Vulnerability and Livelihoods before and after the Haiti Earthquake

Transcription

1 Poliy Researh Working Paper 5850 WPS5850 Vulnerability and Livelihoods before and after the Haiti Earthquake Damien Éhevin Publi Dislosure Authorized Publi Dislosure Authorized Publi Dislosure Authorized Publi Dislosure Authorized The World Bank Latin Ameria and the Caribbean Region Soial Protetion Setor Otober 2011

2 Poliy Researh Working Paper 5850 Abstrat This paper examines the dynamis of poverty and vulnerability in Haiti using various data sets. As living onditions survey data are not omparable in this ountry, we first propose to use the three rounds of the Demographi Health Survey (DHS) available before the earthquake. Deomposing household assets hanges into age and ohort effets, we use repeated ross-setion data to identify and estimate the variane of shoks on assets and to simulate the probability of being poor in the future. Poverty and vulnerability profiles are drawn from these estimates. Seond, we deompose vulnerability to poverty into various soures using a unique survey onduted in 2007 in rural areas. Using two-level modelling of onsumption/inome, we assess the impat of both observable and unobservable idiosynrati and ovariate shoks on households eonomi well-being. Empirial findings show that idiosynrati shoks, in partiular health-related shoks, have larger impat on vulnerability to poverty than ovariate shoks. Third, asset-wealth is haraterized for households after the 2010 earthquake based on a survey designed to provide a rapid assessment of food inseurity in Haiti after the quake. Whereas it is not possible to onfirm the existene of poverty trap, it seems that those households who have lost the most due to the earthquake sueeded in reovering more rapidly from the shok, regardless of the effets of assistane, and probably more in line with oping strategies that are speifi to households. This paper is a produt of the Soial Protetion Setor, Latin Ameria and the Caribbean Region. It is part of a larger effort by the World Bank to provide open aess to its researh and make a ontribution to development poliy disussions around the world. Poliy Researh Working Papers are also posted on the Web at The author may be ontated at damien.ehevin@usherbrooke.a. The Poliy Researh Working Paper Series disseminates the findings of work in progress to enourage the exhange of ideas about development issues. An objetive of the series is to get the findings out quikly, even if the presentations are less than fully polished. The papers arry the names of the authors and should be ited aordingly. The findings, interpretations, and onlusions expressed in this paper are entirely those of the authors. They do not neessarily represent the views of the International Bank for Reonstrution and Development/World Bank and its affiliated organizations, or those of the Exeutive Diretors of the World Bank or the governments they represent. Produed by the Researh Support Team

3 Vulnerability and Livelihoods before and after the Haiti Earthquake Damien Éhevin 1 Keywords: Vulnerability; Poverty; Asset-Wealth; Earthquake; Haiti. JEL Codes: D12; D31; I32; O15. 1 Université de Sherbrooke. damien.ehevin@usherbrooke.a. I gratefully aknowledge Andrea Borgarello, Carlo del Ninno, Nany Gillespie, Franesa Lamanna, Philippe Leite, Ana Maria Oviedo and Ludovi Subran as well as partiipants at workshops in Port-au-Prine and Washington DC (World Bank) for their useful omments and suggestions. I also aknowledge Gary Mathieu from CNSA-Haiti for providing some of the data used in this paper. All errors and opinions expressed in this paper remain mine.

4 1. INTRODUCTION Examining hanges in poverty over time in Haiti poses severe hallenges. An issue ommon to many developing ountries is that survey data are not omparable. In Haiti, eah of the three expenditure or inome surveys olleted in reent years (1986, 1999, and 2001) has a very different design. As a onsequene, the analyses drawn on the basis of these surveys differ in the estimates of poverty inidene and trends (World Bank, 2006). The Demographi and Health Surveys (DHS), designed to be omparable, are of high quality but fail to inlude the expenditure or inome data generally used for poverty estimates. As reliable data are laking in order to trae poverty and vulnerability trends over time, disparate views on the part played by reforms in alleviating ex ante or ex post poverty may arise. Indeed, the basi question of what has happened poverty- and vulnerability-wise over the last deade is far from having satisfatorily been answered. Addressing this issue is a pre-requisite to improving our understanding of the underlying soial and eonomi proesses that have ontributed towards hanges in eonomi well-being in Haiti. Some nationally representative household inome and expenditure surveys have helped to provide a better understanding of living standards. In 1986, monetary poverty statistis (based on stated onsumption expenditure) showed that 59.6% of Haitians were under the poverty line (Pedersen and Lokwood, 2001). This situation only slightly improved in 1999, as 48.0% were then ategorized as poor. In 2001, the HLCS stated that 55.6% of households lived with less than US$1 per day and 76.7% with less than US$2 per day. This survey has not been onduted again sine then. In this paper, we explore different avenues in order to assess the dynamis of poverty in Haiti. First, we use the Demographi Health Surveys (DHS) to analyze the evolution of asset-poverty over time. We also propose a simple and intuitively appealing framework to assess vulnerability to asset-poverty with these data. Seond, we haraterize poverty and vulnerability in Haiti based on a unique survey onduted in 2007 in rural areas. Using two-level modeling of onsumption/inome, we assess the impat of both observable and unobservable idiosynrati and ovariate shoks on households eonomi well-being. Third, we use a post-earthquake survey designed to provide a rapid assessment of food inseurity in Haiti in order to assess the post-earthquake dynamis of asset-poverty. The paper is organized as follows. Setion 2 gives a bakground onerning risks, poverty and oping strategies in Haiti. Setion 3 examines the dynamis of poverty using pre-earthquake data. Setion 4 provides a haraterization of poverty and vulnerability in rural Haiti. In Setion 5, post-earthquake distribution of household asset-wealth is in diretly affeted areas. The last setion onludes. 2. BACKGROUND Like most developing ountries, Haiti faes insidious risks and shoks, inluding droughts, hurrianes, earthquake and eonomi and health shoks. The year 2008 proved partiularly arduous for Haitians, as they simultaneously had to fae a sharp rise in basi 2

5 food and fuel pries, exeptionally bad weather onditions and a major deline in international trade due to the global eonomi risis. On January 12th, 2010, a magnitude 7.0 earthquake struk Haiti. It was the most powerful in over 200 years, ausing thousands of Haitians to be killed, injured, homeless or displaed and infliting tremendous infrastrutural damage to the water and eletriity infrastruture, roads and ports systems in the apital, Port-au-Prine, and its surrounding areas. What is more, although the hurriane season was not partiularly destrutive in 2010, Haiti was struk by a holera epidemi in Otober. Until now, about 230,000 ases were reported, resulting in about 4,500 deaths. As of February 2011, about 3,000 patients per week were admitted for hospitalisation, as opposed to 10,000 at the November peak. USAID/OFDA believe that the disease will most likely be present in the ountry for the next years. Few months after the disaster, the human toll was extremely severe: 2.8 million people were affeted by the earthquake, ausing 222,570 deaths, and 300,572 injuries. 2,3 Over 97,000 houses were destroyed and over 188,000 were damaged. 661,000 people moved to non-affeted regions. Before the earthquake, poverty reahes very high levels in Haiti, with more than half of the population living in extreme poverty (i.e. with less than US$1 a day). Most of these approximately 4.5 million destitute lived in rural areas (about 70%) while the others lived in the metropolitan and other urban areas. Moreover, not only was extreme poverty widespread, but it was also severe. Inome was among the most unequally distributed in the world: aording to the 2001 Household Living Condition Survey, 20% of the poorest got 2% of total inome while 20% of the rihest got 68% of total inome. Multidimensional poverty was also far-reahing: soial indiators suh as literay, life expetany, infant mortality and hild malnutrition showed that poverty was allenompassing in Haiti. Around 4 out of 10 people ould not read and write, nearly half of the population had no aess to health are and more than four-fifths had no lean drinking water. 4 Aording to the 2009 national nutrition survey, hroni malnutrition (stunting) affeted from 18.1% (Port au Prine) to 31.7% (Plateau Central) of 6-59 month old hildren. Chroni malnutrition had to be linked with low aess to basi publi servies (health, eduation, running water, sanitation) and there was evidene that the extremely poor had muh less aess to servies than their non-poor ounterparts (World Bank, 2006). As a onsequene, the under-five mortality rate was twie the regional average and life expetany was about 18 years shorter than the regional average. Malnutrition also had to do with food inseurity in a ountry where food onsumption was the main type of expenditure for Haitian households, so that they stood defenseless when faed with prie flutuations. In 2000, food onsumption made up for 55.1% of the households' real onsumption (IHSI, 2001), with stark ontrasts between areas (64.2% in rural areas and 50.2% in urban ones). What is more, the food-dediated budget oeffiients were muh higher for poorer households and also remained fairly high for riher rural households 2 Soure: United Nation Offie for the Coordination of Humanitarian Affairs (OCHA). 3 Kolbe et al. (2010) estimated that 158,679 people in Port-au-Prine died during the quake or in the six-week period afterwards owing to injuries or illness. 4 Aording to the Household Living Conditions Survey (HLCS),

