Policy Research Working Paper 10825 Understanding Vulnerability to Poverty and Natural Disasters in Latin America and the Caribbean Gustavo J. Canavire Bacarreza Adriana Conconi Sergio Olivieri Monserrat Serio Poverty and Equity Global Practice June 2024 Policy Research Working Paper 10825 Abstract This paper provides the first measures of vulnerability poverty is related to households’ coping and adaptation to poverty in Latin America and the Caribbean at the strategies, and there is great heterogeneity depending on household level looking at natural hazards such as floods, the natural hazard assumed. The evidence generated by landslides, cyclones, earthquakes, and droughts. It consid- this approach helps in understanding the lack of coping ers seven countries in the region: Brazil, the Dominican mechanisms and adaptation strategies among vulnerable Republic, Ecuador, El Salvador, Honduras, Mexico, and households. This understanding can be used to mitigate the Peru. The paper constructs a unique data base that links effects of natural disasters. The evidence-based approach is household surveys from the Socio-Economic Database for useful for identifying specific vulnerable communities in Latin America and the Caribbean database to data on natu- disaster-prone areas. Furthermore, the information can be ral hazards from the ThinkHazard database and Emergency presented in the most disaggregated representative unit that Events Database, which permits the first estimates of how the data allow so that policy makers can build more efficient household exposure to natural events raises the probabil- management policies in their location. ity of being poor. The results suggest that vulnerability to This paper is a product of the Poverty and Equity Global Practice. 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://www.worldbank.org/prwp. The authors may be contacted at gcanavire@worldbank.org, solivieri@worldbank.org, mserio@worldbank.org, and aconconi@worldbank.org 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. Produced by the Research Support Team Understanding Vulnerability to Poverty and Natural Disasters in Latin America and the Caribbean * Gustavo J. Canavire Bacarreza Adriana Conconi Sergio Olivieri Monserrat Serio JEL Classification: Q54, I32, D1. Keywords: Vulnerability, Natural Disasters, Poverty, LAC. * We are very grateful to Carlos Rodriguez-Castelan, Ruth Hill, Ghazala Mansuri, Alejandro de La Fuente, Hugo Ñopo, Lander Bosch, and Mattia Amadio for their valuable discussions and comments. The authors would also like to thank Miki Khanh Doan, Juan Jose Miranda Montero, and William Maloney for their insightful suggestions and revisions. The findings, interpretations, and conclusions in this paper are entirely those of the authors. They do not necessarily represent the views of the World Bank Group or the countries in the study. 1. Introduction Climate change is a threat to the objective of sustainably eradicating poverty. Poor people are more often and harshly affected by climate change as they have fewer mechanisms to cope, more losses, and less support to recover. But even those not in poverty can suffer negative impacts that condition their future welfare prospects. Climate change can affect people's well- being through various channels, and it can also have an impact on natural disasters, leading to further negative consequences for people. The identification of the specific impact of climate change on natural disaster events such as floods, landslides, droughts, and storms is challenging and complex to estimate. However, as Hallegatte (2014) mentions, climate change has played an increasing role over time, dramatically escalating natural hazards, and displaying heterogeneous effects on people's well-being depending on the type of hazard. In Latin America and the Caribbean (LAC), the acceleration of climate change has shown a sharp increase in the frequency and intensity of extreme weather events (World Bank, 2022). In recent years, some of the deadliest natural catastrophes have occurred, including the earthquakes in Haiti (230,000 deaths in 2010 and 2,248 deaths in 2021), the Brumadinho mudslide in Brazil (300 deaths in 2019), and hurricanes Maria (3,059 deaths in 2017) and Eta (183 deaths in 2020) (UNDP, 2022). The number of people in the region affected by natural disasters has also increased substantially over the last decades. Between 2000 and 2019, the Emergency Events Database (EM-DAT) indicates that 41 million people have been affected by floods, while droughts have impacted the lives of approximately 53 million people (Guha- Sapir, et al., 