Health insurance: None 55197 49% 0.50 Income instability: High 50838 34% 0.47 Income instability: Medium 50838 45% 0.50 Employment sector: Primary 48317 37% 0.48 Employment sector: Secondary 48317 19% 0.39

Geographic context Living place area: Urban 54891 72% 0.45 Life Zones: Chala 55623 34% 0.47 Life Zones: Yunga 55623 15% 0.36 Life Zones: Quechua 55623 19% 0.39 Life Zones: Suni 55623 9% 0.28 Life Zones: Puna 55623 3% 0.17 Life Zones: Rupa rupa 55623 6% 0.24 Life Zones: Omagua 55623 13% 0.34

19

Standard Variable N Mean Deviation

Exogenous shocks Household’s head abandonment 54891 4% 0.19 Member’s death 54891 1% 0.11 Employment shock 54891 3% 0.17 Business shock 54891 1% 0.10 Robbery shock 54891 3% 0.16 Natural Disaster shock 54891 8% 0.26 Health shock 54891 13% 0.34 Other shocks 54891 9% 0.28 More than one shock 54891 2% 0.13 Source: ENAHO 2004-2014. Elaborated by authors. In Table 7 we present the average value of the expenditure divided by the poverty line, known as the welfare ratio, for the four categories of households determined in the transition matrices for the last two consecutive years of the ENAHO panel (the patterns observed for these years are repeated with little variation for the previous ones). Households that fell in poverty in 2014 had a per capita expenditure in 2013 36% above the poverty line. The fall is important as the average expenditure of these households once in poverty was 20% lower than the value of the poverty line. In the case of individuals who left poverty, they were 18% below the line in 2013 and once out of poverty in 2014 managed to reach a level of expenditure 36% above the poverty line. Note that households that managed to stay out of poverty in both years had a per capita expenditure 2.3 times the poverty line. Table 7: Household’s Welfare Ratio 2013 – 2014

Welfare Ratio 2013 2014 (Per capita expenditure/poverty line) Non-poor (2013) – Non-poor (2014) 2.26 2.27 Poor (2013) - Non-poor (2014) 0.82 1.36 Non-poor (2013) – Poor (2014) 1.36 0.8 Poor (2013) - Poor (2014) 0.71 0.7 Total Households 1.86 1.88 Source: ENAHO 2013-2014. Elaborated by authors.

Households that remain in poverty spend an average 30% below the poverty line. A significant poverty gap could therefore be an important factor that would make the outflows of poverty less probable, trapping households in that situation.

4.2 Determinants of the vulnerability to poverty In this section we present the results of the estimates of the determinants of vulnerability, which comprise three steps: the estimation of the vulnerability index and its determinants, the calculation of the vulnerability threshold, and the evolution of the relative shares of the vulnerable, poor and non-poor non-vulnerable population over the period 2004-2014.

20

The vulnerability index has been estimated from a binomial logit model and the coefficients are presented in the form of odds ratio insofar their values do not depend on the level of the explanatory variables, which facilitates their interpretation. We decided to estimate the vulnerability index considering the falls in poverty regarding an income level below the poverty line in the second period instead of the poverty-related transitions according to expenditures. The reason for this is that expenditures already integrate the different strategies of response of the households in the face of an adverse shock. Thus, the fluctuations in spending that would have occurred due to a sharp decrease in income are "smoothed" by resorting to savings or borrowing. In addition, a weakness of poverty according to the expenditures is that it considers as an improvement in the welfare an increase of the expense to attend a health problem or the funeral expenses of some member of the household. On the other hand, in countries such as Peru, there are still strong inequalities in health access according to income levels health coverage, so that the average health expenditure of the population around the poverty line does not represent the "socially necessary" expenditure in health. To mitigate this limitation, several authors have proposed to separately estimate the health component of the poverty line, taking into account the degree of satisfaction in health expressed by the households.14 In the first place, we observe that the vulnerability of households depends on factors related to the geographical environment of the living place. Thus, vulnerability increases as households situate on higher life zones. However, the vulnerability gradient is specific; residing in yungas or ravines (500 to 2300 masl) implies a risk 20% higher of falling into poverty compared to the residents of the coast. The risk of poverty reaches its maximum (47% higher than coast) in Suni or Jalca (3550-4000 masl), which is characterized by the scarcity of agricultural land and few possibilities of diversifying productive activities. Households living in the Quechua area (2300 to 3500 masl), who benefit from warm weather and rains from October to December, are 12% more vulnerable to poverty than coastal households. Finally, households living in the high jungle (Rupa Rupa) and lower jungle (Omagua) are less vulnerable than households on the coast (13% and 26%, respectively). Once life zones have been considered, the effect of residing in urban or rural areas has a contrary impact to the one found in the case of chronic poverty: the risk of vulnerability is greater (by 14%) for urban households compared to rural households. The greater vulnerability of urban households could be due to the fact that this variable captures the effect of precarious insertion in the labor market through family microenterprises that frequently break down. Both factors - informal sector employment and business bankruptcies - were explicitly included in different model specifications but were found to be statistically insignificant. The characteristics of housing, which approximates the structural deficiencies of households, are quite significant. Overcrowded households or the ones without toilets have a 58% and 13% higher vulnerability. Household’s wealth, measured by the number of assets, reduces significantly the vulnerability (by 20% for each additional asset) because of the chance to mobilize resources to cope against adverse shocks. Similarly, Internet access