6 (about 50%). Among the fators fostering food inseurity, it should be noted that, on the one hand, a mere 10% of total onsumption in rural areas in ame from subsistene eonomy, and that, on the other hand, an average 52% of the ountry s food availability ame from imports: food imports urrently made up for a quarter of total imports while they only used to represent 18.3% in 1981, and the value of the per apita food imports had sharply inreased. Households being highly dependent on trade for food aess issues, they had beome highly exposed to prie hanges. Consequently, aording to the omprehensive food seurity and vulnerability analysis (CFSVA) 5 that was onduted before the sharp inflation inrease in 2007, 5.9% of rural households suffered from extreme levels of food inseurity while 19.1% of them were affeted to a lesser extent by food inseurity. 6 In total, 25% of these households were in a situation of food inseurity in Otober 2007, that is, just before the prie explosion in Haiti. In order to ope with poverty and food inseurity, households adopt various strategies: they diversify their inome soures, migrate or reeive international remittanes, adopt food restritions strategies, lend money or food, sell part of the household s assets, or renoune ostly ativities (eduation for hildren, et.). In Haiti, these strategies onern differently the poor and the rih: for instane, while remittanes reeived from international migrants represented about 18 perent of Haiti s GDP in 2007, 72% of the rihest households reeived emigrant remittanes, as ompared to only 39% for the poorest quintile. 7 On the other hand, food restrition strategies onerned 45% of poor rural households, who atually stated that they were used in utting on quantities. 8 Food restritions may indue early hildhood malnutrition, with permanent ognitive and psyhomotor onsequenes. Hene, malnutrition may indue diret produtivity loss due to bad physial onditions, indiret produtivity loss due to ognitive and eduation defiits, as well as loss due to inreasing health are osts. For this reason, malnutrition lowers eonomi growth and perpetuates poverty, from mother to hild (Alderman et al., 2002, Behrman et al., 2004). Other ut in expenditure suh as taking hildren out of shool an also have long-term effets on living standards. 3. DYNAMICS OF POVERTY BEFORE THE EARTHQUAKE 3.1.Data and Asset Index Various indiators of well-being are generally used to measure poverty suh as per apita household expenditures or per apita household inome. However, in developing ountries, good quality data on onsumption or inome prove to be hard to find in omparable surveys over time. Sahn and Stifel (2003) have listed several other problems in using household expenditures data suh as measurement errors due to reall data or due to the lak of information onerning pries and deflators. Alternative measures of 5 This study was a joint projet of the World Food Program (WFP) and the National Coordination of Food Seurity Unit (NCFSU). 6 CFSVA (2007). A sore was alulated for food inseurity from data related to diet diversity on the one hand (based on the number of types of food or food groups eaten during the week previous to the survey), and to their onsumption frequeny expressed in number of days during the period of referene on the other hand. 7 HLCS (2001). 8 CFSVA (2007). 4

7 household s well-being suh as the asset index should thus be onsidered. 9 Sahn and Stifel (2003) proposed to onsider three ategories of assets: household durables, housing quality and human apital. 10 The absene of omparable data soures on inome and expenditures over the last deade motivates our use of the Demographi and Health Surveys (DHS) 11 as an alternative instrument for assessing hanges in poverty and vulnerability, relying on an asset index as an alternative metri of welfare. The DHS are provided at three periods in Haiti: 1995, 2000 and It is then possible to ompare the assets over the three surveys. Among household assets, we first onsider liquid assets sine these assets an be sold to purhase basi ommodities in the event of a drop in inome. Seond, we onsider more durable assets suh as housing and eduation, whih an also be aumulated in order to protet households against poverty. Other intangible assets suh as household relations and soial apital may have been taken into aount in the analysis, but they are not available in the data. The asset index is a omposite indiator that is a linear ombination of ategorial variables obtained from a multiple orrespondene analysis: 12 a i K k 1 F 1 kdki, where a i is the value of the asset index for the ith observation, d ki is the value of the kth dummy variable (with k=1,,k) desribing the asset variables onsidered in the analysis (liquid assets as well as housing variables and eduation of the head of the household), and F 1 k is the value of the standardized fatorial sore oeffiient (or asset index weights) of the first omponent of the analysis. 13 Built this way, the asset index an be desribed as the 9 See, for instane, Sahn and Stifel (2000), Filmer and Prithett (2001), Sahn and Stifel (2003), Booysen et al. (2008). 10 This list of assets is not exhaustive and ould be ompleted following Moser (1998) s asset-based approah. In her asset vulnerability framework, Moser (1998) identifies several ategories of assets and illustrates how portfolio management affets vulnerability. Asset management inludes: labor (e.g., the number of earners in the family and their inome level), human apital (eduation and health), produtive assets (suh as housing in urban areas or attle in rural areas), household relations and soial apital. 11 The DHS surveyed households in Haiti s nine departments. These departments were divided into 9 urban and 9 rural strata plus the metropolitan area of Port-au-Prine, amounting to a total of 19 strata. A two-stage sampling proedure was employed to selet a representative sample of the target population. In the first stage, systemati sampling with probability proportional to the size of the strata was used to selet 317 ommunities as lusters or primary sampling units (PSUs). In the seond stage of sampling, households in eah of the PSUs were systematially sampled. 12 See Benzéri (1973) or, more reently, Asselin (2009). 13 Alternatively, Sahn and Stifel (2000) used fator analysis, and Filmer and Prithett (2001) used prinipal omponent analysis to measure their asset index. In referene to these methodologies, multiple orrespondene analysis an be viewed as a prinipal omponent analysis applied to a ontingeny table with the hi2-metri being used on the row/olumn profiles, instead of the usual Eulidean metri. Multiple orrespondene analysis provides information similar in nature to those produed by fator analysis and is less restritive than prinipal omponent analysis. 5

8 best regressed latent variable on the K asset primary indiators, sine no other explained variable is more informative (Asselin, 2009). Next, the methodology is developed in order to ompare distributions of the asset index over time. The data sets for several years are then pooled and asset weights are estimated using fator analysis for the pooled sample. We obtain: a i( t) K k1 F 1k d ki( t) where the fatorial sore oeffiients F 1 k are supposed to be onstant over time. Results from multiple orrespondene analysis for pooled data sets (Demographi and Health Surveys 1995, 2000 and 2005) are presented in Table 1. Several wealth items have been used: liquid assets (radio, television, refrigerator, biyle, motoryle, ar), housing harateristis (tap water, surfae water, flush toilet, no toilet, eletriity, rudimentary floor, finished floor) and head of household s eduation (no eduation, primary eduation, seondary eduation and tertiary eduation). Table 1. Asset index weights for pooled data Asset variables Weights % Inertia Liquid assets Radio Television Refrigerator Biyle Motoyle Car Housing Tap water Surfae water Flush toilet No toilet Eletriity Rudimentary floor Finished floor Head of household s eduation No eduation Primary eduation Seondary eduation Tertiary eduation Partial inertia 21.5 Soure: Own omputations using DHS 1995, 2000,

9 Weights have signs onsistent with interpretation of the first omponent as an assetpoverty index. Contribution of having no eduation appears to be partiularly high (17.7%). Having no toilet also ontributes in a large extent to inertia (19.7%). Having aess to surfae water ontributes to 21.5% of inertia. Other items ontribute to less than 10% of inertia. 3.2.Other Welfare Indies Inome Determination Other indies than the asset index an be used in order to approximate well-being. Firstly, eonomists generally onsider that total expenditure or inome should be favoured. However, in developing ountries, national surveys sometimes do not provide suh information on households. It is even more diffiult to get it on a regular basis. Let us start with a log linear model of inome determination: ln y i( t) t x' i( t) t t ei ( t) t y i t ) t where ( is the inome of household i(t) at time t, i ( t t is a vetor of explanatory variables and e i ( t ) t is an error term that is supposed to be independent and identially distributed. As proposed for instane by Stifel and Christiaensen (2007), it is possible to alulate ln yi( t k) t k x' i( t k) t k t k ei ( t k) t k, for all integers k, using estimates of ei ( t k ) t k and t k drawn from the estimated distributions of e i ( t ) t and t obtained from the previous equation. In doing so, we suppose that t k and t have the same distribution. This method is diretly inspired from poverty mapping methodology (f. Elbers et al., 2003). It is then possible to ompare several predited distributions of inome over time even if these distributions are not observed in eah time period. This is atually the ase when using, on the one hand, the Household Living Conditions Survey (HLCS), whih is the most reent national household survey, onduted in 2001 by the Haitian Statistial Offie (IHSI), and whih inludes modules on inome, health, eduation, and other household s assets; and, on the other hand, the Demographi Health Survey (DHS), a nationally representative household survey onduted every 5 years (1995, 2000, 2005) that provides data for a wide range of indiators in the areas of population, health, nutrition and other individual and household variables like assets and eduation. Finally, the ombination of ˆ, along with the available variables xi ( t k ) t k, yields : t k ln yˆ ˆ x ) i( t k) t k x' i t k t k t k eˆ ( ) i( t k) t k eˆ and i( t k ) t k 7

