Vulnérabilité et moyens de subsistance avant et après le tremblement de terre en Haïti

Policy Research Working Paper 5850 Vulnerability and Livelihoods before and after the Haiti Earthquake Damien Échevin The World Bank Latin America and the Caribbean Region Social Protection Sector October 2011 WPS5850 Public Disclosure Authorized Public Disclosure Authorized Public Disclosure Authorized Public Disclosure Authorized Public Disclosure Authorized Public Disclosure Authorized Public Disclosure Authorized Public Disclosure Authorized Produced by the Research Support Team Abstract The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Policy Research Working Paper 5850 This paper examines the dynamics of poverty and vulnerability in Haiti using various data sets. As living conditions survey data are not comparable in this country, we first propose to use the three rounds of the Demographic Health Survey (DHS) available before the earthquake. Decomposing household assets changes into age and cohort effects, we use repeated cross-section data to identify and estimate the variance of shocks on assets and to simulate the probability of being poor in the future. Poverty and vulnerability profiles are drawn from these estimates. Second, we decompose vulnerability to poverty into various sources using a unique survey conducted in 2007 in rural areas. Using two-level modelling of consumption/income, we assess the impact This paper is a product of the Social Protection Sector, Latin America and the Caribbean Region. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at damien.echevin@usherbrooke.ca. of both observable and unobservable idiosyncratic and covariate shocks on households’ economic well-being. Empirical findings show that idiosyncratic shocks, in particular health-related shocks, have larger impact on vulnerability to poverty than covariate shocks. Third, asset-wealth is characterized for households after the 2010 earthquake based on a survey designed to provide a rapid assessment of food insecurity in Haiti after the quake. Whereas it is not possible to confirm the existence of poverty trap, it seems that those households who have lost the most due to the earthquake succeeded in recovering more rapidly from the shock, regardless of the effects of assistance, and probably more in line with coping strategies that are specific to households. Vulnerability and Livelihoods before and after the Haiti Earthquake Damien Échevin 1 Keywords: Vulnerability; Poverty; Asset-Wealth; Earthquake; Haiti. JEL Codes: D12; D31; I32; O15. 1 Université de Sherbrooke. Email: damien.echevin@usherbrooke.ca . I gratefully acknowle dge Andrea Borgarello, Carlo d el Ninno, Nancy Gillespie , Francesca Lamanna , Philippe Leite, Ana Maria Oviedo and Ludovic Subran as well as participants at workshops in Port - au - Prince and Washington DC (World Bank) for their useful comments and suggestions. I also acknowledge 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. 2 1. INTRODUCTION Examining changes in poverty over time in Haiti poses severe challenges. An issue common to many developing countries is that survey data are not comparable. In Haiti, each of the three expenditure or income surveys collected in recent years (1986, 1999, and 2001) has a very different design. As a consequence, the analyses drawn on the basis of these surveys differ in the estimates of poverty incidence and trends (World Bank, 2006). The Demographic and Health Surveys (DHS), designed to be comparable, are of high quality but fail to include the expenditure or income data generally used for poverty estimates. As reliable data are lacking in order to trace 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 basic question of what has happened poverty- and vulnerability-wise over the last decade is far from having satisfactorily been answered. Addressing this issue is a pre-requisite to improving our understanding of the underlying social and economic processes that have contributed towards changes in economic well-being in Haiti. Some nationally representative household income and expenditure surveys have helped to provide a better understanding of living standards. In 1986, monetary poverty statistics (based on stated consumption expenditure) showed that 59.6% of Haitians were under the poverty line (Pedersen and Lockwood, 2001). This situation only slightly improved in 1999, as 48.0% were then categorized as poor. In 2001, the HLCS stated that 55 .6 % of household s lived with less than US$1 per day and 76 .7% with less than US$2 per day. This survey has not be en conducted again since then. In this paper, we explore different avenues in order to assess the dynamics of poverty in Haiti. First, we use the Demographic 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 . Second, we characterize poverty and vulnerability in Haiti bas ed on a unique survey conducted in 2007 in rural areas. Using two - level modeling of consumption/income, we assess the impact of both observable and unobservable idiosyncratic and covariate shocks on households’ economic well - being. Third, we use a post - ear thquake survey designed to provide a rapid assessment of food insecurity in Haiti in order to assess the post - earthquake dynamics of asset - poverty . The paper is organized as follows. Section 2 gives a background concerning risks, poverty and coping strategies in Haiti. Section 3 examines the dynamics of poverty using pre-earthquake data. Section 4 provides a characterization of poverty and vulnerability in rural Haiti. In Section 5, post-earthquake distribution of household asset-wealth is in directly affected areas. The last section concludes. 