14 Cordero, Herrera & Yamada (2003)

21 reduces the risk of poverty by 34%. This effect may have several channels, one of them probably related information on economic opportunities (job opportunities, niche markets, etc.). Different demographic structures have an impact on the ability of households to avoid episodes of poverty. The most important protective factor is the inverse of the dependency ratio, since the greater the proportion of income preceptors in the household, the lower the vulnerability (each percentage point decreases in 72% the risk of falling). Once controlled by this rate of economic dependency, nuclear households have a 20% lower vulnerability to poverty than extended households. In other words, the presence of other relatives does not provide protection against poverty. Single-parent households as well as female-headed households have a lower risk of having a downward transition (21% and 5%, respectively). This result is consistent with the evidence of developing countries as well as in static approaches of poverty in Peru. One explanation is that these variables are not strictly endogenous, but rather determined by other factors. Being able to separate from the spouse requires, for a woman, a minimum of economic independence, a certain level of family empowerment, a small number of young children, etc. In this sense, the presence of the spouse is not a protective factor against the risk of falling into poverty even though it may constitute an additional source of income. However, this relationship is more complex, considering the case of the abandonment by the household’s head, which implies a 56% increase in the risk of becoming. In this case, it is no longer a question of a woman’s possible choice by having met the conditions for her independence but an adverse exogenous situation. Not only does the presence or absence of the head of the household matter, but his characteristics play an important role. These characteristics are related to the life cycle, ethnicity, human and social capital, and the insertion in the labor market. In this way, the ethnicity of the chief, measured by mother tongue, is associated with a 14% higher risk of falling in poverty. Households whose head are in the age cohort with the highest labor participation rates (before the age of 46) are also the most vulnerable (+41%) compared to those who are retired from productive activities (those over 65). Those who have achieved more work experience (between 45 and 65 years) are the least vulnerable, although this effect is relatively small (-5% risk of falling into poverty). The years of education of the household’s head, directly linked to their productive capacities, constitute the main individual characteristic associated with a lower vulnerability. In this sense, having reached the highest level of education reduces vulnerability by 60% compared to those with only primary level. Secondary education also represents a significant reduction in vulnerability (-35%) compared to those chiefs who did not exceed the primary level of education. The risks of temporarily falling into poverty have traditionally been associated in the literature with adverse shocks suffered by households, while chronic poverty has been linked to deficits in physical and human capital endowments. This hypothesis can be contrasted with the data because the survey questionnaire includes, besides the characteristics and endowments of the households, questions about different adverse idiosyncratic and covariant shocks suffered by the households. Thus, we included in the regression controls for adverse shocks from natural disasters, victimization by criminal

22 acts, bankruptcy of the family business, loss of employment and other shocks. From the follow-up panel of individuals within the same household we could also add shocks in health variables as well as demographic shocks. Out of a total of eight individual adverse shocks considered in the regression, in addition to the aforementioned abandonment of the head, only natural disasters proved to have a negative impact on their own, increasing the vulnerability of the household by 35%. This shock mainly affects the population living in areas exposed to such risks (water, flood, frost, etc.); which is mainly found in the highlands and whose main source of income comes from agriculture. A natural disaster can have a negative effect not only on current incomes but also on future incomes by damaging the productive assets of the household as well as its houses and other assets that could have served to face other shocks. Although the resilience of households appears to be sufficient to deal with other shocks individually, this does not occur in terms of the accumulation of several adverse shocks because in this situation we observe that vulnerability increases by 33%. However, the structural variables are those who have the greatest impact on the vulnerability of households. First, a high instability of income reported by households has a considerable incidence on the risk of poverty. Households with very unstable incomes have a vulnerability almost 50% higher than households with stable incomes, while for households with more or less unstable incomes, this probability is reduced to 25% (see Figure 1). The competition of the market faced by workers is highly atomized. This is due to the fact that most workers are self-employed whose production units are very small and have low productivity. The low added value generated does not allow them to hire salaried workers (the average number of workers per unit of production is around 1.5), so that their income situates them around the poverty line and with a high likelihood of falling below it. Figure 1: Odds Ratio – Income Instability