10 Based on this model, we will use a simple way of prediting ln yˆ by using i( t k ) tk x' i( t k) t k ˆ t. However, we should reognize that this short ut of the model will result in an underestimate of the variane of the distribution of the predited value of inome. 14 Health and Nutrition Index Seondly, Sahn and Stifel (2002) suggest using a height-for-age z-sore (HAZsore) in order to approah well-being. This sore an be stated as follows: Hi H HAZ sorei H median where H i is height for hild i, H median is the median height for a healthy and wellnourished hild from the referene population of the same age and gender and H is the standard deviation from the mean of the referene population. By onvention, a hild whose HAZ-sore falls below -2 is lassified as malnourished (stunting). Note that in the health and nutrition literature the HAZ-sore is generally onsidered as a reliable indiator of hroni malnutrition. This sore in Haiti is relatively high, with about one hild under 5 years old out of four being onerned by stunting or hroni malnutrition. To go one step further, in order to determine the health and nutritional status of hildren, we onsider a health prodution funtion: where x it is onsumption, hit h( xit, Zit, Cit, uit ) Z it is a vetor of household and individual harateristis, C it is u is unobserved heterogeneity. To apply a vetor of ommunity-level harateristis, and it this model empirially, we use the HAZ-sore for h it and, in the absene of data onerning onsumption, we will use predited inome or asset index as proxies for x it. Note that the ontinuous index h it an also be onsidered as a latent variable, sine we ould lass the hildren into two groups: one group whose HAZ-sore is below -2 and one whose HAZ-sore is above -2, with -2 being the malnutrition poverty threshold. 14 Note that one important drawbak of the methodology onerns the alulation of standard errors of the estimates (Tarozzi and Deaton, 2009). Indeed, although the methodology has been forefully advoated and onsiderably enhaned by Elbers et al. (2003), it is still ritiized, in partiular beause it relies on assumptions that are virtually untestable. This approah has for instane been used by the World Bank (2006) to ompare welfare over time in Haiti. The estimates show a small deline in extreme poverty over time nationally, from 60% in 1986 to 54% in Estimates based on the US$2-a-day poverty line show trends broadly similar to those for US$1-a-day poverty rates. The US$2-a-day headount estimates show a deline from 84% to 78%. However, given the large margin of error in the estimates, the hange has not been proved to be statistially signifiant. 8

11 Sensitivity Validation of the Asset Index We examine to what extent the asset index overlaps with other indies, i.e. the extent to whih one ats as an indiator for the other. One possible way of examining this is to define a poverty threshold for predited inome and one for assets. The proportion of people lassified as poor under both thresholds an then be examined and ompared with those lassified as poor under only one threshold and with those not lassified as poor under either threshold. However, the results yielded may be sensitive to the threshold that was seleted. Alternatively, the reeiver operating harateristi (ROC) urve provides a useful proedure for this omparison. It is arguable that the area under the ROC urve gives a more intuitive summary of the extent to whih two dimensions of welfare are orrelated in the sense of identifying the poor. Figure 1 suggests that asset-based poverty is a good indiator of inome-poverty (when using predited inome as a proxy for well-being). With an area below the ROC urve of around 0.85, this suggests that targeting low-asset households is going to alleviate muh of (though not all) poverty as measured with the predited inome, and vie-versa. The ROC urve methodology states as follows. Let us onsider inome-poor households that are below a ertain threshold (that is, US$2 when onsidering poverty and US$1 when onsidering extreme poverty). If the asset index assigns someone as poor who is also poor under the inome-poverty definition then this is alled a true positive (TP), also alled sensitivity. If it signals as poor someone who is not poor under the inome definition, it is a false positive (FP), also alled (1 speifiity), whih is also known as a type I error (i.e. poor people lassified as non-poor). If it signals someone as non-poor even though this person is poor under the inome definition, it is a false negative (FN). Finally true negatives (TN) are those who are lassified as non-poor under both definitions. Figure 1. Asset-based poverty and predited inome-poverty Speifiity Area under ROC urve = Soure: Own omputations using HLCS 2001 and DHS 2005 Table 2 summarizes the results together with Spearman rank orrelations between HAZ-sore and alternative measures of well-being. As for the area under the ROC urve, it is diffiult to settle, from this analysis, on whih of these two indies is the best preditor 9

12 for the health and nutrition welfare index. Indeed, they seem to have omparable power in targeting hronially malnourished hildren. Table 2. Correlations between HAZ-sore and alternative measures of well-being Predited inome Asset index Area under ROC urve Spearman rank orrelation Soure: Own omputations using HLCS 2001 and DHS 2000 (2005 in bold) In a last analysis of orrespondene between welfare indies, we use the methodology proposed by Sahn and Stifel (2003). In order to assess the explanatory power of the asset index and the predited inome in prediting well-being, separate models of health and nutritional status are estimated onditioned on (i) the log of predited per apita household inome, (ii) the log of household asset index, (iii) both the log of predited per apital household inome and the log of household asset index. The probit regression model is fitted using an indiator whose value is one when the hild is malnourished (HAZ-sore under -2) and zero otherwise. One the models are run, we use them to predit hild HAZsores and ompare the rank orrelations and ROC urves between the fitted nutritional outomes and the atual measured outomes. Table 3. Probit estimates Predited inome Asset index HAZ-sore Predited inome & Asset index (iii) only (i) only (ii) Elastiity z-statisti Pseudo R Soure: Own omputations using HLCS 2001 and DHS 2000 (2005 in bold) Table 3 shows that the pseudo R-square is approximately the same for the model with asset index and for the model with predited inome: it is slightly higher in 2005 for the former. Looking at the measures of orrelations between atual and fitted values of the health and nutrition index in Table 4 shows that fitted values are better orrelated to atual values when asset index is used as a regressor. Using both indies in a regression does not signifiantly improve the orrelations. In onlusion, these findings suggest that analysts are not worse off, and may be better off, onditioning hild nutrition models on the asset index rather than predited inome in their effort to predit nutritional outomes and target programs. 10

13 Table 4. Correlations between atual and fitted HAZ-sore Fitted HAZ-sore with Predited inome Asset index Predited inome Atual HAZ-sore only (i) only (ii) and asset index (iii) Area under ROC urve Spearman rank orrelation Soure: Own omputations using HLCS 2001 and DHS 2000 (2005 in bold) 3.4.Evolution of Asset-Poverty over Time For the purpose of the temporal omparison of assets, all of the household asset indies used in the analysis are alulated on an individual basis by dividing indies by household size. In order to fator in asset-related eonomies of sale within the household, indies were also alulated on a per household basis and for assets divided by the square root of household size. Results did not qualitatively hange with the use of these different definitions of asset indies, so that they prove to be robust to the hoie of equivalent sales. Asset-based poverty headount (P0), poverty gap (P1), and poverty severity (P2) indies are presented in table 5 for various thresholds. The asset-based poverty lines are the 20 th, 40 th, 60 th and 80 th perentiles of the 1995 distribution of the asset index. In Table 5, results have been split aording to a rural/urban division so that the evolution of the index an be observed in different areas. It appears that asset-based poverty dereased between 1995 and 2000 and remained fairly onstant between 2000 and Moreover, it mainly delined in the North-East, Artibonite, Centre, and Grand Anse departments, espeially in rural areas. 15 These trends are omparable to those observed from hroni malnutrition rates, whih have also dereased over time from 32% in 1995 to 23% in 2000 and 24% in Table 5. Changes in asset-based poverty between 1995 and 2005 Poverty headount (P0) Poverty gap (P1) Poverty severity (P2) Perentile th National Urban Rural th National Urban Rural th National Urban Rural th National Urban Rural Soure: Own omputations using DHS 1995, 2000, These results are not reported here but are available upon request. 11