2022). In this study, we look at natural disasters and the mechanisms of households to cope with them. This study is framed within a climate risk framework, with a specific focus on the household perspective. As per the terminology of the Intergovernmental Panel on Climate Change (IPCC), disaster risk is the result of the interplay between hazard, exposure, and vulnerability dimensions. Hazard refers to the possible future occurrence of natural or human-induced physical events that may have adverse effects on vulnerable and exposed elements (White, 1973; UNDRO, 1980; Cardona, 1990; UNDHA, 1992; Birkmann, 2006b). Exposure refers to the inventory of elements in an area in which hazard events may occur (Cardona, 1990; UNISDR, 2004, 2009b). And vulnerability refers to the propensity of exposed elements such as human beings, their livelihoods, and assets to suffer adverse effects when hazard events impact. Moreover, for international organizations such as the World Bank, the Intergovernmental Panel on Climate Change (IPCC), the European Commission, and the ISO standards, among others, the term vulnerability encompasses a variety of concepts and elements, including the sensitivity or susceptibility, the weaknesses, and the lack of capacity to cope and adapt that favor the potential adverse effects. As previously mentioned, there is a connection between climate change and the increased intensity and frequency of natural hazards. These hazards can impact the well-being of households, potentially leading to poverty. Vulnerability to poverty of a household is understood as the probability that households will fall into poverty. It is an ex-ante condition, a latent variable that is not directly observable but depends on the future household welfare. The primary distinction between poverty and vulnerability to poverty is that the latter 2 considers the household’s prospects due to the adverse effects from shocks such as natural disasters, their exposure, and the capacity to mitigate or recover from them. One of the best ways to understand the capacity to recover, bearing in mind the household's current characteristics and future shocks, is to involve the well-being dynamics perspective in the vulnerability to poverty assessment through long-term panel data (Chaudhuri, 2003). However, even though the quality and the number of panel data have increased in Latin America and the Caribbean, long-term panel databases still need to be available in the region. Nonetheless, the literature has built measures of vulnerability to poverty in the context of cross-sectional data (Chaudhuri et al., 2002; Chaudhuri, 2003; Dercon, 2006; Bérgolo et al., 2010; Hill & Porter, 2017; Skoufias, Vinha, and Beyene, 2021, among others). In this study, we develop vulnerability to poverty measures under a climate risk framework using cross-sectional data. To do this, we model vulnerability to poverty as a function of variables that can be related to hazards, exposure, sensitivity, adaptative capacity, and household demographic characteristics. Based on a micro-framework, we model the probability of being poor, including current households' coping mechanisms and observable aggregate measures that are proxies of exposure and natural shocks. Our analysis aimed to cover as many countries as possible, focusing on those where administrative-level data on hazards and exposure could be aggregated due to the survey design at the district level. Ultimately, we included seven countries in our analysis: Brazil, the Dominican Republic, Ecuador, El Salvador, Honduras, Mexico, and Peru. Our approach presents differences from others. Typically, natural disaster literature measures disasters from risk metrics and maps, implicitly including household adaptation capacity (Ward et al., 2020; Cremen et al., 2022, among others). In contrast, Chaudhuri's (2003) and earlier works measure vulnerability to poverty by estimating the consumption process and drawing inferences about future consumption prospects, such as mean and variance, based on household conditions and unobserved idiosyncratic shocks without taking into account aggregate shocks. To do this, the authors model the idiosyncratic shocks (household-level shocks) through the variance of the disturbance term of the linear consumption equation. Skoufias, Vinha, and Beyene (2021) account for unobserved idiosyncratic and community covariate shocks modeling consumption through multilevel models and the variance of the disturbance term of the model. Hill and Porter (2017), for their vulnerability computation, work with objective measures of aggregate and idiosyncratic shocks. They