2. BACKGROUND Like most developing countries, Haiti faces insidious risks and shocks, including droughts, hurricanes, earthquake and economic and health shocks. The year 2008 proved particularly arduous for Haitians, as they simultaneously had to face a sharp rise in basic 3 food and fuel prices, exceptionally bad weather conditions and a major decline in international trade due to the global economic crisis. On January 12th, 2010, a magnitude 7.0 earthquake struck Haiti. It was the most powerful in over 200 years, causing thousands of Haitians to be killed, injured, homeless or displaced and inflicting tremendous infrastructural damage to the water and electricity infrastructure, roads and ports systems in the capital, Port-au-Prince, and its surrounding areas. What is more, although the hurricane season was not particularly destructive in 2010, Haiti was struck by a cholera epidemic in October. Until now, about 230,000 cases 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 country for the next years. Few months after the disaster, the human toll was extremely severe: 2.8 million people were affected by the earthquake, causing 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-affected regions. Before the earthquake, poverty reaches 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. Income was among the most unequally distributed in the world: according to the 2001 Household Living Condition Survey, 20% of the poorest got 2% of total income while 20% of the richest got 68% of total income. Multidimensional poverty was also far-reaching: social indicators such as literacy, life expectancy, infant mortality and child malnutrition showed that poverty was all- encompassing in Haiti. Around 4 out of 10 people could not read and write, nearly half of the population had no access to health care and more than four-fifths had no clean drinking water. 4 According to the 2009 national nutrition survey, chronic malnutrition (stunting) affected from 18.1% (Port au Prince) to 31.7% (Plateau Central) of 6-59 month old children. Chronic malnutrition had to be linked with low access to basic public services (health, education, running water, sanitation) and there was evidence that the extremely poor had much less access to services than their non-poor counterparts (World Bank, 2006). As a consequence, the under-five mortality rate was twice the regional average and life expectancy was about 18 years shorter than the regional average. Malnutrition also had to do with food insecurity in a country where food consumption was the main type of expenditure for Haitian households, so that they stood defenseless when faced with price fluctuations. In 2000, food consumption made up for 55.1% of the households' real consumption (IHSI, 2001), with stark contrasts between areas (64.2% in rural areas and 50.2% in urban ones). What is more, the food-dedicated budget coefficients were much higher for poorer households and also remained fairly high for richer rural households 2 Source: United Nation Office for the Coordination of Humanitarian Affairs (OCHA). 3 Kolbe et al. (2010) estimated that 158,679 people in Port - au - Prince died during the quake or in the six - week period afterwards owing to injuries or illness. 4 According to the Household Living Conditions Survey (HLCS), 2001. 4 (about 50%). Among the factors fostering food insecurity, it should be noted that, on the one hand, a mere 10% of total consumption in rural areas in 1999-2000 came from subsistence economy, and that, on the other hand, an average 52% of the country’s food availability came from imports: food imports currently made up for a quarter of total imports while they only used to represent 18.3% in 1981, and the value of the per capita food imports had sharply increased. Households being highly dependent on trade for food access issues, they had become highly exposed to price changes. Consequently, according to the comprehensive food security and vulnerability analysis (CFSVA) 5 that was conducted before the sharp inflation increase in 2007, 5.9% of rural households suffered from extreme levels of food insecurity while 19.1% of them were affected to a lesser extent by food insecurity. 6 In total, 25% of these households were in a situation of food insecurity in October 2007, that is, just before the price explosion in Haiti. In order to cope with poverty and food insecurity, households adopt various strategies: they diversify their income sources, migrate or receive international remittances, adopt food restrictions strategies, lend money or food, sell part of the household’s assets, or renounce costly activities (education for children, etc.). In Haiti, these strategies concern differently the poor and the rich: for instance, while remittances received from international migrants represented about 18 percent of Haiti’s GDP in 2007, 72% of the richest households receive d emigrant remittances, as compared to only 39% for the poorest quintile . 7 On the other hand , food restriction strategies concern ed 45% of poor rural households , who actual ly stated that they we re used in cut ting on quantities . 8 Food restrictions may induce early childhood malnutrition , with permanent cognitive and psychomotor consequences. Hence , malnutrition may induce direct productivity loss due to bad physical condition s, indirect productivity loss due to cognitive and education deficits, as well as loss due to increasing health care costs. For this reason, malnutrition lowers economic growth and perpetuates poverty, from mother to child ( Alderman et al., 2002, Behrman e t al. , 2004). Other cut in expenditure such as taking children out of school can also have long - term effects on living standards. 