14 Aggregate hanges in asset-based poverty follow from the relative gains or losses of the poor and vulnerable within speifi setors or groups as opposed to population shifts between these groups (Ravallion and Huppi, 1991, Sahn and Stifel, 2000). A methodology of deomposition of the hange in asset-poverty an be stated as follows. Let us onsider P a poverty measure for two distributions at time t and, and two setors u (urban area) and r (rural area), so that: P t P ( P t u P u ) n u ( P t r P r ) n r r ju ( n t j n j ) P j r ju ( P t j P j )( n t j n ). j The first two omponents are the within omponents: they show how asset-poverty in eah of the residene areas (urban and rural) ontribute to the aggregate hange of assetpoverty between t and. The third omponent is the between omponent: it is the ontribution of hanges in the distribution of the population aross two groups. The final omponent is a residual omponent that is a measure of orrelation between population shifts and hanges in asset-poverty within the groups. Table 6 presents the deomposition of the hange of the asset-based index headount ratio between 1995 and This deomposition suggests that intra-rural effets aount for most of the hanges when the poverty line is hosen under the 80 th perentile. Migration explains about 25% of the hange and its ontribution to the hange generally delines when the poverty line gets higher. Finally, the ontribution of delining asset-poverty in urban areas is low for low poverty lines and reahes nearly half of the hange when the poverty line is fixed at the level of the th perentile. Table 6. Deomposition of hanges in asset-based poverty between 1995 and 2005 Headount Deomposition Poverty line (Within) (Between) (Interation) (perentile in 1995) Change Urban Rural Migration Crossed effet 20th th th th Soure: Own omputations using DHS 1995, 2005 Other deompositions of the hange in asset-poverty an be ahieved by splitting the population into different groups of households aording to eduation and gender of the head of household, and aording to the presene of hildren under 5 years old in the household (see Table 7). It appears that the no or primary eduation group aounts for most of the hange in asset-poverty, all the more so as lower asset-poverty line is hosen. The same statement an be made for households with male head or with under 5 hildren: households with these harateristis experiened a larger derease in asset-poverty between 1995 and As a result of this analysis, we should emphasize that households with higher probability of being poor may have experiened a sharper derease of assetpoverty over the last deade. This should thus be kept in mind when analyzing poverty in a more stati manner. 12

15 3.1.Measuring Vulnerability 16 Asset Based Approah There are several arguments in favour of an asset-based approah to vulnerability. Firstly, sine vulnerability is a dynami onept, we an onsider that onsumption-poverty or inome-poverty measurements, beause they are stati, are of limited use in apturing omplex external fators affeting the poor as well as their response to eonomi diffiulty (Moser, 1998). Seondly, owning assets redues the risk for households to fall into poverty as a result of maroeonomi volatility (de Ferranti et al., 2000). Hene, aumulating assets be they liquid or not (e.g., durable goods and housing), material or not (by fostering eduation, health, family and soial networks) helps people to insure themselves against falling into poverty and to ope with risks and shoks. Asset aumulation should thus be onsidered as a major fator in risk management. Nevertheless, though an asset index an be a good proxy for living standards in order to measure poverty 17, two problems arise when using household wealth as an indiator of well-being in order to measure vulnerability to poverty. On the one hand, if assets are used for onsumption-smoothing, then an asset-based approah overestimates vulnerability sine assets an flutuate whereas onsumption does not. On the other hand, if assets are not used to smooth onsumption, the approah would underestimate vulnerability. So, knowing whether an asset-based approah deviates from a more standard onsumption-based approah is mainly an empirial question. 18 Besides, we ould ask whether, in some irumstanes, an asset-based approah is not preferable when it omes to measuring vulnerability. Indeed, let us onsider the most interesting and realisti ase where produtive assets ontribute towards the inome generation proess and an also serve as buffer-stok in order to fae a non-antiipated drop in inome (Deaton, 1991, Carroll, 1992). Empirially though, many studies find little evidene supporting the buffer-stok hypothesis in developing ountries. 19 For instane, Deron (1998) shows that, given subsistene onstraints and agent heterogeneity, rih households will aumulate assets more quikly than poor ones who will pursue low-risk, low-return ativities. Interestingly enough, the evidene suggests that households with lower endowments are less likely to own attle and returns to their endowments are lower. So, in presene of imperfet markets for redit and insurane, few households are able to smooth their onsumption. What is more, when assets are mainly made up of produtive assets, selling these assets would indue a permanent loss in inome for the household who 16 See Ehevin (2010a) for a more omplete version of this setion and appliation to other ountries. 17 Sahn and Stifel (2003) show that an asset index obtained from a fator analysis on household assets using multipurpose surveys from several developing ountries is a valid preditor of hild health and nutrition and, thus, long term poverty. 18 Ehevin (2010a) provides suh empirial evidene using Ghana Living Standard Surveys. 19 See, among others, Rosenzweig and Wolpin (1993), Morduh (1995), Fafhamps et al. (1998), Kazianga and Udry (2006), and Hoddinott (2006). 13

16 ould then fall into a poverty trap. 20 For this reason, poor households will prefer to smooth their assets instead of smoothing their onsumption. 21 An asset-smoothing behaviour might be a desirable strategy for households to avoid falling into poverty traps. As pointed out by Zimmerman and Carter (2003) who build on Deron (1998) s approah by inorporating the role of endogenous asset prie risks, portfolio strategies an bifurate between rih and poor households. In this setting, poor agents respond to shoks by using onsumption to buffer assets when they get lose to a ritial asset threshold. 22 Eonometri Framework Let us quantify vulnerability to poverty by onsidering the probability to be poor in the future that is having predited future inome or assets below a pre-defined threshold, onditional on household harateristis and exogenous shoks. This probability an be stated as follows: vˆ it it1 it it 1 it1 Pr( a z x, x, a ), where a it 1 is household i welfare (using per apita asset index as a proxy) at time t+1, x it x are vetors of household harateristis at time t and t+1 respetively that are not and it 1 used in the definition of ohort, and z is a given threshold. This probability is modelled using pseudo panel data. Indeed, in the absene of panel data, repeated ross-setion data an be grouped together by age ohort, eduation, and geographi groups in order to implement the methodology. So, the welfare index an be modelled in logarithm as follows: 23 it it t it ln a x, where supersript denotes ohort group. It is assumed that the residual term it an be deomposed into an individual speifi effet i and an error term it as follows: it i it, 20 Zimmerman and Carter (2003) and Carter and Barrett (2006), among others, have analyzed the existene of poverty traps when households are involved in various asset aumulation dynamis. 21 Note that if households are able to diversify their portfolio of assets into risky and safe assets, then in presene of redit onstraints they will hoose to lower the proportion of risky assets held in order to smooth inome over time (Morduh, 1994). 22 The empirial evidene onerning the existene of suh asset-poverty traps and thresholds are mixed with some authors finding evidene of its existene: see, for instane, Lybbert et al. (2004), Adato et al. (2006), Barret et al. (2006) or Carter et al. (2007). Carter and May (1999, 2001) also provide evidene of poverty traps although they are differently theoretially grounded. 23 Bourguignon and Goh (2004) proposed a similar method for assessing vulnerability to poverty, although relying on earning dynamis. 14

17 where i an be modelled either as a fixed effet or as a random effet and to follow a martingale that is it is supposed it it it 1, with it denoting an innovation term that is supposed to be normally, independently and 2 identially distributed, with mean zero and variane t. Grouping households together by ohorts gives the possibility to estimate the model with repeated ross-setion surveys. Estimating this model by fousing on seond-order moments as in Deaton and Paxson 2 (1994) yields estimates of t 1 that an diretly be used to predit the degree of household vulnerability in ohort. Indeed, by first drawing a value ~ in the normal 2 distribution with mean zero and variane in t+1 for household i in ohort : vˆ it Pr( a it 1 it z x, x it 1, a it 1 ˆt 1 ln z x ) it 1, we obtain the probability to beome poor it 1 ˆ t 1 it ln a ˆ t 1 ˆ ~ x where (.) denotes the umulative density of the standard normal distribution. Assuming, for simpliity sake, that x ˆ it t x ˆ 1 1 itt gives ~ ln z ln a ˆ it it 1 2 v it Pr( ait 1 z xit, xit 1, ait 1), where ˆ t 1 is the estimator of the ˆ t 1 2 slope of the age profile for the asset disturbane term variane t. Indeed, we propose to deompose the residual variane into age and ohort effets as follows: it t it 1, 2 t t u at t, where is a onstant, t is a ohort effet, at is an age effet, and u t is an error term whih is supposed to be independent and identially distributed and of mean zero. Then, assuming that the ohort effet is time invariant as it should asymptotially be the ase (Verbeek, 2008), we estimate the first differene (from t to t+1) of age effets that is ˆ ˆ 2 for eah ohort in order to get. at 1 at ˆt1 Following the previous methodology, the estimation steps to obtain the vulnerability index an be summarized as follows: Step 1. Create a pseudo panel from repeated ross-setion surveys. The rationale for this is to hoose time-invariant harateristis to group households in eah survey into ohorts. 24 The number of ells onstituted equals the number of ohorts multiplied by the number of periods/surveys available for the analysis. Cell size 24 A ohort is typially defined by the year of birth, eduation level and loalization. 15