Logit Estimation vs. Base category Stable Income

1.48

Very unstable 1.38

1.25

More or less unstable 1.30

1.00 1.20 1.40 1.60 1.80 Odds Ratio

Income Expenditures

Source: ENAHO 2004-2014. Elaborated by authors. This variability affects both agricultural producers and urban workers in informal microenterprises as the ones employed in sectors heavily exposed to cyclical variations

23 and market saturation effects. In fact, household’s head in the primary sector of activity (mainly agriculture) have the highest level of vulnerability (41% higher than workers in the tertiary sector), while it decreases to 18% for workers in the secondary sector (manufacture). This also reveals that few households have protection mechanisms because of reduced access to credit or other risk pooling mechanisms to reduce their vulnerability. The insurance and protection mechanism against adverse shocks count on the reduction of vulnerability; in particular those related to health shocks. The Figure 3 shows the relative impact of different health insurance situations and highlights, in particular, the fact that the type of insurance has a differentiated impact on vulnerability. Households insured by EsSalud or other private insurance are 55% and 66% less vulnerable to poverty, respectively. Households with SIS insurance coverage are more vulnerable than uninsured households. However, this seemingly paradoxical effect could simply mean that there is an inverse causality: SIS coverage is well focused, with priority being given to the poor and vulnerable households. In this case, the targeting criteria used by SIS identifies, besides the poor, also vulnerable households that accumulate various vulnerability factors and that have not been observed in the survey or not included in the model specification. Figure 2: Odds Ratio – Strategies against shocks

Logit Estimation vs. Base category No Shock

1.02

Savings expense and loan 1.21

0.72

Only saving expenses 0.83

1.13

Only loan 1.28

0.50 1.00 1.50 2.00 2.50 Odds Ratio

Income Expenditures

Source: ENAHO 2004-2014. Elaborated by authors. The vulnerability of a household to an adverse shock depends not only on the exposure to risk factors, but also on the ability to respond ex ante in terms of strategies to avoid adverse shock - such as productive diversification on different ecological floors in the case of farmers. Also, it depends on ex post response capacity in terms of mitigation strategies that reduce the consequences of the shock. Evidently, the resilience of households that suffered one or more adverse shocks will depend both on the severity and the resources available to the household to deal with it. Savings and access to credit are considered as the main instruments of households to mitigate the impact of shocks. However, according to the results of the estimated regression, these strategies have coefficients with opposite signs. While recourse to

24 savings reduces vulnerability by 28% for those households who suffered a shock, the ones that had to borrow to cope with the shock increased their vulnerability by 13% (see Figure 2). This unexpected result could be explained by the fact that indebtedness, usually from informal sources with usurious interest rates, is a strategy of last resort used by households with no assets and no previous savings. Debt accumulation can have a negative impact on vulnerability in a future period to the extent that its reimbursement may involve arbitrage in detriment of intermediate consumption expenditure and investment in household businesses, thus reducing their income to the point of causing a downward transition.

Figure 3: Odds Ratio – Insurance holding

Logit Estimation vs. Base category No Insurance

1.27

SIS insurance 1.26

0.45

EsSalud insurance 0.60

0.34

Other insurance 0.43

0.00 0.50 1.00 1.50 Odds Ratio

Income Expenditures

Source: ENAHO 2004-2014. Elaborated by authors. In this sense, having durable goods and other assets reduce the risk of poverty by 20%. The unexpected result of the use of credit may be due to market imperfections in the Peruvian case. Informal workers and farmers do not have access to formal credit markets (absence of collateral secured by title deeds, incapacity to prove regular and sufficient income, etc.) so they must resort to informal lenders who charge fees of usurious interest. The periodical payments ends up reducing in the short term the productive activities of the households (making it impossible, for example, the purchase of inputs) that should operate in favor of the repayment of the loan in order to preserve the financial credibility. Being able to resort the help of people outside the household will depend on the density, diversification and importance of social networks. In the estimation of vulnerability factors we have included a variable related to the "social capital" of the household, represented by the participation in different types of associations. Not belonging to any association implies a risk 11% higher than in the case of households that participate in any kind of association. The set of variables considered in the model have a high predictive capacity. The goodness of fit evaluated from the confusion matrix indicates that the model correctly predicts 86% of the falls into poverty in 86% (and 44% of the cases that remained out of poverty).