3. DYNAMICS OF POVERTY BEFORE THE EARTHQUAKE 3.1.Data and Asset Index Various indicators of well-being are generally used to measure poverty such as per capita household expenditures or per capita household income. However, in developing countries, good quality data on consumption or income prove to be hard to find in comparable surveys over time. Sahn and Stifel (2003) have listed s everal other problems in using household expenditures data such as measurement errors due to recall data or due to the lack of information concerning prices and deflators. Alternative measures of 5 This study was a joint project of the World Food Program (WFP) and the National Coordination of Food Security Unit (NCFSU). 6 CFSVA (2007). A score was calculated for food insecurity 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 consumption frequency expressed in number of days during the period of reference on the other hand. 7 HLCS (2001). 8 CFSVA (2007). 5 household’s well -being such as the asset index should thus be considered. 9 Sahn and Stifel (2003) proposed to consider three cat egories of assets: household durables, housing quality and human capital. 10 The absence of comparable data sources on income and expenditures over the last decade motivates our use of the Demographic and Health Surveys (DHS) 11 as an alternative instrument for assessing changes in poverty and vulnerability, relying on an asset index as an alternative metric of welfare. The DHS are provided at three periods in Haiti: 1995, 2000 and 2005. It is then possible to compare the assets over the three surveys. Among household assets, we first consider liquid assets since these assets can be sold to purchase basic commodities in the event of a drop in income. Second, we consider more durable assets such as housing and education, which can also be accumulated in order to protect households against poverty. Other intangible assets such as household relations and social capital may have been taken into account in the analysis, but they are not available in the data. The asset index is a composite indicator that is a linear combination of categorical variables obtained from a multiple correspondence analysis: 12    K k ki k i d F a 1 1 , where i a is the value of the asset index for the i th observation, ki d is the value of the k th dummy variable (with k=1,…,K) describing the asset variables considered in the analysis (liquid assets as well as housing variables and education of the head of the household), and k F 1 is the value of the standardized factorial score coefficient (or asset index weights) of the first component of the analysis. 13 Built this way, the asset index can be described as the 9 See, for instance, Sahn and Stifel (2000), Filmer and Pritchett (2001 ), Sahn and Stifel (2003), Booysen et al. (2008). 10 This list of assets is not exhaustive and could be completed following Moser (1998)’s asset - based approach. In her asset vulnerability framework, Moser (1998) identifies several categories of assets and i llustrates how portfolio management affects vulnerability. Asset management includes: labor (e.g., the number of earners in the family and their income level), human capital (education and health), productive assets (such as housing in urban areas or cattl e in rural areas), household relations and social capital. 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 - Prince, amounting to a total of 19 st rata. A two - stage sampling procedure was employed to select a representative sample of the target population. In the first stage, systematic sampling with probability proportional to the size of the strata was used to select 317 communities as clusters or primary sampling units (PSUs). In the second stage of sampling, households in each of the PSUs were systematically sampled. 12 See Benzécri (1973) or, more recently, Asselin (2009). 13 Alternatively, Sahn and Stifel (2000) used factor analysis, and Filmer an d Pritchett (2001) used principal component analysis to measure their asset index. In reference to these methodologies, multiple correspondence analysis can be viewed as a principal component analysis applied to a contingency table with the chi2 - metric bei ng used on the row/column profiles, instead of the usual Euclidean metric. Multiple correspondence analysis provides information similar in nature to those produced by factor analysis and is less restrictive than principal component analysis. 6 best regressed latent variable on the K asset primary indicators, since no other explained variable is more informative (Asselin, 2009). Next, the methodology is developed in order to compare distributions of the asset index over time. The data sets for several years are then pooled and asset weights are estimated using factor analysis for the pooled sample. We obtain:    K k t ki k t i d F a 1 ) ( 1 ) ( where the factorial score coefficients k F 1 are supposed to be constant over time. Results from multiple correspondence analysis for pooled data sets (Demographic and Health Surveys 1995, 2000 and 2005) are presented in Table 1. Several wealth items have been used: liquid assets (radio, television, refrigerator, bicycle, motorcycle, car), housing characteristics (tap water, surface water, flush toilet, no toilet, electricity, rudimentary floor, finished floor) and head of household’s education (no education, primary education, secondary education and tertiary education). Table 1. Asset index weights for pooled data Asset variables Weights % Inertia Li quid assets Radio 0.310 2.1 Television 0.976 7.4 Refrigerator 1.146 4.5 Bicycle 0.462 1.4 Motocycle 0.807 0.6 Car 1.216 2.2 Housing Tap water 0.392 2.0 Surface water - 1.145 21.5 Flush toilet 1.150 2.6 No toilet - 1.076 19.7 Electricity 0.805 8.0 Rudimentary floor - 0.590 0.1 Finished floor 0.351 2.9 Head of household’s education No education - 0.912 17.7 Primary education - 0.005 0.0 Secondary education 0.938 6.0 Tertiary education 1.309 1.5 Partia l inertia 21.5 Source: Own computations using DHS 1995, 2000, 2005 7 Weights have signs consistent with interpretation of the first component as an asset- poverty index. Contribution of having no education appears to be particularly high (17.7%). Having no toilet also contributes in a large extent to inertia (19.7%). Having access to surface water contributes to 21.5% of inertia. Other items contribute to less than 10% of inertia. 