18 should be large enough in order to minimize the bias arising from using pseudo panel data and not genuine panel data. 25 Step 2. Estimate the residual variane of the logarithm of the asset index within eah ell of the pseudo panel orresponding to ohort at time t. Pratially speaking, we regress for eah ell at the household level the logarithm of the asset index on a set of variables (inluding gender dummy, age and age squared, eduation dummies, household size, number of hildren under 5 years old, urbanization dummy or loalisation dummies) and estimate the residuals. The residual variane over ohorts orresponds to the variane of the residuals of the previous regression. Step 3. Regress the residual variane on ohort dummies and a polynomial funtion of age. Then, draw the estimated age effets on a graph to obtain the age-profile of the residual variane. 26 Estimate the slope of this age-profile for eah ohort 2 whih represents the estimated variane of the shoks faed by household,. Step 4. Draw a value 2 ˆt 1 it 1 ˆt 1 ~ in the normal distribution with mean zero and variane within eah ohort and ombine it with the estimated oeffiients of the observable harateristis to predit the vulnerability index it xit 1 vˆ for eah household i at time t belonging to ohort. For that purpose, an be predited deterministially from x it by inrementing age or assuming that harateristis are time invariant. Creation of a Pseudo-Panel In order to have a look at the dynami of asset-poverty, we regroup households from the DHS into homogeneous ohorts: households whose heads have the same date of birth (we define five-year ohorts), the same level of eduation (no eduation, primary and seondary and more) and the same plae of residene (ten departments and urban/rural distintion) are regrouped into ells. After regrouping some low-sized ells, 261 ells were onstituted for eah year of the DHS dataset, with an average size of around 150 households and 950 individuals in eah ell. Aggregate Estimates Our estimates of the vulnerability index follow the different steps realled previously. First, log per apita asset index is regressed on various household s harateristis suh as log of household size, age of the head and its square, eduation and gender of the head, loation and the presene of hildren under 5 years old. Residuals are 25 As exposed by Verbeek and Nman (1992), the bias in the standard within estimator based on pseudo panel data is dereasing with the number of individuals in eah ell, more than with the number of ells. However, Verbeek (2008) notes that there is no general rule to judge whether ell size is large enough. Deaton (1985) also suggests that measurement error dereases as a funtion of the size of the ells. 26 As in Deaton and Paxson (1994), we an normalize so that the fitted age effet at, for instane, age equals the average residual variane of the logarithm of the asset index for year-olds over all ohorts. 16

19 estimated from these regressions. Seond, we alulate for eah ell the variane of the residuals of the first-stage household-level regression. Third, we regress the residual variane on ohort dummies (reated by rossing household head date of birth, eduation and loation dummies) and a polynomial funtion of age (generally of two degrees or more if statistially signifiant). From the age profile of the residual variane, we alulate the slope whih is an estimate of the variane of asset shoks. Note that this slope should neessary be positive (i.e. the amplitude of shoks grows with age) sine the estimated variane should always be positive. This is generally the ase. However, when it is not, ontiguous ells have been regrouped for the estimates. Finally, one the variane of shoks is estimated for eah ohort then the last estimation step onsists in drawing values of shoks within the standard normal distribution and estimating the household vulnerability index using oeffiient estimates. Poverty and vulnerability rates are reported in Table 8 where a household is onsidered as poor when its asset index is below the 80 th perentile of the 1995 distribution of asset index. An extremely poor household is one whose asset index is below the 40 th perentile of the 1995 distribution of asset index. A household is onsidered as vulnerable if the probability to be poor or extremely poor is higher than 0.5, whereas it is onsidered as highly vulnerable for a probability higher than 0.8. At the national level, poverty headount (71.5%) is not different from the estimated fration of the population who is vulnerable to poverty. Moreover, 34.5% of the population is extremely poor, while is 34.7% vulnerable to extreme poverty. Whatever the threshold indeed, poverty and vulnerability do not appear very different from eah other. If we look at non-poor people, 10.6% are vulnerable to poverty. Among the population that is vulnerable to poverty, 95.8% are poor, and among the vulnerable to extreme poverty, 85.7% are estimated to be extremely poor. 3.2.Poverty and Vulnerability Profiles Aording to previous results, the harateristis of those who are estimated to be poor should not be very different from the harateristis of those who are estimated to be vulnerable to poverty. Table 9 presents the distribution of poor and vulnerable groups aross various harateristis. We find lear similarities between the poor and the vulnerable. Indeed, poor and vulnerable groups are mostly rural. Relative to their share in the population (60.2%), rural households are over-represented among individuals who are poor (74.7%) or extremely poor (93.0%) and among those who are vulnerable to poverty (74.2%) or extreme poverty (91.9%). Other ategories are over-represented among the poor and vulnerable groups: 41.8% of individuals live in a household where the head has no eduation, while 53.6% among the poor (77.3% among the extreme poor) and 53.5% among the vulnerable to poverty (74.7% among the vulnerable to extreme poverty). There are more malnourished hildren under 5 years old among the extremely poor (35.3%) or vulnerable to extreme poverty (34.9%) than in the whole population (23.2%). There are also less 5-11 year old hildren who attend shool among the extremely poor (71.7%) or vulnerable to extreme poverty (72.6%) than in the whole population (83.4%). Interestingly, there are more latating or pregnant women among the extremely poor (47.3%) or vulnerable to extreme poverty (45.9%) than in the whole population (34.6%). 17

20 When looking at the harateristis of the ommunity, we find important disrepanies, sine there are fewer extremely poor or vulnerable to extreme poverty with aess to basi servies like primary shool (respetively 80.1% and 81.3% have aess against 89.7% in the whole population), first yle seondary shool (respetively 13.4% and 14.9% have aess against 39.1% in the whole population), seond yle seondary shool (respetively 4.7% and 5.8% have aess against 31.0% in the whole population), the market (respetively 12.7% and 14.8% have aess against 40.8% in the whole population), hospitals (respetively 1.9% and 2.7% have aess against 14.8% in the whole population), health entres (respetively 9.1% and 10.8% have aess against 29.8% in the whole population), drugstores (respetively 21.7% and 22.9% have aess against 37.2% in the whole population) and dotors' offies (respetively 2.2% and 3.5% have aess against 28.6% in the whole population). Overall, we note that vulnerable people have aess to basi servies relatively more often than the poor, sine some of them are atually non poor. 18

21 Table 7. Deomposition of hanges in asset-based poverty between 1995 and 2005 Headount Deomposition aording to eduation groups Deomposition aording to gender groups Deomposition aording to hildren groups Poverty line (Within) (Between) (Interation) (Within) (Between) (Interation) (Within) (Between) (Interation) (per. in 1995) Change No or primary Seondary or more Crossed effet Female head Male head Crossed Effet Without under 5 With under 5 Crossed effet 20th th th th Soure: Own omputations using DHS 1995, 2005 % Number of individuals ('00,000) Table 8. Asset-poverty and vulnerability to asset-poverty Number of households ('00,000) Fration poor Mean vulnerability to poverty Fration vulnerable to poverty Fration highly vulnerable to poverty Fration extremely poor Mean vulnerability to extreme poverty Fration vulnerable to extreme poverty Fration highly vulnerable to extreme poverty Overall Non poor Poor Extremely poor Non vulnerable to poverty Vulnerable to poverty Vulnerable to extreme poverty Highly vulnerable to poverty Highly vulnerable to extreme poverty Soure: Own omputations using DHS 19

22 Table 9. Poor and vulnerable groups Overall Non poor Poor Extremely poor Vulnerable to poverty Vulnerable to extreme poverty Highly vulnerable to poverty Highly vulnerable to extreme poverty N % N % N % N % N % N % N % N % ('00,000) Overall Household and individual harateristis (from DHS 2005) Region West Southeast North Northeast Artibonite Center South Grand-Anse Northwest Port-au-Prine Area of residene Urbain Rural Head of households Male Female Eduation of head of household No eduation Primary Seondary and above years old Malnutrition (stunting) Mortality

23 0-2 years old Malnutrition (stunting) Mortality years old Attend shool Do not attend shool years old Attend shool Do not attend shool years old years old Female head over 50 years old over 60 years old Monoparental Female Latating and pregnant women Community harateristis (from DHS 2000)* Primary shool With Without First yle seondary shool With Without Seond yle seondary shool With Without Market With Without Hospital With Without Health Center 21

24 Drugstore Dotor's offie With Without With Without With Without Soure: Own omputations using DHS. Note: *ommunity harateristis are available in 2000 not in 2005 in the DHS. Results are omputed using sample weights. 22