seek exogenous sources of shocks to estimate causal effects. They use crop losses as a measure of rainfall shocks and control for the probability of drought in each community using the historical distribution of droughts, considering the timing of the shock as exogenous. As idiosyncratic shocks, they include the household member's death or job loss. Other studies have introduced climate change and natural disaster impacts taking a macro-level perspective (Hallegatte et al., 2016, 2017; Jafino et al., 2020, among others). Our focus is to analyze the household's vulnerability to poverty in a context of climate risk at a micro-level perspective by analyzing the household's socioeconomic conditions and adaptation capacity, as well as the potential natural disasters that could impact them and their exposure based on their location of residence. To estimate vulnerability to poverty, we make an effort to include aggregate shocks explicitly in our model. In the region, the data reveals differences in natural hazard occurrences among subregions; for instance, in the Caribbean and Central America, storms, including tropical 3 cyclones such as hurricanes, are more prevalent. In South America, floods occur more frequently, and droughts have an important impact affecting people. Following EM-DAT, we found that the five most common and impactful natural disasters in the region are floods, tropical storms (including tropical cyclones such as hurricanes), droughts, earthquakes, and landslides. 2 Thus, we use objective measures of these five hazards in the model, but unlike Hill and Porter (2017), we do not model or draw the distribution of these future shocks. Our cross-sectional measures indicate the potential damage from natural hazards based on a probabilistic analysis which is interpreted as a proxy of future aggregate shocks. We also do not include idiosyncratic shocks in the model. Although, we explore them when we perform the Chaudhuri procedure as a robustness exercise. Finally, this study aims to provide disaggregated data to policy makers to build policies to reduce vulnerability to poverty due to natural hazards in LAC. In this paper, we address the effects of climate risk in the region, which is a relatively unexplored aspect in the region. While there have been some macro-level papers relying on simulations such as Jafino et al. (2020), Solano-Rodríguez et al. (2021), Banerjee et al. (2021) or on agricultural and livestock productivity such as Reyer et al. (2017) and Lachaud et al. (2022) simulations; our main contribution lies in a micro-level perspective focusing on households. Caruso (2017) explores natural disasters’ long-run effects in the region in the 20th century at the micro level exploiting census data. And other studies focus on individual countries within the region or specific types of natural hazards (e.g., Albert, Bustos, and Ponticelli, 2023; Acevedo et al., 2023; Rude and Robayo-Abril, 2023; Aguilar-Gomez et al., 2022). However, there is a lack of comprehensive understanding regarding climate risk in the overall region in the present context. We contribute to addressing this gap by choosing various natural hazards across different countries in the region. The paper's main contribution is the analysis of vulnerable households' coping and adaptation mechanisms at the micro-level. The paper is structured as follows. The next section presents the conceptual framework and the methodology. Section 3 describes the data, and section 4 presents the results. Section 5 summarizes the main findings and presents a final discussion. 2. Measuring vulnerability to poverty including natural hazards in LAC Methodology Building on the definition of vulnerability, we propose a model to measure vulnerability to poverty, including natural disasters and exposure in the region. Although a household's vulnerability is an ex-ante condition and not directly observable until the adverse shock occurs (Ravallion, 1988; Chaudhuri, 2003), it is possible to summarize it as the probability of the household's welfare falling below a minimum standard in the future. 2 Hurricanes are a type of tropical cyclones in the Atlantic Ocean. From now on, we will refer to cyclones instead of hurricanes to follow the terminology of the aggregate natural hazard datasets used in the analysis. Following, ThinkHazard! methodology, cyclones, tropical storms, hurricanes, and typhoons, despite their distinct names, all describe the same hazard: a non-frontal storm system characterized by a low-pressure center, spiral rain bands, and strong winds. Thus, we decide to use the term cyclones to encompass a broader scope of this hazard. 