25

Table 8: Vulnerability Regressions - Logit Estimation

Logit Estimation – Odds Ratio Income Expenditures Dependent Variable: Vulnerable Household Urban (vs. rural) 1.140** 1.09 Life zones: Yunga (vs. chala) 1.200*** 1.125 Life zones: Quechua (vs. chala) 1.125* 1.235*** Life zones: Suni (vs. chala) 1.480*** 1.437*** Life zones: Puna (vs. chala) 1.239* 1.719*** Life zones: Rupa Rupa (vs. chala) 0.863 0.795** Life zones: Omagua (vs. chala) 0.737*** 0.622*** Dependency rate inverse 0.278*** 0.325*** Household’s assets 0.795*** 0.773*** NBI2: Overcrowded household (vs. not crowded) 1.578*** 1.587*** NBI3: House without toilet (vs. has toilet) 1.128** 1.158*** Internet Access (vs. no access) 0.678*** 0.491*** Single parent household (vs. no single parent) 0.789*** 0.683*** Nuclear household (vs. extended household) 0.801*** 0.707*** Age cohort: 46 years or fewer (vs. 65 or more) 1.419*** 1.079 Age cohort: Bewteen 46 - 65 years (vs. 65 or more) 0.95 0.905 Gender: male (vs. female) 0.95 1.068 Indigenous mother tongue (vs. non indigenous tongue) 1.133** 1.084 Education: Secondary level (vs. No level/Primary) 0.650*** 0.662*** Education: Tertiary (vs. No level/Primary) 0.389*** 0.391*** Health insurance: SIS (vs no insurance) 1.266*** 0.479*** Health insurance: EsSalud (vs no insurance) 0.448*** 0.342*** Health insurance: others (vs no insurance) 0.345*** 0.793*** Income instability: high (vs. stable income) 1.478*** 1.381*** Income instability: medium (vs. stable income) 1.243*** 1.300*** Employment sector: Secondary (vs. tertiary sector) 1.183*** 1.396*** Employment sector: Primary (vs. tertiary sector) 1.408*** 1.293*** Does not belong to associations (vs. belongs any assoc.) 1.111** 1.134** Only employment shock (vs. no shock) 1.099 0.836 Only business shock (vs. no shock) 0.879 1.015 Only robbery shock (vs. no shock que) 0.941 0.855 Only natural disaster shock (vs. no shock) 1.352*** 1.204** Only other shock (vs. no shock) 1.147 0.918 More than one shock (vs. no shock) 1.330** 1.152 Only health shock (vs. no shock) 0.998 0.889* Household’s head abandonment 1.560*** 1.06 Member’s death 1.128 1.751*** Strategy: Only savings expense (vs. no shock) 0.717*** 0.833* Strategy: Only loan (vs. no shock) 1.13 1.278** Strategy: Savings expense and loan (vs. no shock) 1.014 1.208 % corrected predictions for Non poor - Poor 85.82 80.52 % corrected predictions for Non poor - Non poor 44.25 46.05 Year Fixed Effects Yes Yes Prob > chi2 0.000 0.000 Count R2 0.591 0.598 McKelvey R2 0.394 0.390 Pseudo R2 0.194 0.184 Observations 25027 24204 Note: Significance is denoted by * = 10%; ** = 5%; *** = 1%. Odds ratios are presented, standard errors were obviated by space constraints and the base categories for dichotomic variables are presented in parentheses. The included confusion matrix takes as cut-off point 푝̂푖푡 > 0.1 for the correct predictions Source: ENAHO 2004-2014. Elaborated by authors.

26

4.3 Vulnerability Lines From the logit regressions we estimate the probability of falling into poverty for each of the households that were non-poverty in the initial period. The predicted probability is what Dang & Lanjouw, (2014) name as an index of insecurity or vulnerability. Figure 4 shows the kernel density of the vulnerability index of the households that experienced a downward transition and those that stayed out of poverty. This curve reaches a density peak around risk values of 10% in the case of households that became poor, while it presents values close to zero in the case of Households that remained out of poverty, thus meaning that most of the vulnerable households are above this threshold. The results do not change significantly when, instead of considering the expenditures, the vulnerability indexes are estimated from the income poverty transitions (see Figure 5). Considering that, on average, households experiencing an unfavorable transition to poverty over the period 2004-2014 represent about 10% of non-poor households and that the peak density is around that value, it is plausible to fix the vulnerability threshold at values close to 10%. Table 9: Vulnerability Lines at Lima’s prices and real values 2014