25 4. POVERTY AND VULNERABILITY IN RURAL HAITI 27 In order to fully haraterize the determinants of poverty and vulnerability in rural Haiti, a unique survey an be used to assess the impat of idiosynrati and ovariate shoks on eonomi well-being (suh as household onsumption or inome). This household survey on Haitian food seurity and vulnerability has been onduted in 2007 in rural areas. The number of households is around 3,000 distributed in 228 ommunities. It ontains quantitative information on household onsumption, prodution, inome and assets as well as a good deal of qualitative information on pereived shoks, oping strategies, response apaity and other risks. 4.1.Methodology Vulnerability to Shoks In this setion, we explore the relationships between onsumption or inome, on the one hand, and various idiosynrati and aggregate ovariates on the other hand. We suppose that households are imperfetly insured against shoks and have limited aess to redit. So, assuming uninsured exposure to risk, we an write: ln y X S S X, (1) j where y is the onsumption of household i in ommunity j, S is a vetor of observable shoks, harateristis, the error term. X is a vetor of household is j is a ommunity speifi effet and In the above equation, two parameters are of partiular interest. First, we should assess whether is signifiantly different from zero that is whether observable shoks have signifiant impat on eonomi well-being. Seond, in order to asertain whether observable shoks have different impats depending on household and ommunity harateristis, we should also assess whether is signifiantly different from zero. Community speifi effet j an be modelled either as a fixed effet or a random effet. In what follows, we will see how to model this unobservable omponent within a two-level linear random oeffiient model. Finally, we should take into aount the possibility that the error term an be orrelated with observable household harateristis and shoks so that parameters estimates might be biased. Following Datt and Hoogeveen (2003), equation (1) parameters estimates are used to measure the impat of the observable shoks on poverty. First, the ounterfatual welfare index ( y ) is derived from the differene between atual onsumption ( y ) and the estimated impat of observable shoks that is: 27 See Ehevin, D. (2010b) for a omplete version of this setion. 23

26 y y [exp(lnyˆ ) exp(lnyˆ S 0)]. (2) Seond, we measure the impat of shoks on poverty by a poverty gap (PG): * PG Pr(lny ln z) Pr(lny ln z). (3) This poverty gap will inform us about the extent to whih shoks affet poverty so that poliy should be implemented to redue the impat of shoks on soial welfare. Parameters estimates in equation (1) an be used as measures of vulnerability sine they inform us about oping mehanisms. However, we don t learn muh from these parameters about the variability of shoks, so we do not know their vulnerability inidene. Vulnerability as Expeted Poverty One step further, we an define vulnerability to poverty as the probability of falling into poverty when one s onsumption/inome falls below a predefined poverty line. Furthermore, households will be onsidered as vulnerable when the probability to be poor in the future is below a hosen vulnerability threshold. In order to estimate suh a probability, Chaudhuri et al. (2002) proposed to estimate the expeted mean and variane in onsumption using rosssetional data or short panel data. Let us define vulnerability for individual i in ommunity j by: vˆ Pr(lny ln z X ln z ln yˆ ) ˆ, (4) where (.) denotes the umulative density of the standard normal; z is the poverty line; ln ŷ is the expeted mean of log per apita onsumption and apita onsumption. 2 ˆ is the estimated variane of log per As in Christiaensen and Subbarao (2005), the onditional mean and variane ould be expressed from equation (1) as: y X X ES X E E ln, (6) 2 2 ln y X X V S X V. (7) One of the main strong points of Chaudhuri s approah resides in the fat that it is rather straightforward to implement on various types of datasets. One limitation of this approah when it is applied to a single ross-setion is that it annot take the temporal variability of parameters into aount. Moreover, vulnerability estimates using ross-setions usually prove to rely on partial observation of the loal ovariate and idiosynrati shoks experiened by households (f. j 24

27 Christiansen and Subbarao, 2005), whih implies omitted variables or reverse ausality biases. By taking into aount both observable and unobservable shoks our approah thus build on previous literature by providing a larger spetra of possible shoks endured by households. What is more, a two-level modelling approah will allow us to assess the impat of shoks at a ommunity level whih is an appropriate level to analyse risk-sharing behaviours (f. Suri, 2010). A Multilevel Deomposition Analysis Our methodologial approah is based on a two-level linear random oeffiient model where y is the onsumption of household i in ommunity j, x is a vetor of household ovariates (suh as households harateristis, self-reported shoks and their interations) and is a vetor of ommunity ovariates. One writes: w j ln y 0 j 1 j x j j w j j u 0 j w, 1 j,, (8) where the error term u reflets unobserved heterogeneity of household onsumption and the error terms 0 j and 1 j represent unobserved heterogeneity of onsumption aross ommunities. Given previous equations we get: ln y 10 11w j x 0 j jx u 00 01w j 1, (9) where the equation an be deomposed into a fixed part and a random part. For identifiation purposes, we assume that the ovariates x and w are exogenous with E x, w 0, E x, w 0 and Eu x w,, 0 1 j j, j 0 j 1 j j 0 j j. This model an be estimated using standard statistial software suh as Stata s gllamm ommand (Rabe-Hesketh and Skrondal, 2008). In ontrast with Günther and Harttgen (2009), we will both onsider observable and unobservable shoks as soures of vulnerability, whereas Günther and Harttgen do not onsider observable shoks in their analysis. Using this multilevel random oeffiient model, we an deompose the total onditional variane into two spatial levels: household and ommunity. So, using equation (9) and following Chaudhuri et al. (2002), we estimate the expeted unobservable idiosynrati variane, 2 2 ovariate variane ˆ and total variane ˆ of household onsumption using the estimated 0 j oeffiients from the following regressions: u 0 j ˆ 2 u u j ( u x 0 w, 0 0 j ) w x w, j 0 2 x 1 j 3 2 j j w x w. 3 j (10) 25

28 Using variane estimates from the above equations, we will provide measures of vulnerability aording to the different soures of vulnerability. First, we are onerned with vulnerability indued by strutural (or permanent) poverty, that is the fration of vulnerable households whose expeted mean onsumption ln ŷ is already below the poverty line ln z. Seond, we will measure vulnerability indued by risk, that is the fration of vulnerable households whose expeted mean onsumption ln ŷ lies above the poverty line ln z. As in Chaudhuri et al. (2002), a household is onsidered as vulnerable if the estimated vulnerability index is greater than the vulnerability threshold of Identifiation Issues A problem assoiated with the estimation of equations (1) and (9) is that idiosynrati and ovariate observable shoks are potentially endogenous for at least three reasons. First, sine the shoks are self-reported by the households in the questionnaire, it might be reported with errors. Hene, it is possible that households with a ertain level of onsumption or welfare onsider an event to be a shok, while others with a different level of onsumption or welfare may not. Seond, if onsumption levels influene the likelihood of exposure to the shok then reverse ausality may arise. For instane, health shok has not the same probability of ourring depending on household onsumption/inome level. This problem is most likely to happen with idiosynrati shoks. Community shoks are less likely to be influened by household onsumption or inome. Third, shoks may be orrelated to the error term beause of unobserved heterogeneity. Unobserved fators may indeed influene both exposure to shoks and onsumption/inome. For example, riher households may better irrigate their lands. If irrigation is not observed, the estimated impat of drought on onsumption delines may be upwardly biased. These soures of estimation bias are diffiult to take into aount with ross-setional data. However, as proposed in Datt and Hoogeveen (2003), a potential solution is the use of instrumental variables (IV) estimation. Instrumental variables are onstruted as ommunity means of shok variables leaving out the urrent household. These instruments are valid if households are more likely to report a shok when neighbours also report that shok, although neighbours affeted by the shok do not influene other household s eonomi well-being in another way than through the household s self-reported shok. 4.2.Data The vulnerability and food seurity survey was onduted in Haiti in Otober and November 2007 on approximately 3,000 households living in 228 rural ommunities. This survey has been realized by the National Coordination of Food Seurity Unit with the partnership of the World Food Program. A ommunity-related omponent was added to the household omponent of the survey, in onnetion with infrastrutures and aessibility to basi soial servies. So, this survey ontains quantitative information on household onsumption expenditures, prodution, inome and assets as well as a good deal of qualitative information on pereived shoks, oping strategies and other hazards. Our empirial study will thus try to assess vulnerability by using both sets of data quantitative and qualitative. 26