3.2.Other Welfare Indices Income Determination Other indices than the asset index can be used in order to approximate well-being. Firstly, economists generally consider that total expenditure or income should be favoured. However, in developing countries, national surveys sometimes do not provide such information on households. It is even more difficult to get it on a regular basis. Let us start with a log linear model of income determination: t t i t t t i t t i e x y ) ( ) ( ) ( ' ln    where t t i y ) ( is the income of household i (t) at time t , t t i x ) ( is a vector of explanatory variables and t t i e ) ( is an error term that is supposed to be independent and identically distributed. As proposed for instance by Stifel and Christiaensen (2007), it is possible to calculate k t k t i k t k t k t i k t k t i e x y          ) ( ) ( ) ( ' ln  , for all integers k , using estimates of k t k t i e   ) ( and k t   drawn from the estimated distributions of t t i e ) ( and t  obtained from the previous equation. In doing so, we suppose that k t   and t  have the same distribution. This method is directly inspired from poverty mapping methodology ( cf. Elbers et al., 2003). It is then possible to compare several predicted distributions of income over time even if these distributions are not observed in each time period. This is actually the case when using, on the one hand, the Household Living Conditions Survey (HLCS), which is the most recent national household survey, conducted in 2001 by the Haitian Statistical Office (IHSI), and which includes modules on income, health, education, and other household ’s assets; and, on the other hand, the Demographic Health Survey (DHS) , a nationally representative household survey conducted every 5 years (1995, 2000, 2005) that provides data for a wide range of indicator s in the areas of population, health, nutrition and other individual and household variables like assets and education. Finally, the combination of k t k t i e   ) ( ˆ and k t   ˆ , along with the available variables k t k t i x   ) ( , yields : k t k t i k t k t k t i k t k t i e x y          ) ( ) ( ) ( ˆ ˆ ' ˆ ln  8 Based on this model, we will use a simple way of predicting k t k t i y   ) ( ˆ ln by using t k t k t i x  ˆ ' ) (   . However, we should recognize that this short cut of the model will result in an underestimate of the variance of the distribution of the predicted value of income. 14 Health and Nutrition Index Secondly, Sahn and Stifel (2002) suggest using a height-for-age z-score (HAZ- score) in order to approach well-being. This score can be stated as follows: H median i i H H score HAZ     where i H is height for child i , median H is the median height for a healthy and well- nourished child from the reference population of the same age and gender and H  is the standard deviation from the mean of the reference population. By convention, a child whose HAZ-score falls below -2 is classified as malnourished (stunting). Note that in the health and nutrition literature the HAZ-score is generally considered as a reliable indicator of chronic malnutrition. This score in Haiti is relatively high, with about one child under 5 years old out of four being concerned by stunting or chronic malnutrition. To go one step further, in order to determine the health and nutritional status of children, we consider a health production function: ) , , , ( it it it it it u C Z x h h  where it x is consumption, it Z is a vector of household and individual characteristics, it C is a vector of community-level characteristics, and it u is unobserved heterogeneity. To apply this model empirically, we use the HAZ-score for it h and, in the absence of data concerning consumption, we will use predicted income or asset index as proxies for it x . Note that the continuous index it h can also be considered as a latent variable, since we could class the children into two groups: one group whose HAZ-score is below -2 and one whose HAZ-score is above -2, with -2 being the malnutrition poverty threshold. 14 Note that o ne important drawback of the methodology concerns the calculation of standard errors of the estimates (Tarozzi and Deaton, 2009). Indeed, although the methodology has been forcefully advocated and considerably enhanced by Elbers et al. (2003), it is still criticized, in particular because it relies on assumptions that are virtually untestable. This approach has for instance been used by the World Bank (2006) to compare welfare over time in Haiti. The estimates show a small decline in extreme poverty over ti me nationally, from 60% in 1986 to 54% in 2001. 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 headcount estimates show a decline from 84% to 78%. However, given the large ma rgin of error in the estimates, the change has not been proved to be statistically significant. 