4 Defining household welfare is challenging and complex, and there is no perfect welfare indicator to measure it. Nevertheless, absolute poverty is usually a reasonable approximation of household welfare, correlating with vulnerability to natural hazards. Thus, we approximate household welfare using a monetary poverty measure, and we measure households' vulnerability to absolute poverty (poverty from now on) in a setting where it is possible that households suffer damage from natural hazards. In LAC, monetary poverty is usually computed over household incomes and not consumption since all countries in the region collect income information more widely and periodically, and only a few collect frequently consumption information. In addition, it is essential to recognize that the vulnerability concept encompasses the future well-being of the household, which is dynamic. One approach to achieving this is by utilizing long panel data. However, periodic panel datasets are not available in the region, and to overcome this, we propose a model based on a cross- sectional study. Since the concept of vulnerability includes the susceptibility to be harmed and the capacity to cope and adapt, we measure household vulnerability as the propensity of being poor conditioned to observed natural hazards' potential damage, community exposure, and current coping and adaptation strategies. We extend Chaudhuri's (2003) and Skoufias et al. (2021) approach, where shocks are unobserved, and explicitly introduce aggregate shocks through observed natural hazard variables in the model. We also incorporate some elements of Hill and Porter's (2017) study working at the household level. And we exploit the spatial variability of the cross-section data as much as the survey design allows (i.e., for some countries, we go down at the district level, and for others, at the state/province level). Thus, we propose to estimate the probability of being vulnerable to poverty using the following Probit model: (ℎ,, = 1 | , , , ) = ( + , + , + ℎ,, + ℎℎ,, + + ℎ,, ), where ℎ represents households, indicates districts whenever the data permits (and states when the data does not allow a more disaggregate unit), and regions. The variable is a dummy that takes value 1 if the household is poor (i.e., their per capita household income does not exceed the international poverty line of USD 6.85 per day adjusted by PPP 2017). ( ) is a vector of aggregate variables, including indexes of potential damage from natural hazards and the number of natural events occurred in the past. ( ) is a vector of aggregate variables, including the presence of people (natural logarithm of density), infrastructure (percentage of urbanization area), and resources (percentage of cropland area). ( ) is a vector of household variables related to coping mechanisms and adapting strategies at the household level, including income diversification and occupation, social assistance, housing conditions, access to the Internet, and household head's education level. ℎ () includes the size of the household, household head's age, squared age, and gender. And ℎ,, is the error term. 5 We estimate the equation for each country individually and use it to predict the probability of households falling into poverty. We begin by considering one natural hazard at a time and then estimate the full model by taking into account all five natural hazards. Measuring vulnerability index Once we predict each household's probability of falling into poverty, we classify each household as vulnerable if it has a predicted probability of being poor greater than 0.50. The choice of the threshold of 0.50 is arbitrary; nevertheless, it has been used in several studies for its simplicity and intuitive interpretation (Chaudhuri, 2003, among others). Likewise, we perform the analysis considering other thresholds too. After classifying households as vulnerable or not, we compute the incidence of vulnerability using the headcount ratio. In the cases where household data allows state representation, we compute vulnerable ratios at state levels in maps. It is important to note that this measure of household vulnerability does not account for the depth and severity of the vulnerability. The possibility of overcoming the damage caused by natural disasters depends predominantly on households' coping and adaptation mechanisms. Hence, we provide conditional and unconditional profiles to explore the coping and adaptation mechanisms of vulnerable households. For example, evidence shows that the same natural hazards affect people far more than others when their livelihoods depend on