Value of Vulnerability Line Proportion of the Poverty Line Unconditional Income in Income in Lines Income Expenditures value of Income Expenditures value of expenditures expenditures Threshold 5% 903.08 661.19 671.66 2.36 1.73 1.75 Threshold 10% 690.88 571.93 551.29 1.80 1.49 1.44 Threshold 15% 645.84 516.04 511.31 1.69 1.35 1.34 Threshold 20% 582.58 491.55 451.16 1.52 1.28 1.18 Threshold 25% 569.90 470.62 429.30 1.49 1.23 1.12 Note: Note: The poverty line considered for comparison was the non-extreme poverty line of Lima province of 2014. Source: ENAHO 2004-2014. Elaborated by authors. Table 9 presents the point estimation of the unconditional vulnerability lines calculated from the logit regression previously presented, the 10% threshold chosen and the lines calculated with other threshold values to consider their sensitivity. The change in the decomposition of the population given the lines to different values of the threshold are presented in Annex A.4. The values of the lines are calculated at Lima’s prices level of 2014 as well as a proportion of the Lima’s 2014 poverty line, both in term of income, expenditures and Income in value of expenditures. For this third estimation that uses income and expenditures, we estimate the equation of the determinants of vulnerability considering poverty transitions regarding the household’s income as a dependent variable, while in the second stage the value of the line is determined using the level of expenditure. That is, the line ends up being the value of per capita household expenditure for those households with a probability between 9% and 11% of falling into poverty by a low income level.

27

Figure 4: Vulnerability Index - Expenditures

Logit Expenditures 15

10

Non Poor - Non Poor Non Poor - Poor

Density

5

0 0 .2 .4 .6 .8 Estimated Probability

Source: ENAHO 2004-2014. Elaborated by authors.

Figure 5: Vulnerability Index - Income

Logit Income 15

10

Non Poor - Non poor Non Poor - Poor

Density

5

0 0 .2 .4 .6 .8 1 Estimated Probability

Source: ENAHO 2004-2014. Elaborated by authors.

28

4.4 Vulnerability and macroeconomic cycle In order to specify the potential macroeconomic effect in the vulnerability, we have included in the pooled regression a dichotomous variable for each of the panel years. As it is observed in Table 10, the probabilities of falling into poverty for each year show a downward trend compatible with an increase in the line necessary to consider the household as vulnerable. Considering poverty transitions by income or expenditure does not change this result. Table 10: Year Fixed Effects – Vulnerability Regression

Year Fixed Effects Odds Ratio Expenditures Income 2005 0.552*** 0.649*** (0.068) (0.085) 2007 0.512*** 0.543*** (0.060) (0.068) 2008 0.436*** 0.544*** (0.052) (0.068) 2009 0.421*** 0.436*** (0.049) (0.056) 2010 0.440*** 0.464*** (0.052) (0.059) 2011 0.430*** 0.396*** (0.050) (0.050) 2012 0.334*** 0.284*** (0.040) (0.037) 2013 0.362*** 0.411*** (0.043) (0.052)

Note: Significance is denoted by * = 10%; ** = 5%; *** = 1%. Odds ratio are presented. Source: ENAHO 2004-2014. Elaborated by authors. It is interesting to note that the sharp deceleration of growth in 2014 has traduced in an increase in the risk of falling into poverty for that year, reversing the downward trend as shown in Figure 6. Consequently, the vulnerability of households have grown because of this macroeconomic (covariant) effect affecting every households, regardless of their characteristics and the specific (idiosyncratic) adverse shocks they may have experienced. The sharp slowdown in the growth rhythm in 2008/2009 had only a slight impact on the vulnerability index, while the recent deceleration in the GDP growth over several consecutive periods has translated in a greater vulnerability for all households. This would seem to indicate that short-term unfavorable macroeconomic shocks would not have a significant impact on vulnerability; while lasting unfavorable macroeconomic contexts, such as the deceleration experienced by Peru, would imply a significant increase in the vulnerability of households to poverty.

29

Figure 6: Macroeconomic Effect in the Vulnerability of Household and GDP growth

70% 9.00%

8.00% 60% 7.00% 50% 6.00%

40% 5.00%

30% 4.00% 3.00% 20% 2.00% 10% 1.00%

0% 0.00% 2005 2007 2008 2009 2010 2011 2012 2013

Riesgo Caída Ingresos Riesgo Caída Gastos PBI (Var %)

Note: The fixed effects are plotted taking into account the left axis, while GDP growth is plotted according to the right secondary axis. Source: ENAHO 2004-2014 y MEF. Elaborated by authors.