9 3.3.Validation of the Asset Index We examine to what extent the asset index overlaps with other indices, i.e. the extent to which one acts as an indicator for the other. One possible way of examining this is to define a poverty threshold for predicted income and one for assets. The proportion of people classified as poor under both thresholds can then be examined and compared with those classified as poor under only one threshold and with those not classified as poor under either threshold. However, the results yielded may be sensitive to the threshold that was selected. Alternatively, the receiver operating characteristic (ROC) curve provides a useful procedure for this comparison. It is arguable that the area under the ROC curve gives a more intuitive summary of the extent to which two dimensions of welfare are correlated in the sense of identifying the poor. Figure 1 suggests that asset-based poverty is a good indicator of income-poverty (when using predicted income as a proxy for well-being). With an area below the ROC curve of around 0.85, this suggests that targeting low-asset households is going to alleviate much of (though not all) poverty as measured with the predicted income, and vice-versa. The ROC curve methodology states as follows. Let us consider income-poor households that are below a certain threshold (that is, US$2 when considering poverty and US$1 when considering extreme poverty). If the asset index assigns someone as poor who is also poor under the income- poverty definition then this is called a ―true positive‖ (TP), also called ―sensitivity . ‖ If it signals as poor someone who is not poor under the income definition, it is a ―false positive‖ (FP), also called ―(1 – specificity), ‖ which is also known as a type I error (i.e. poor people classified as non-poor). If it signals someone as non-poor even though this person is poor under the income definition, it is a ―false negative‖ (FN). Finally ―true negatives‖ (TN) are those who are classified as non-poor under both definitions. Figure 1. Asset-based poverty and predicted income-poverty 0.00 0.25 0.50 0.75 1.00 Sensitivity 0.00 0.25 0.50 0.75 1.00 1 - Specificity Area under ROC curve = 0.8535 Source: Own computations using HLCS 2001 and DHS 2005 Table 2 summarizes the results together with Spearman rank correlations between HAZ-score and alternative measures of well-being. As for the area under the ROC curve, it is difficult to settle, from this analysis, on which of these two indices is the best predictor 10 for the health and nutrition welfare index. Indeed, they seem to have comparable power in targeting chronically malnourished children. Table 2. Correlations between HAZ-score and alternative measures of well-being Predicted income Asset index Area under ROC curve 0.6087 0.6521 0.6002 0.6415 Spearman rank correlation 0.2133 0.2606 0 .1879 0.2230 Source: Own computations using HLCS 2001 and DHS 2000 (2005 in bold) In a last analysis of correspondence between welfare indices, we use the methodology proposed by Sahn and Stifel (2003). In order to assess the explanatory power of the asset index and the predicted income in predicting well-being, separate models of health and nutritional status are estimated conditioned on (i) the log of predicted per capita household income, (ii) the log of household asset index, (iii) both the log of predicted per capital household income and the log of household asset index. The probit regression model is fitted using an indicator whose value is one when the child is malnourished (HAZ-score under -2) and zero otherwise. Once the models are run, we use them to predict child HAZ- scores and compare the rank correlations and ROC curves between the fitted nutritional outcomes and the actual measured outcomes. Table 3. Probit estimates HAZ - score Predicted income o nly (i) Asset index only (ii) Predicted inc ome & A sset index (iii) Elasticity - 0.3304 - 0.3778 - 0.2895 - 0.3424 - 0.2158 - 0.2177 - 0.1879 - 0.2416 z - statistic - 8.38 - 6.37 - 8.96 - 7.53 - 5.03 - 3.50 - 5.01 - 4.61 Pseudo R2 0.0334 0.0539 0.0327 0.0552 0.0367 0.0598 Source: Own computations 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 predicted income: it is slightly higher in 2005 for the former. Looking at the measures of correlations between actual and fitted values of the health and nutrition index in Table 4 shows that fitted values are better correlated to actual values when asset index is used as a regressor. Using both indices in a regression does not significantly improve the correlations. In conclusion, these findings suggest that analysts are not worse off, and may be better off, conditioning child nutrition models on the asset index rather than predicted income in their effort to predict nutritional outcomes and target programs. 11 Table 4. Correlations between actual and fitted HAZ-score Fitted HAZ - score with Actual HAZ - score Predicted income only (i) Asset index only (ii) Predicted income and asset index (iii) Area under ROC curve 0.5978 0.5985 0 .6033 0.6627 0.6669 0.6730 Spearman rank correlation 0.1889 0.1861 0.1981 0.2551 0.2496 0.2652 Source: Own computations using HLCS 2001 and DHS 2000 (2005 in bold) 3.4.Evolution of Asset-Poverty over Time For the purpose of the temporal comparison of assets, all of the household asset indices used in the analysis are calculated on an individual basis by dividing indices by household size. In order to factor in asset-related economies of scale within the household, indices were also calculated on a per household basis and for assets divided by the square root of household size. Results did not qualitatively change with the use of these different definitions of asset indices, so that they prove to be robust to the choice of equivalent scales. Asset-based poverty headcount (P0), poverty gap (P1), and poverty severity (P2) indices are presented in table 5 for various thresholds. The asset-based poverty lines are the 20 th , 40 th , 60 th and 80 th percentiles of the 1995 distribution of the asset index. In Table 5, results have been split according to a rural/urban division so that the evolution of the index can be observed in different areas. It appears that asset-based poverty decreased between 1995 and 2000 and remained fairly constant between 2000 and 2005. Moreover, it mainly declined in the North- East, Artibonite, Centre, and Grand’Anse departments, especially in rural areas. 