specific physical assets such as livestock, their consumption is closer to subsistence levels, and they cannot rely on savings to smooth out the impacts or even if they are going through a critical moment in life for the development of human capital, the impacts on education and health can extend the risks (Hallegatte et al., 2017). The conditional profiles show the percentage of vulnerable households lacking the coping mechanisms included as controls in the regressions. The unconditional profiles explore the lack of coping mechanisms of the households outside of the model. In particular, we assess the percentage of vulnerable households reporting only one sector of activity and only one income earner in the household. We look at the percentage of vulnerable households without legal title ownership of the dwelling, whereas it is crucial to have dwelling and land titles to enhance investments in housing infrastructure to protect it from disasters. We also compute the percentage of vulnerable households where at least one member does not have medical insurance because health costs can potentially affect the family's future well-being. Finally, we explore the percentage of vulnerable households without remittances; households with remittances are more likely to smooth their consumption due to this external source of income. Further, we compute the percentage of vulnerable households with children younger than 5 years old, vulnerable households with children less than 12 years old attending school, and vulnerable households with all their members being elders (65 years old or older). These measures aim to identify groups facing greater risk from natural disasters. The first two groups correspond to children in different stages of life who may suffer more from reductions in their investments in human capital. The third group faces a significant risk due to their low potential capacity to generate an income flow that allows them to restock their assets and wealth better. 6 Next, we explore whether there are differences in coping and adaptation mechanisms among poor and non-poor vulnerable households. To do this, we calculate the above household profiles, differentiating them by poor status. In addition, we carried out a counterfactual exercise. In this exercise, we estimated the households' vulnerability in a scenario where all households face the lowest possible hazards and compare the estimated vulnerability rates with the actual vulnerability incidence. Thus, we set to all households the minimum hazard they can face, that is, the minimum number of events that occurred among districts and a value of very low damage (more details are presented in Appendix B). 3. Data and descriptive statistics The datasets for the analysis are the national Household Surveys of each country of the study from the SEDLAC database. This database provides harmonized information of coping and adaptation mechanisms at the household level and poverty status for different countries of the region. The analysis focuses on seven countries: Brazil, the Dominican Republic, Ecuador, El Salvador, Honduras, Mexico, and Peru. We use household survey data, pre-dating Covid- 19, from 2018 for Mexico and from 2019 for the rest of the countries. Except for Brazil survey, the surveys allow for the identification of the district where households reside. This enables the imputation of hazard and exposure aggregate data at the district level for each household. In the case of Brazil, the survey design led to impute at the state level. However, despite the imputation being carried out at the district level in most countries, the results need to be analyzed at the level of survey representation, which is at the state level. To identify poor households, we utilize a monetary poverty measure assuming a poverty line of 6.85 dollars a day for the per capita total household income adjusted by 2017 PPP. The household poverty rate is 20% in Brazil, 15% in the Dominican Republic, 21% in Ecuador, 23% in El Salvador, 45% in Honduras, 26% in Mexico, and 26% in Peru (Table 1). Table 1. Household poverty headcount ratio by country. Dominican Country Brazil Ecuador El Salvador Honduras Mexico Peru Rep. Poverty adjusted by PPP 2017 Poor HH (6.85 U$S) 20% 15% 21% 23% 45% 26% 26% Poor HH (3.65 U$S) 8% 2% 6% 5% 23% 7% 8% Poor HH (2.15 U$S) 4% 1% 1% 1% 11% 2% 2% Note: Households are poor if their total income per capita falls below the a-day International Poverty Line (adjusted by 2017 PPP and CPI prices). Source: authors' calculations based on the SEDLAC database. The selection and construction of variables of households' coping