29 Prior to the 2010 earthquake, the rural population of Haiti represented about 60% of the total population. These households are partiularly vulnerable to natural shoks suh as droughts, floods and hurrianes. They also fae other risks and shoks suh as eonomi and health shoks, animal disease 28, rime and violene. When looking at the shoks faed by rural households in Haiti in Table 10, we find that many households fae ovariate shoks suh as: inrease in food pries, ylones, floods, droughts and irregular rainfall; many of those shoks have an impat upon inome or upon both inome and assets, and less often upon assets only. On the other hand, among the worst shoks delared by the households, most of them are idiosynrati shoks: they have to do with disease, asualties or death of a household member (for 42.5% of them) or animal diseases (14.0%); the worst ovariate shoks are ylones, floods, droughts and inrease in food pries whih onern around 26.3% of the households. Table 11 presents summary statistis for variables used in the analysis. Consumption and inome are expressed in Gourdes. The agriultural index is a omposite indiator whih is a linear ombination of ategorial variables obtained from a multiple orrespondene analysis (f. Asselin, 2009). Variables onsidered in the analysis are the number of lands, animals and agriultural materials owned by the household. The ommunity index is a linear ombination of ommunity basi infrastruture and aess to market variables (roads, aess to elementary or seondary shools, health entres, markets, eletriity and ell phone). A sore of inome diversity has also been built from the various inome soures earned by the household. As four main inome soures are delared by the household, the inome diversity variable (ID) is defined 1 4 as k 2 ID i 1 2 k s 1 i, where household i. This sore equals 0 when only one soure of inome is delared by the household. It averages 0.17 in the studied population. k s i is the share of the kth inome soure in total inome of As reported in Table 11, many heads of household are working in agriultural ativities (54%) and about one quarter of them have no job. Another important soure of inome is trade. Note also that about three quarters of households are land owners. 28 Haiti has had several ovariate shoks on animal and plant diseases in reent history. However, delaration of households onerned here their own animals. 27

30 % Table 10. Shoks in rural Haiti % affeted by this shok Inome only Assets only Both % reporting this shok as the worst shok* Inome only Assets only Both Inrease in food pries Cylone, Flood Drought Irregular rainfall Disease/Aident of a household member Animal diseases Crop diseases Rarity of basi foodstuffs on the market Inrease in seed pries Drop in relative agriultural pries Drop in wages Human epidemia Death of an household member Inrease in fertilizer pries Drop in demand Inseurity (theft, kidnapping) New household member Cessation of transfers from relatives/friends Loss of job or bankrupty Equipment, tool breakdown Others Soure: Own omputations using Haitian Vulnerability and Food Seurity Survey, Notes: The sum of the three olumns "inome only", "assets only" and "both" do not sum to 100% due to non response or don't know or no impat. *Do not sum to 100% due to non response or don't know. 28

31 Table 11. Desriptive statistis Mean SE Household variables Log of onsumption Log of inome Agriultural index Inome diversity Household size Number of hildren Age of head Male head Years of shooling (head) Ativity of head No job Agroalimentary Industry Constrution Trade Servies Other ativity Community variables Average years of shooling Land owners Community index Soure: Own omputations using Haitian Vulnerability and Food Seurity Survey, Results Regression Results We use self-reported shoks in order to estimate their impat on onsumption and inome. Table 12 presents OLS estimates and GLLAMM estimates. Both models are estimated with log onsumption and log inome. Our preferred speifiation regroups a large set of explanatory variables suh as household harateristis, regional dummies, ommunity harateristis, interation between household harateristis and ommunity harateristis, shoks variables, interation between shoks variables and household harateristis, interation between shoks variables and ommunity harateristis. Estimating the two-level linear random oeffiient model (GLLAMM) allows us to deompose the variane of the residuals into an idiosynrati variane and a ovariate variane. 29

32 Table 12. Regression results Consumption (in log) Inome (in log) OLS GLLAMM OLS GLLAMM Coeff. P-value Coeff. P-value Coeff. P-value Coeff. P-value Coeff. P-value Coeff. P-value Interept Household variables Agriultural index Number of adults over 50 years Number of adults years Number of adults years Number of hildren years Number of hildren 6-11 years Number of infants 3-5 years Number of infants 0-2 years Age of head Age of head 2 / Male head Years of shooling of head No job Inome diversity Land owner Region North West North North East Artibonite Centre West Grande'Anse Nippes South Southeast ref ref ref ref ref ref Community variables % Land owners Community index Average years of shooling Household * Community variables 30

33 Average years of shooling * Agriultural index Average years of shooling * Number of hildren % Land owner * Agriultural index % Land owner * Household size % Land owner * Age of head Community index * Agriultural index Shok variables Idiosynrati health shok Idiosynrati disease shok New household member Loss of inome Covariate limate shok Covariate health shok Covariate eonomi shok Inseurity shok Shok * Household variables Idiosynrati health shok * Nb adults 15 and more Idiosynrati health shok * Age of head Idiosynrati health shok * Years of shooling of head Idiosynrati health shok * Inome diversity Idiosynrati health shok * Land owner Idiosynrati disease shok * Nb adults 15 and more Idiosynrati disease shok * Age of head Idiosynrati disease shok * Years of shooling of head Idiosynrati disease shok * Inome diversity Idiosynrati disease shok * Land owner New household member * Nb adults 15 and more New household member * Age of head New household member * Years of shooling of head New household member * Inome diversity New household member * Land owner Loss of inome * Nb adults 15 and more Loss of inome * Age of head Loss of inome * Years of shooling of head Loss of inome * Inome diversity Loss of inome * Land owner Covariate limate shok * Nb adults 15 and more Covariate limate shok * Age of head Covariate limate shok * Years of shooling of head Covariate limate shok * Inome diversity

34 Covariate limate shok * Land owner Covariate health shok * Nb adults 15 and more Covariate health shok * Age of head Covariate health shok * Years of shooling of head Covariate health shok * Inome diversity Covariate health shok * Land owner Covariate eonomi shok * Nb adults 15 and more Covariate eonomi shok * Age of head Covariate eonomi shok * Years of shooling of head Covariate eonomi shok * Inome diversity Covariate eonomi shok * Land owner Inseurity shok * Nb adults 15 and more Inseurity shok * Age of head Inseurity shok * Years of shooling of head Inseurity shok * Inome diversity Inseurity shok * Land owner Shok * Community variables Idiosynrati health shok * % Land owners Idiosynrati health shok * Community index Idiosynrati disease shok * % Land owners Idiosynrati disease shok * Community index New household member * % Land owners New household member * Community index Loss of inome * % Land owners Loss of inome * Community index Covariate limate shok * % Land owners Covariate limate shok * Community index Covariate health shok * % Land owners Covariate health shok * Community index Covariate eonomi shok * % Land owners Covariate eonomi shok * Community index Inseurity shok * % Land owners Inseurity shok * Community index Idiosynrati variane Covariate variane Number of households Number of ommunities R2 or Pseudo-R Soure: Own omputations using Haitian Vulnerability and Food Seurity Survey,

35 In the regressions, shoks variables were regrouped into broad ategories: idiosynrati health shoks (disease/aident or death of a household member), idiosynrati disease shoks (animal and rop diseases), new household member, loss of inome (drop in wages, essation of transfers from relatives/friends, loss of job or bankrupty, equipment/tool breakdown), ovariate limate shoks (ylone, flood, drought and irregular rainfall), ovariate health shoks (human epidemia), ovariate eonomi shoks (inrease in food pries, rarity of basi foodstuffs on the market, inrease in seed pries, drop in relative agriultural pries, inrease in fertilizer pries, drop in demand), health shoks (human epidemia), inseurity shoks (theft, kidnapping). In Table 12, OLS estimates without shoks interating with harateristis shows the mean impat of shoks. Their impat is generally negative exept when hosting new household members (positive impat). In partiular, idiosynrati and ovariate health shoks have large and signifiant negative effets on both onsumption and inome. Regression results in Table 12 also help us haraterizing vulnerable groups by differentiating the impat of shoks on well-being aording to different household and ommunity harateristis. Idiosynrati health shoks. The negative impat of this shok is redued in absolute term when many households own lands in the ommunity. This may be due to the fat that idiosynrati health shok an be mutually insured within riher ommunities. Idiosynrati disease shoks. The signifiant positive parameter on the number of more than 15 years old household members shows that the idiosynrati disease shok onerning rops or animals signifiantly inreases the produtivity of adults who may have to ompensate for this kind of losses. In other words, the presene of a larger number of 15+ year old has a positive effet in reduing the impat from an animal/plant disease shok. Furthermore, idiosynrati disease shok signifiantly dereases the benefits of inome diversity for household eonomi well-being. New household member. On the one hand, the positive impat of aommodating a new member in the household is redued when the head of the household is higher eduated or for greater diversity of inome. The positive impat also dereases with the ommunity index (aess to basi infrastrutures). On the other hand, the household benefit more from a new member in ase of land ownership. Loss of inome. The negative impat of a loss of inome appears to be signifiantly redued when the head of the family is older. Covariate limate shoks. The negative impat of ovariate limate shok is signifiantly redued with inome diversity. The impat of this shok is further negative when many households own lands in the ommunity. Covariate health shoks. The negative impat of ovariate health shok is signifiantly redued when the household owns a land and when the head is older. 33