4.5 Population’s decomposition and evolution using poverty and vulnerability lines The estimation of a poverty and a vulnerability line both in monetary terms allows us to divide the group of households into the following categories: Extreme poor, non-extreme poor, vulnerable non-poor and non-poor non-vulnerable. The sharp reduction in total poverty between 2004 and 2014 (from 55.6% to 22.7%) is composed by a reduction in 20 percentage points of non-extreme poverty and by a decrease of more than 11 points in extreme poverty. A question that arise from this considerable reduction in the poverty rates is if the households that emerged from poverty became vulnerable households with a high risk of falling back into poverty or if they escaped further from the poverty line becoming non-vulnerable non-poor households with low risk of falling back. Once the vulnerability index and line were estimated using the critical threshold of 10%, we decompose the population into four groups: extreme poor, poor, vulnerable non-poor and non-vulnerable. This decomposition has been calculated for both income and expenditure of households and for each year between 2004 and 2014 using all the observations of the ENAHO survey in its cross-sectional version. Once the vulnerability line has been estimated, this line is compared to the household expenditure or income levels to identify the vulnerable population. The vulnerability line represents in some way the "ceiling" above which it is very unlikely that a household will fall into poverty in the next period. The lower limit below which a household is no longer vulnerable to poverty is the monetary poverty line. This allows us to calculate the disaggregation for the whole population and for the entire period examined.

30

Considering income as a welfare indicator does not change the general conclusions since they show similar trends, although the size of the vulnerable population is slightly larger in the case of the income decomposition. From Figure 7, we note that, while poverty was significantly reduced, the proportion of vulnerable households has been growing, particularly since 2006. In 2004, the beginning of the expansionary cycle, vulnerable households accounted for 19.9% of the population, a proportion which was multiplied by a factor of 1.7. Considering reduction of 31.7 percentage points in the poverty rate between 2004 and 2014, a quarter of these households (24.6%) became part of the vulnerable population.

Figure 7: Vulnerability Evolution – Unconditional Expenditure Line 10%

100

90 21.4 22.5 26.4 31.4 34.2 35.0 38.3 80 40.1 42.4 43.4 43.9 70 19.9 22.0 60 24.5 26.2 50 28.5 31.5 30.9 32.1 40 31.8 32.7 33.4 42.3 39.8 30 35.4 31.2 26.5 20 24.0 23.1 21.5 19.8 19.2 18.4 10 16.4 15.8 13.8 11.3 10.9 9.5 7.6 6.3 6.0 0 4.7 4.3 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Extreme Poor Non-extreme Poor Vulnerable Non-Poor Non-vulnerable

Source: ENAHO 2004-2014. Elaborated by authors. The composition of the non-poor population changed drastically. If in 2004 the vulnerable represented 52.3% of the non-poor, this proportion rose to 68.8% by 2014. The slowdown in economic growth in recent years no longer led to a significant reduction in the incidence of poverty and weakened the population that emerged from poverty in the previous years. The expansion the social programs coverage, by targeting only the population living in poverty, has left the vulnerable population unattended and have not developed specific policies aimed at reducing the vulnerability (in particular unemployment insurance mechanisms, bankruptcy of family businesses, the consequences of natural disasters, access to formal credit, etc.). In addition to household decomposition over the decade under review, Figure 9 presents the symbolization of regional poverty and vulnerability levels by 2014. This outcome at the regional level can result in an extremely useful instrument for a geographical targeting for future public policies designed to reduce the fall into poverty of vulnerable households.

31

Figure 8: Vulnerability Evolution – Unconditional Income Line 10%

100

90 23.0 21.6 26.5 30.7 33.7 80 35.6 38.4 40.5 43.6 43.9 45.0 70 25.3 25.2 60 27.6 29.5 50 31.5 33.6 34.3 33.4 40 32.3 33.2 33.0 32.7 32.7 30 28.9 25.4 20.9 20 19.1 17.9 17.0 15.3 15.1 14.9 10 19.0 20.6 17.0 14.4 14.0 11.8 9.4 9.1 8.7 7.8 7.2 0 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Extreme Poor Non-extreme Poor Vulnerable Non-Poor Non-vulnerable

Source: ENAHO 2004-2014. Elaborated by authors. As we can see from the maps, the situation of poverty and vulnerability, whether by income or expenditure, is not directly or inversely correlated with clarity in the regional levels. In broad outline, we note that vulnerability shows considerably lower levels in the coastal regions, while the southern highlands concentrate several regions with a high proportion of vulnerable households. In this way, the coexistence of a strong incidence of both phenomena is evident, for example, in the regions of Loreto and Puno. However, other regions show an inverse correlation, where reduced levels of poverty are obtained at the end of the period of strong growth, although considerable levels of vulnerability remain for the households that escaped from deprivation. This would be the case of Ica, Tacna and Ucayali. Finally, we have another group of regions where poverty levels are still high in comparison to the national average so that a large proportion of households turn out to be chronic poor resulting in a difficult escape from the deprivation situation. In this set of regions characterized by their strong rurality are Amazonas, Apurimac, Ayacucho, Cajamarca, Cerro de Pasco and Huancavelica, which historically have suffered higher levels of deprivation.

32

Figure 9: Vulnerability and Poverty Maps by Regions

Source: ENAHO 2004-2014. Elaborated by authors.