mechanisms, adaptation strategies, and socio-demographic characteristics are detailed in Appendix A. Here Table 2 provides basic information on all seven Latin American countries of coping and demographic variables ( and ) included in the model. Statistics suggest a low occupation protection among households and a lack of income diversification to deal with natural hazards. The level of informality is high; on average, in Ecuador and Peru, half of the household heads have an informal job. Moreover, in Ecuador, Honduras, and Peru, nearly one household in four reports 7 that the household head's principal occupation is agriculture. And, in most countries, except El Salvador and Honduras, households benefit from social protection that can be adapted as a safety net to respond to natural disasters. Moreover, the results show, on average, a significant proportion of fragile infrastructure and buildings susceptible to natural disasters and limited access to toilet facilities. In El Salvador and Mexico, the percentage of households living in fragile dwellings is to around one-quarter, and the percentage of households with access to toilet facilities linked to the sewer ranges, on average, from 23 percent to 68 percent, depending on the country. Table 2. Average of households with coping and adaptation mechanisms. Domini- El Hondu- Brazil can Ecuador Mexico Peru Salvador ras Category Variable Rep. HH head is salaried in 38.9% 38.4% 43.3% 40.6% 35.8% 56.8% 33.9% the main occupation HH head is informal in 27.0% 36.6% 50.6% 39.6% 42.8% 32.5% 51.5% the main occupation The non-labor income Income 194.98 116.47 115.15 88.19 39.55 89.32 100.81 per capita PPP 2017 diversification Ln. non-labor income and 2.75 2.92 2.94 2.14 1.67 2.23 3.55 per capita PPP 2017 occupation Labor income as a protection 67.1% 74.7% 73.3% 75.6% 78.6% 77.9% 71.8% share of total income Dependency rate 1.8 1.7 1.7 2.1 2.2 1.8 1.9 HH head's principal income is from 6.4% 8.4% 23.8% 14.3% 23.9% 12.8% 26.6% agriculture Social HH receives benefits 19.3% 26.3% 18.6% 1.9% 12.2% 28.9% 47.2% protection from social assistance Dwelling's materials of 3.2% 5.9% 19.7% 25.6% 11.4% 25.8% 12.6% low quality Toilet facilities linked Housing 67.7% 23.2% 64.0% 35.6% 38.3% 62.1% 73.0% to sewer Access to electricity 99.8% 99.1% 99.0% 97.6% 94.1% 99.5% 95.6% Overcrowding 3.5% 3.9% 28.9% 8.7% 17.0% 10.6% 8.9% Car 49.4% 20.9% 23.3% 20.2% 23.6% 42.5% 11.8% Assets Fridge 98.1% 87.9% 80.8% 75.7% 72.4% 86.3% 54.5% ownership Computer 41.7% 16.7% 41.7% 16.7% 17.2% 27.4% 32.7% Washing machine 65.9% 78.7% 41.0% 21.1% . 67.9% 30.5% TICs Access to Internet 37.7% 25.9% 36.4% 23.3% 27.4% 40.6% 35.9% Years of education of Education 8.5 8.7 8.6 6.8 6.2 8.8 8.9 HH head Number of hh members 3.0 3.1 6.0 3.5 4.1 3.6 3.6 Demographics Age household head 48 49 51 50 50 50 53 Male household head 52.8% 63.1% 71.3% 62.9% 65.8% 71.3% 68.8% Source: authors' calculations based on the SEDLAC database. For hazard variables ( ), we exploit two types of data, the number of occurrences of natural disasters in the region in the past and the potential damage expected from them. For the first one, we use the number of natural disasters that occurred in 2000-2020 at the state level from the EM-DAT of the Université Catholique de Louvain, Belgium. This database offers a 8 worldwide collection of natural disasters. Figure 1 shows the distribution and the number of natural disasters in LAC during 1900-2021. The distribution indicates that some natural disasters are more common than others. Thus, the study focuses on the five most common events: floods, tropical cyclones, earthquakes, droughts, and landslides, leaving aside epidemics led by biological processes. Figure 1. Number of natural disasters occurred in LAC during 1900-2021 by subtype. 40% 1,040 35% 30% 25% 595 20% 15% 271 10% 181 171 144 72 65 56 93 59 5% 10 3 0% Mass movement (dry) Drought Earthquake Volcanic activity Epidemic Insect infestation Tropical cyclone Flood Extreme temperature Convective storm Landslide Wildfire Other storms Hydrological Meteorological Geophysical Biological Climatological Source: EM-DAT, CRED / UCLouvain Also, the data shows differences by subregions, both in the types of disasters they face and how they impact the people of each country (Figure 2). For example, in the Caribbean, cyclones are more common. In Central America, both cyclones and floods are similarly significant, while in South America, floods represent almost half of all shocks, followed by landslides though out by far. In Table 3, we can see differences in each country, too. Figure 2. Distribution of natural disasters during 