36 Covariate eonomi shoks. The negative impat of aggregate eonomi shok is signifiantly redued when the household owns a land. Inseurity shoks. The negative impat of inseurity shok is signifiantly inreased in absolute term when the household owns a land. The impat of this shok is signifiantly less negative when many households own lands in the ommunity. Simulation Results Previous estimates of equation (9) with GLLAMM are used to simulate the impat of self-reported idiosynrati and ovariate shoks on both poverty and vulnerability. Table 13 presents the results. We define two poverty thresholds: one is hosen so that 80% of the households are poor; another one is hosen so that 40% are onsidered as extremely poor. What is more, a household is onsidered as vulnerable if the estimated vulnerability index is greater than the vulnerability threshold of People are thus onsidered as vulnerable to poverty when they are more likely to fall into poverty in any period over two onseutive periods than to not be poor, that is (1 P) 2 0.5, where P is the probability to fall below the poverty line. So, previous ondition an be rewritten as P For simulation purposes, the poverty line is hosen so that 80% (resp. 40%) of households have expeted mean onsumption/inome below it. As a result, mean vulnerability appears to be respetively 63% and 46% for onsumption and 67% and 45% for inome. Using a vulnerability threshold of 0.29, vulnerability rates are respetively 98% and 87% for onsumption and 96% and 76% for inome. Simulations exerises first onsist in estimating the poverty rate and the vulnerability rate without observable idiosynrati shoks (olumn 2 in Table 13) or without ovariate shoks (olumn 3). As reported in Table 12, shoks whih have the largest impat on onsumption and inome are health shoks, be they household or ommunity shoks. So, most of the impat of observable shoks ould be attributed to these partiular shoks. On the ontrary, loss of inome has very little impat on poverty and vulnerability to poverty. Without observable idiosynrati shoks (olumn 2 in Table 13), the onsumptionpoverty rate falls to 28% and the onsumption-extreme poverty rate is estimated to be 6%. So, the poverty gap, as it is defined by equation (3), orresponds to 52 perentage points, whereas the extreme poverty gap is 34 perentage points. Without observable ovariate shoks, poverty dereases less: the poverty gap is 10 perentage points and the extreme poverty gap is 11 perentage points. We also simulate the impat of observable shoks on the vulnerability rates. This impat is twofold: observable shoks have an impat on the mean (as stated in equation (6)) as well as on the variane of onsumption/inome (as stated in equation (7)). On the one hand, the perentage of households with mean onsumption/inome below the poverty line is what we all poverty indued vulnerability. On the other hand, the perentage of 34

37 households with mean onsumption/inome above the poverty line that would fall into poverty due to onsumption/inome variability is what we all risk indued vulnerability. Simulations results of the impat of shoks on vulnerability are as follows. Firstly, the impat of observable idiosynrati shoks (in partiular, observable idiosynrati health shoks) on the vulnerability rate is large, whereas ovariate shoks have little impat on it. Without observable idiosynrati shoks, the rate of vulnerability to poverty (resp. to extreme poverty) is estimated to be 64% (resp. 28%), ompared to 98% (resp. 87%) with these shoks, whih represents a 34 perentage points (resp. 58 perentage points) fall. Without observable ovariate shoks, the rate of vulnerability to poverty (resp. extreme poverty) is estimated to be 95% (resp. 73%), ompared to 98% (resp. 87%) with these shoks, whih represents a 3 perentage points (resp. 14 perentage points) fall. We also have simulated the impat of idiosynrati and ovariate observable shoks on household inome. The results are very similar to previous ones (see Table 13). Seondly, Table 13 shows that observable idiosynrati and ovariate shoks have larger impat on the mean than on the variane of onsumption/inome. This is partiularly true when onsidering observable idiosynrati shoks. Indeed, the ratio between poverty indued and risk indued vulnerability that equals 4.42 with shoks is sharply dereased in the absene of observable shoks. This ratio is even lower without observable idiosynrati shoks (0.35) than without observable idiosynrati shoks (0.77). So, one possible interpretation of those results is that the main impat of shoks is to inrease poverty permanently rather than transitorily. Finally, one should estimate the impat of unobservable idiosynrati or ovariate shoks on vulnerability. By onstrution, unobservable shoks have no impat on mean onsumption or mean inome. However, they influene the variability of both onsumption and inome. So, we estimate vulnerability rates using either unobservable shoks or observable shoks as soures of onsumption/inome variability. Table 13 indiates that unobservable idiosynrati shoks have more influene on households vulnerability than unobservable ovariate shoks. Indeed, 96% of households are vulnerable to unobservable idiosynrati shoks (80% when onsidering vulnerability to extreme poverty), whereas they are 82% to be vulnerable to unobservable ovariate shoks (44% when onsidering vulnerability to extreme poverty). By ontrast, observable idiosynrati shoks have the same influene on households vulnerability than observable ovariate shoks. Indeed, the ratio of idiosynrati to ovariate vulnerability is 1.00 (1.04 when onsidering vulnerability to extreme poverty) for observable shoks, whereas it is 1.17 (respetively 1.83) for unobservable shoks. 35

38 Table 13. Vulnerability deomposition and simulations Consumption Inome Without Without Without Without Fatual idiosynrati ovariate Fatual idiosynrati ovariate (1) shoks (2) shoks (3) (1) shoks (2) shoks (3) Poverty rate* Mean vulnerability Vulnerability rate** Poverty indued vulnerability Risk indued vulnerability Poverty indued/risk indued vulnerability Idiosynrati vulnerability (unobserved) Covariate vulnerability (unobserved) Idiosynrati/Covariate vulnerability (unobserved) Idiosynrati vulnerability (observed) Covariate vulnerability (observed) Idiosynrati/Covariate vulnerability (unobserved) (Extreme) Poverty rate* Mean vulnerability Vulnerability rate** Poverty indued vulnerability Risk indued vulnerability Poverty indued/risk indued vulnerability Idiosynrati vulnerability (unobserved) Covariate vulnerability (unobserved) Idiosynrati/Covariate vulnerability (unobserved) Idiosynrati vulnerability (observed) Covariate vulnerability (observed) Idiosynrati/Covariate vulnerability (observed) Soure: Own omputations using Haitian Vulnerability and Food Seurity Survey, Notes: *The poverty line is hosen so that 80% (resp. 40%) of households have expeted mean onsumption below it. The poverty rate is the perentage of households whose expeted mean onsumption is below the poverty line. **The vulnerability threshold is 29%. 36

39 5. POST-EARTHQUAKE CHARACTERIZATION OF ASSET-WEALTH Data Soures and Methodology A post-earthquake food seurity-oriented survey was onduted in June 2010 by the CNSA in ollaboration with its main partners (ACF, FEWS-Net, Oxfam GB, FAO, UNICEF and WFP). The sampling used for the household survey is a probabilisti luster method, using two stages: (i) enumeration setions (geographial areas) and amps and (ii) households ensus data is used to selet the enumeration setions, with a probability proportional to population size. Eight households are then seleted randomly in eah setion. Camps are seleted using the International Organization for Migration (IOM) data; the number of amps seleted was proportional to the size of the ommunes. The sampling method yielded 1901 interviewed households, loated in the disaster areas (amp and nonamp sites) as well as in some non-diretly affeted areas. Geographi strata overed by the EFSA II survey are presented in Figure 2. To randomly selet households, different methods were used for the urban households, the rural households and the amps. For urban households, survey investigators observe and mark the loation of households on a street map that does not ontain soioeonomi infrastruture, and the households are randomly seleted. For rural households, previously mapped buildings are randomly seleted using enumeration setion maps, and households living in those buildings are interviewed; if there are no households inside, then the losest household is seleted. For amps, survey investigators start from the entre of the amp and walk towards the outside in a different randomly seleted paths. They number eah household enountered in the way, and randomly selet two households to interview. For all three types of sampling, when multiple households are found living in the same building or tent, a single household is randomly seleted Assets Based on the June survey, an asset index is alulated using a wider set of preearthquake dihotomous variables, namely some durable goods not delared in the February survey and aess to basi utilities. Table 14 reports both weights and ontributions to inertia. Weights have signs onsistent with interpretation of the first omponent as an asset-poverty index. In diretly affeted areas, ontribution to inertia of lighting appears to be partiularly high (26.7%). Water soure also ontributes in a large extent to inertia (18.9%). Having tools or material for fishery, agriultural prodution and handiraft ontributes to 12.2% of the inertia explained by the first omponent of the analysis. Other items ontribute to less than 10% of inertia eah. 29 This setion is an exerpt from Ehevin (2011). 37

40 Figure 2. Geographi strata overed by the EFSA II survey Soure: CNSA (2010b). 38