33

5. Sensitivity Analysis In order to estimate the vulnerability lines chosen as the main estimate, several critical choices were required during the estimation process. Therefore, this section seeks to show the robustness of our results to the choice of (i) income vs expenditures, (ii) different vulnerability thresholds, (iii) unconditional vs conditional models, (iv) the regression technique, (v) pooled vs annual regressions, (vi) the model specification, and (vii) the caliper’s size. Due to the word count limitations, we include the sensitivity analysis in an online appendix available at: https://drive.google.com/file/d/0B3SAO4f5fpoNT3BzZzdLQmlKalk/view?usp=sharing 6. Conclusions The spectacular reduction of poverty during the last decade of rapid growth and its subsequent deceleration in the last four years has led to questioning whether this economic expansion has allowed the country to consolidate a middle class freed from the risk of falling into poverty or if those households that left the poverty situation would return to it with a less favorable macroeconomic context. In order to answer this question, it is necessary to observe the same households in consecutive periods in a representative sample, which has been possible thanks to the design of the ENAHO survey. All biannual panels were considered together during the period 2004-2014, which allowed for observations of poverty transitions for more than 50,000 households. The option of using biannual panels covering a long period (2004-2014) has also made it possible to specify the specific contribution of macroeconomic factors associated with changes in the vulnerability of the households. It has been shown that the vulnerability index has a behavior that could be described as anti-cyclical. The deceleration of growth, particularly in the last two years, has resulted in an increase in the vulnerability index, thus breaking the downward trend observed during the period of rapid and sustained growth. This means that, by slowing down macroeconomic growth, poverty reduction is slowed down while the vulnerability of households to poverty increases. Our results suggest that the vulnerability of Peruvian households to poverty is "structural" in its nature. It is mainly related to the characteristics of the labor market insertion (primary sector and informal microenterprises in urban areas) generating very high income instability, to the demographic structure of the household, the level of education, the ethnicity condition, and the geographic environment that defines the productive opportunities and reflects the density and the presence of the State. Adverse shocks, factors that have traditionally been considered as the distinguishing factor of vulnerable population, only have a significant impact in the case of major shocks (natural disasters) or when they occur cumulatively. Insurance and protection mechanism against adverse shocks play an important role in reducing the vulnerability of households, particularly those related to health shocks. Being covered by the ESSALUD or other insurance reduces by more than a half the risk of falling into poverty. Although the SIS insurance targeting criteria are focused on poor households, it also achieves a good targeting towards vulnerable households, especially those

34 households that present factors of structural vulnerability not necessarily linked to health shocks. This is probably due to this double causality that having SIS is associated to a greater vulnerability in the econometric estimations. Individual strategies based on the resources accumulated by households (savings), by mitigating or neutralizing the impact of an adverse shock, reduce vulnerability to poverty by almost a third (-28%); while recurring to debt appears as a factor that aggravates rather than attenuates the vulnerability of households. Not being part of any kind of association increases the vulnerability to poverty by limiting access to a potential network of solidarity and risk pooling. The results show that the vulnerable population represents about 30% of the total population and almost half of the non-poor population. Social policies must consider, in addition to the population living in poverty, the non-poor population but with a high risk of becoming poor that we have identified as vulnerable households. The insurance mechanisms against adverse shocks should be extended not only in the health domain but also in the domain of productive activities in order to mitigate the effects of the strong income instability in atomized markets with low entry costs. Access to formal credit could also reverse the negative role played by the cost and access conditions characteristic of informal credit. The great pending challenge is the implementation of insurance mechanisms for the vast majority of households whose income comes from informal employment and still does not have the coverage of health insurances.

35

References

Calvo, Cesar, and Stefan Dercon 2005 Measuring Individual Vulnerability. Economics Series Working Paper, 229. University of Oxford, Department of Economics. Vulnerability to Individual and Aggregate Poverty. Social Choice and Welfare 41(4): 721–740.

Cameron, Adrian Colin, and P. K. Trivedi 2009 Microeconometrics Using Stata. College Station, Tex: Stata Press.

Carter, Michael R, Munenobu Ikegami, and Chronic Poverty Research Centre 2007 Looking Forward: Theory-Based Measures of Chronic Poverty and Vulnerability. Great Britain: Chronic Poverty Research Centre.

Celidoni, Martina, and Isabella Procidano 2015 Identification Precision of Vulnerability to Poverty Indexes: Does Information Quantity Matter? Social Indicators Research 121(1): 93–113.

Chaudhuri, Shubham 2003 Assessing Vulnerability to Poverty: Concepts, Empirical Methods and Illustrative Examples. Working Paper. Columbia University.

Christiaensen, L. J., and Kalanidhi Subbarao 2005 Towards an Understanding of Household Vulnerability in Rural Kenya. Journal of African Economies 14(4): 520–558.