15 These trends are comparable to those observed from chronic malnutrition rates, which have also decreased over time from 32% in 1995 to 23% in 2000 and 24% in 2005. Table 5. Changes in asset-based poverty between 1995 and 2005 Poverty headcount (P0) Poverty gap (P1) Poverty severity (P2) Percentile 1995 2000 2005 1995 2000 2005 1995 2000 2005 20th National 0.20 0.15 0.17 0.139 0.111 0.125 0.1175 0.0966 0.1088 Urban 0.01 0.01 0.02 0.007 0.004 0.009 0.0052 0.0030 0.0066 Rural 0.31 0.23 0.28 0.214 0.172 0.201 0.1811 0.1499 0.1762 40th National 0.40 0.31 0.34 0.258 0.198 0.221 0.2068 0.1586 0.1797 Urban 0.06 0.02 0.06 0.026 0.011 0.028 0.0165 0.0072 0.0195 Rural 0.59 0.48 0.53 0.389 0.304 0.348 0.3145 0.2446 0.2854 60th National 0.60 0.52 0.51 0.378 0.305 0.323 0.2951 0.2311 0.2524 Urban 0.21 0.14 0.18 0.072 0.038 0.066 0.0399 0.0187 0 .0394 Rural 0.82 0.74 0.73 0.552 0.456 0.492 0.4395 0.3518 0.3930 80th National 0.80 0.75 0.71 0.520 0.450 0.450 0.4064 0.3367 0.3491 Urban 0.56 0.46 0.45 0.206 0.151 0.172 0.1082 0.0701 0.0952 Rural 0.94 0.91 0.89 0.698 0.620 0.6 34 0.5751 0.4882 0.5166 Source: Own computations using DHS 1995, 2000, 2005 15 These results are not reported here but are available upon request. 12 Aggregate changes in asset-based poverty follow from the relative gains or losses of the poor and vulnerable within specific sectors or groups as opposed to population shifts between these groups (Ravallion and Huppi, 1991, Sahn and Stifel, 2000). A methodology of decomposition of the change in asset-poverty can be stated as follows. Let us consider P a poverty measure for two distributions at time t and  , and two sectors u (urban area) and r (rural area), so that:               r u j j t j j t j r u j j j t j r r t r u u t u t n n P P P n n n P P n P P P P ) )( ( ) ( ) ( ) (          . The first two components are the within components: they show how asset-poverty in each of the residence areas (urban and rural) contribute to the aggregate change of asset- poverty between t and  . The third component is the between component: it is the contribution of changes in the distribution of the population across two groups. The final component is a residual component that is a measure of correlation between population shifts and changes in asset-poverty within the groups. Table 6 presents the decomposition of the change of the asset-based index headcount ratio between 1995 and 2005. This decomposition suggests that intra-rural effects account for most of the changes when the poverty line is chosen under the 80 th percentile. Migration explains about 25% of the change and its contribution to the change generally declines when the poverty line gets higher. Finally, the contribution of declining asset-poverty in urban areas is low for low poverty lines and reaches nearly half of the change when the poverty line is fixed at the level of the 1995 80 th percentile. Table 6. Decomposition of changes in asset-based poverty between 1995 and 2005 Headcount Decomposition Poverty line (Wit hin) (Between) (Interaction) (percentile in 1995) 1995 2005 Change Urban Rural Migration Crossed effect 20th 0.201 0.172 - 0.028 0.016 - 0.021 - 0.011 0.003 40th 0.400 0.339 - 0.062 - 0.002 - 0.043 - 0.019 0.002 60th 0.600 0.513 - 0.088 - 0.01 1 - 0.056 - 0.022 0.002 80th 0.800 0.711 - 0.089 - 0.041 - 0.032 - 0.014 - 0.002 Source: Own computations using DHS 1995, 2005 Other decompositions of the change in asset-poverty can be achieved by splitting the population into different groups of households according to education and gender of the head of household, and according to the presence of children under 5 years old in the household (see Table 7). It appears that the no or primary education group accounts for most of the change in asset-poverty, all the more so as lower asset-poverty line is chosen. The same statement can be made for households with male head or with under 5 children: households with these characteristics experienced a larger decrease in asset-poverty between 1995 and 2005. As a result of this analysis, we should emphasize that households with higher probability of being poor may have experienced a sharper decrease of asset- poverty over the last decade. This should thus be kept in mind when analyzing poverty in a more static manner. 13 3.1.Measuring Vulnerability 16 Asset Based Approach There are several arguments in favour of an asset-based approach to vulnerability. Firstly, since vulnerability is a dynamic concept, we can consider that consumption - poverty or income - poverty measure ments, because they are static, are of limited use in capturing complex external factors affecting the poor as well as their response to economic difficulty (Moser, 1998). Secondly, owning assets reduces the risk for households to fall into poverty as a re sult of macroeconomic volatility (de Ferranti et al., 2000). Hence , accumulating assets―be they liquid or not (e.g. , durable goods and housing), material or not (by fostering education, health, family and social networks)―helps people to insure themselves against falling into poverty and to cope with risks and shocks. Ass et accumulation should thus be considered as a major factor in risk management. Nevertheless, though an asset index can be a good proxy for living standards in order to measure poverty 17 , two problems a rise when using household wealth as an indicator of we ll - being in order to measure vulnerability to poverty . On the one hand, if assets are used for consumption - smoothing, then an asset - based approach overestimates vulnerability since assets can fluctuate whereas consumption does not. On the other hand, if as sets are not used to smooth consumption, the approach would underestimate vulnerability. So, knowing whether an asset - based approach deviates from a more standard consumption - based approach is mainly an empirical question. 