1900-2021 by subtype. Source: EM-DAT, CRED / UCLouvain For potential damage, we use the index of ThinkHazard! Platform. This platform provides data on the potential damage of hazards triggered by geophysical, hydrologic, and meteorologic processes. The Thinkhazard! index (TH index) classifies hazards as high (potentially severe damaging events), medium (potentially damaging events), low (potentially damaging events are less likely to occur but still possible), and very low (potentially damaging effects are unlikely to occur). ThinkHazard levels are derived from hazard maps applying the spatial distribution of hazard intensity (e.g., flood depth, ground 9 shaking) at a given frequency or 'return period' based on a probabilistic analysis (more details are presented in Appendix A). Although, the potential damage captured through the TH indexes is not a projection, it allows us to observe which types of natural events could affect more in the areas where households reside. But it is worth noting that TH scores do not represent hazard risks per se, they are only physical hazards from probabilistic models, and their use for risk-related assessment is discouraged. ThinkHazard! methodology provides a general view of the hazards for a given location to promote disaster and climate resilience. It is a simple flagging system highlighting the hazards in an area. Moreover, the scores emphasize the likelihood of different natural hazards affecting areas. Thus, the data can show some uncertainties in specific areas that fail to cover the full range of hazard categories and intensities (see more detail in ThinkHazard! Platform, 2022). The TH indexes are available from level 2 of administrative units (ADM 2) of FAO's classification. In the case of Brazil, the index is used at level 1 because the geographic variable available in the Household Survey is only at the state level. To capture the differences among administrative units of level 2, we use the average of the indexes of level 2 (see Appendix A). So, this country's index cannot be compared with the other countries' TH indexes. We use data from local areas to control for exposure variables ( ). By definition, exposure is the presence of people, livelihoods, species or ecosystems, environmental functions, services, resources, infrastructure, or economic, social, or cultural assets in places and settings that could be adversely affected by events (IPCC, 2014). Following this definition, we include population density (people per sq. km of the district or state area), percentage of cropland, and urbanization of the district or state area. The data source of type of land is from the MODIS 500m Landcover Dataset NASA LP DAAC at the USGS EROS Center, retrieved from Google Earth Engine Data Catalog based on the 2018 Images LC type 1 with 17 classes of land use. We employ class 12 for the cropland area (where at least 60% of the area is cultivated), class 13 for the urban area (at least 30% is impervious surface area, including building materials, asphalt, and vehicles), and the sum of all the classes for the total area of the administrative units at level 2. Table 3 summarizes the average and standard deviation of the main variables of hazards and exposure at administrative level used in the analysis. The results indicate that Central America and Mexico face an elevated potential damage from cyclones and a higher frequency of occurrences of this hazard. In South America, floods are more prevalent, but countries like Peru and Ecuador evidence on average a higher potential damage from landslides than floods. 10 Table 3. Aggregate statistics of natural hazards and exposure by country. Brazil Dominican Rep. Ecuador El Salvador Honduras Mexico Perú VARIABLES Obs. Mean SD Obs. Mean SD Obs. Mean SD Obs. Mean SD Obs. Mean SD Obs. Mean SD Obs. Mean SD Hazards TH index at ADM 2 Floods 27 2.66 0.7 134 3.10 1.1 219 2.07 1.2 211 1.74 1.1 225 1.97 1.2 879 1.98 1.2 181 1.64 1.1 Landslides 27 1.53 0.4 135 2.90 0.8 222 3.27 1.0 225 3.23 0.8 227 3.48 0.7 883 2.45 1.2 182 3.36 0.8 Earthquake 27 1.19 0.3 136 3.00 0.0 223 3.28 0.5 225 4.00 0.0 230 3.29 0.5 881 2.09 1.0 182 3.23 0.5 Cyclones 5 1.00 0.0 22 4.00 0.0 82 1.00 0.0 2 4.00 0.0 28 4.00 0.0 903 3.94 0.3 4 1.00 0.0 Droughts 27 1.59 0.7 136 3.00 0.0 223 1.00 0.0 225 3.00 0.0 230 2.11 0.3 903 2.31 0.9 182 1.41 0.7 EM-DAT: Number of events (2000-2021) at ADM 1 Floods 23 6.87 6.36 32 6.75 2.51 24 5.08 3.49 14 5.50 1.70 16 6.69 2.63 32 3.19 3.16 25 9.20 3.74 Landslides 23 0.39 0.94 32 0.00 0.00 24 0.29 0.55 14 0.07 0.27 16 0.00 0.00 32 0.25 0.80 25 0.56 0.87 Earthquake 23 0.04 0.21 32 0.13 0.34 24 0.67 0.82 14 1.86 