Cordero, L, J Herrera, and G Yamada 2003 Líneas de Pobreza Ajustadas Por La Necesidad de Salud: El Caso Peruano. Documento Para Discusión. Organización Panamericana de la Salud.

Cruces, Guillermo, Peter Lanjouw, Leonardo Lucchetti, et al. 2011 Intra-Generational Mobility and Repeated Cross-Sections: A Three-Country Validation Exercise. Policy Research Working Papers. The World Bank.

Dang, Hai-Anh H., and Peter F. Lanjouw 2013 Measuring Poverty Dynamics with Synthetic Panels Based on Cross-Sections. Policy Research Working Paper Series, 6504. The World Bank. 2014 Welfare Dynamics Measurement. Policy Research Working Paper, 6944. World Bank. 2015 Toward a New Definition of Shared Prosperity: A Dynamic Perspective from Three Countries. Policy Research Working Papers. The World Bank.

36

Dang, Hai-Anh, Peter Lanjouw, Jill Luoto, and David McKenzie 2011 Using Repeated Cross-Sections to Explore Movements into and Out of Poverty. Policy Research Working Papers. The World Bank. Dercon, Stefan, John Hoddinott, and Tassew Woldehanna 2005 Shocks and Consumption in 15 Ethiopian Villages, 1999--2004. Journal of African Economies 14(4): 559–585.

Fitzgerald, John, Peter Gottschalk, and Robert Moffitt 1998 An Analysis of Sample Attrition in Panel Data: The Michigan Panel Study of Income Dynamics. The Journal of Human Resources 33(2): 251–299.

Foster, James E., Indranil Dutta, and Ajit Mishra 2010 On Measuring Vulnerability to Poverty. Working Paper, 2010–13. The George Washington University, Institute for International Economic Policy.

Haq, Rashida 2012 Shocks as a Source of Vulnerability: An Empirical Investigation from Pakistan. Poverty and Social Dynamics Paper Series, 2012:06. Pakistan Institute of Development Economics.

Hoddinott, John, and Agnes Quisumbing 2003 Methods for Microeconometric Risk and Vulnerability Assessments. Social Protection and Labor Policy and Technical Notes, 29138. The World Bank.

Jha, Raghbendra, Katsushi S Imai, and Imai Gaiha Raghav 2009 Poverty, Undernutrition and Vulnerability in Rural India Public Works versus Food Subsidy. [Manchester]: Chronic poverty research centre (CPRC).

Ligon, Ethan, and Laura Schechter 2002 Measuring Vulnerability: The Director’s Cut. World Institute for Development Economics Research: United Nations University, World Institute for Development Economics Research.

Long, J. Scott, and Jeremy Freese 2001 Regression Models for Categorical Dependent Variables Using Stata. College Station, Tex: Stata Press.

López-Calva, Luis F., and Eduardo Ortiz-Juarez 2014 A Vulnerability Approach to the Definition of the Middle Class. The Journal of Economic Inequality 12(1): 23–47.

37

Morduch, Jonathan 1994 Poverty and Vulnerability. The American Economic Review 84(2): 221–225.

Morduch, Jonathan, and Gisele Gamanou 2002 Measuring Vulnerability to Poverty. Helsinki: United Nations University, World Institute for Development Economics Research.

Narayan, Deepa 2000 Voices of the Poor: Can Anyone Hear Us?, vol.1. World Bank Publications. Pritchett, Lant, Sudarno Sumarto, and Asep Suryahadi 2000 Quantifying Vulnerability to Poverty: A Proposed Measure, Applied to Indonesia. Policy Research Working Papers. The World Bank.

Ravallion, Martin 2016 The Economics of Poverty: History, Measurement, and Policy. Oxford University Press.

Skoufias, Emmanuel, and Agnes Quisumbing 2004 Consumption Insurance and Vulnerability to Poverty: A Synthesis of the Evidence from Bangladesh, Ethiopia, Mali, Mexico and Russia. Social Protection Discussion Paper Series, 401. World Bank.

Tesliuc, E, and K Lindert 2002 Vulnerability: A Quantitative and Qualitative Assessment. Guatemala Poverty Assessment Program. Washington, DC: World Bank.

Wagstaff, Adam, and Magnus Lindelow 2014 Are Health Shocks Different? Evidence from a Multishock Survey in Laos. Health Economics 23(6): 706–718.

World Bank 2013 World Development Report 2014: Risk and Opportunity - Managing Risk for Development. The World Bank. APPENDICES Due to the word count limitations, we include the appendices in an online appendix document available at: https://drive.google.com/file/d/0B3SAO4f5fpoNWmtUN2lNaHlVTUU/view?usp=sharing

38