18 Besides, we could ask whether, in some circumstances, an asset-based approach is not preferable when it comes to measuring vulnerability. Indeed, let us consider the most interesting and realistic case where productive assets contribute towards the income generation process and can also serve as buffer-stock in order to face a non-anticipated drop in income (Deaton, 1991, Carroll, 1992). Empirically though, many studies find little evidence supporting the buffer-stock hypothesis in developing countries. 19 For instance, Dercon (1998) shows that, given subsistence constraints and agent heterogeneity, rich households will accumulate assets more quickly than poor ones who will pursue low - risk, low - return activities. Interestingly enough, the evidence suggests that households with lower endowme nts are less likely to own cattle and returns to their endowments are lower. So, in presence of imperfect markets for credit and insurance, few households are able to smooth their consumption. What is more, when assets are mainly made up of productive asse ts, selling these assets would induce a permanent loss in income for the household who 16 See Echevin (2010a) for a more complete version of this section and application to other countries. 17 Sahn and Stifel (2003) show that an asset index obtained from a factor analysis on household assets using multipurpose surveys from several developing countries is a valid predictor of child health and nutrition and, thus, long term poverty . 18 Echevin (2010 a ) provides such empirical evidence using Ghana Living Standard Surveys. 19 See, among others, Rosenzweig and Wolpin (1993), Morduch (1995), Fafchamp s et al. (1998), Kazianga and Udry (2006), and Hoddinott (2006). 14 could then fall into a poverty trap. 20 For this reason, poor households will prefer to smooth their assets instead of smoothing their consumption. 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 Dercon (1998)’s approach by incorporating the role of endogenous asset price risks, portfolio strategies can bifurcate between rich and poor households. In this setting, poor agents respond to shocks by using consumption to buffer assets when they get close to a critical asset threshold. 22 Econometric Framework Let us quantify vulnerability to poverty by considering the probability to be poor in the future that is having predicted future income or assets below a pre-defined threshold, conditional on household characteristics and exogenous shocks. This probability can be stated as follows: ) , , | Pr( ˆ 1 1 1 c it c it c it c it c it a x x z a v      , where 1  it a is household i welfare (using per capita asset index as a proxy) at time t+1 , it x and 1  it x are vectors of household characteristics at time t and t+1 respectively that are not used in the definition of cohort c , and z is a given threshold. This probability is modelled using pseudo panel data. Indeed, in the absence of panel data, repeated cross-section data can be grouped together by age cohort, education, and geographic groups in order to implement the methodology. So, the welfare index can be modelled in logarithm as follows: 23 c it c t c it c it x a     ln , where superscript c denotes cohort group. It is assumed that the residual term c it  can be decomposed into an individual specific effect c i  and an error term c it  as follows: c it c i c it      , 20 Zimmerman and Carter (2003 ) and Carter and Barrett (2006), among others, have analyzed the existence of poverty traps when households are involved in various asset accumulation dynamics. 21 Note that if households are able to diversify their portfolio of assets into risky and safe a ssets, then in presence of credit constraints they will choose to lower the proportion of risky assets held in order to smooth income over time (Morduch, 1994). 22 The empirical evidence concerning the existence of such asset - poverty traps and thresholds ar e mixed with some authors finding evidence of its existence : see, for instance, Lybbert et al. (2004), Adato et al. (2006), Barret et al. (2006) or Carter et al. (2007). Carter and May (1999, 2001) also provide evidence of poverty traps although they are d ifferently theoretically grounded . 23 Bourguignon and Goh (2004) proposed a similar method for assessing vulnerability to poverty, although relying on earning dynamics. 15 where c i  can be modelled either as a fixed effect or as a random effect and c it  is supposed to follow a martingale that is c it c it c it       1 , with c it  denoting an innovation term that is supposed to be normally, independently and identically distributed , with mean zero and variance 2 ct   . Grouping households together by cohorts gives the possibility to estimate the model with repeated cross-section surveys. Estimating this model by focusing on second-order moments — as in Deaton and Paxson (1994) — yields estimates of 2 1  ct   that can directly be used to predict the degree of household vulnerability in cohort c . Indeed, by first drawing a value c it 1 ~   in the normal distribution with mean zero and variance 2 1 ˆ  ct   , we obtain the probability to become poor in t+1 for household i in cohort c :                        1 1 1 1 1 1 1 ˆ ~ ˆ ln ˆ ln ) , , | Pr( ˆ ct c it c t c it c it c t c it c it c it c it c it c it x a x z a x x z a v      , where (.)  denotes the cumulative density of the standard normal distribution. Assuming, for simplicity sake, that c t c it c t c it x x   ˆ ˆ 1 1    gives                    1 1 1 1 1 ˆ ~ ln ln ) , , | Pr( ˆ ct c it c it c it c it c it c it c it a z a x x z a v    , where 2 1 ˆ  ct   is the estimator of the slope of the age profile for the asset disturbance term variance 2 ct   . Indeed, we propose to decompose the residual variance into age and cohort effects as follows: ct at ct ct u          2 , where  is a constant, ct  is a cohort effect, at  is an age effect, and ct u is an error term which is supposed to be independent and identically distributed and of mean zero. Then, assuming that the cohort effect is time invariant as it should asymptotically be the case (Verbeek, 2008), we estimate the first difference (from t to t+1 ) of age effects ― that is at at   ˆ ˆ 1   ― for each cohort in order to get 2 1 ˆ  ct   . Following the previous methodology, the estimation steps to obtain the vulnerability index can be summarized as follows:  Step 1. Create a pseudo panel from repeated cross-section surveys. The rationale for this is to choose time-invariant characteristics to group households in each survey into cohorts. 24 The number of cells constituted equals the number of cohorts multiplied by the number of periods/surveys available for the analysis. Cell size 24 A cohort is typically defined by the year of birth, education level and localization. 16 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 variance of the logarithm of the asset index within each cell of the pseudo panel corresponding to cohort c at time t . Practically speaking, we regress for each cell at the household level the logarithm of the asset index on a set of variables (including gender dummy, age and age squared, education dummies, household size, number of children under 5 years old, urbanization dummy or localisation dummies) and estimate the residuals. The residual variance over cohorts corresponds to the variance of the residuals of the previous regression.  Step 3. Regress the residual variance on cohort dummies and a polynomial function of age. Then, draw the estimated age effects on a graph to obtain the age-profile of the residual variance. 26 Estimate the slope of this age-profile for each cohort c which represents the estimated variance of the shocks faced by household, 2 1 ˆ  ct   .  Step 4. Draw a value c it 1 ~   in the normal distribution with mean zero and variance 2 1 ˆ  ct   within each cohort c and combine it with the estimated coefficients of the observable characteristics to predict the vulnerability index c it v ˆ for each household i at time t belonging to cohort c . For that purpose, c it x 1  can be predicted deterministically from c it x by incrementing age or assuming that characteristics are time invariant. Creation of a Pseudo-Panel In order to have a look at the dynamic of asset-poverty, we regroup households from the DHS into homogeneous cohorts: households whose heads have the same date of birth (we define five-year cohorts), the same level of education (no education, primary and secondary and more) and the same place of residence (ten departments and urban/rural distinction) are regrouped into cells. After regrouping some low-sized cells, 261 cells were constituted for each year of the DHS dataset, with an average size of around 150 households and 950 individuals in each cell. Aggregate Estimates Our estimates of the vulnerability index follow the different steps recalled previously. First, log per c apita asset index is regressed on various household’s characteristics such as log of household size, age of the head and its square, education and gender of the head, location and the presence of children under 5 years old . Residuals are 25 As exposed by Verbeek and Nijman (1992), the bias in the standard within estimator based on pseudo panel data is decreasing with the number of individuals in ea ch cell, more than with the number of cells. However, Verbeek (2008) notes that there is no general rule to judge whether cell size is large enough. Deaton (1985) also suggests that measurement error decreases as a function of the size of the cells. 26 As i n Deaton and Paxson (1994), we can normalize so that the fitted age effect at, for instance, age 35 - 40 equals the average residual variance of the logarithm of the asset index for 35 - 40 year - olds over all cohorts. 17 estimated from these regressions. Second, we calculate for each cell the variance of the residuals of the first - stage household - level regression. Third, we regress the residual variance on cohort dummies (created by crossing household head date of birth, education and locat ion dummies) and a polynomial function of age (generally of two degrees or more if statistically significant). From the age profile of the residual variance, we calculate the slope which is an estimate of the variance of asset shocks. Note that this slope should necessary be positive (i.e. the amplitude of shocks grows with age) since the estimated variance should always be positive. This is generally the case. However, when it is not, contiguous cells have been regrouped for the estimates. Finally, once th e variance of shocks is estimated for each cohort then the last estimation step consists in drawing values of shocks within the standard normal distribution and estimating the household vulnerability index using coefficient estimates. Poverty and vulnerability rates are reported in Table 8 where a household is considered as poor when its asset index is below the 80 th percentile of the 1995 distribution of asset index. An extremely poor household is one whose asset index is below the 40 th percentile of the 1995 distribution of asset index. A household is considered as vulnerable if the probability to be poor or extremely poor is higher than 0.5, whereas it is considered as highly vulnerable for a probability higher than 0.8. At the national level, poverty headcount (71.5%) is not different from the estimated