0.66 16 0.50 0.73 32 0.78 1.39 25 1.60 1.53 Cyclones 23 0.09 0.29 32 11.66 1.77 24 0.00 0.00 14 5.64 1.28 16 4.00 2.22 32 6.63 5.97 25 0.00 0.00 Droughts 23 1.57 1.53 32 0.00 0.00 24 0.29 0.46 14 2.07 1.07 16 3.38 1.78 32 0.50 0.67 25 1.20 0.41 Exposure at ADM 2 Density 27 75.7 119.9 135 350 1173 25 84 80 225 599 1290 230 124 135 902 746 2094 182 68 296 Ln. density 27 3.4 1.5 135 4.8 1.2 25 3.9 1.3 225 5.6 1.1 230 4.4 0.9 902 4.9 1.9 182 3.0 1.4 Area 27 314,318 376,072 135 325 302 223 1,131 1,884 225 86 91 230 387 743 902 1,462 3,201 182 6,829 13,195 % cropland 27 5% 7% 135 6% 11% 223 3% 10% 225 5% 11% 230 2% 6% 902 18% 25% 182 2% 4% % urbanization 27 1% 2% 135 3% 9% 223 2% 6% 225 2% 8% 230 0% 1% 902 4% 15% 182 1% 4% Note: density in Brazil and Ecuador is at the state level. The statistics are computed before imputations. Source: authors' calculations based on ThinkHazard! EM-DAT, and SEDLAC database. 11 4. Results Table 4 presents the estimated coefficients from the hazard term of the full Probit model that includes all the natural hazards, controls for regional fixed effects, and standard errors clustered within administrative units. It also displays the estimated average marginal effects revealing heterogeneous effects among natural hazards across countries. In the case of Brazil, both floods and droughts present a direct correlation with household poverty. Conversely, earthquakes, cyclones, and landslides positively correlate with poor households in the Dominican Republic. Poverty in Ecuador and Peru correlates positively with the number of landslides and in El Salvador with the number of droughts, earthquakes, and cyclones. In Honduras, the hazards that positively relate to poverty are earthquakes and droughts. Finally, cyclones and floods are positively associated with household poverty in Mexico. Moreover, the correlations presented in Table 4 lead to some interesting conjectures. For example, when both ThinkHazard! and EM-DAT coefficients are positive, such as cyclones in Mexico, the most common hazards coincide with the most disruptive hazards regarding poverty status suggesting a double high association. But, in some cases, significant estimations show opposite signs in hazard variables, e.g., floods in Brazil show a positive correlation with the ThinkHazard! index and a negative correlation with the EM-DAT variable. ThinkHazard! is based on the likelihood of each region experiencing a hazard with enough intensity to cause damage. Meanwhile, the EM-DAT measures provide the frequency of past events. When the coefficient of ThinkHazard! is positive and the coefficient of EM- DAT is negative, it suggests, on the one hand, that hazards with more intensity of causing damage are more likely to have a greater impact on communities than expected. On the other hand, more frequency (controlling already for intensity) could imply that communities that have experienced these events in the past could develop some forms to prevent and adapt between occurrences, preventing a continuous cycle of poverty. Conversely, when the coefficient of ThinkHazard! is negative and the coefficient of EM-DAT is positive, it is not necessarily the severity of the natural hazard that plunges households into current poverty but rather the frequency of such events. The constant recurrence of the past of these hazards could create a cycle of economic setbacks for families, leading to the positive relationship of the present. Something interesting to note in this case is that the higher potential damage from hazards is associated with areas where more affluent households reside. Although these households may have more resources to cope and adapt, they face a higher risk of losing valuable assets and goods. Also, the estimations indicate the relevance of coping mechanisms (although these estimates are not shown in Table 4, they are similar to those estimated with the initial poverty model presented in Appendix A and are available upon request). For instance, having an informal household head, a high dependency ratio, and a primary income from agriculture increases a household's vulnerability to falling into poverty, and larger households are also more likely to be vulnerable. Receiving benefits from social assistance programs is positively correlated with the household's vulnerability. On the other hand, households with salaried household heads, those with a larger share of labor income, better housing conditions, and assets are less likely to be vulnerable. 12 Table 4. Probit model estimations including all hazards. Brazil Dominican Republic Ecuador El Salvador Honduras Mexico Peru Mg. Mg. Mg. Mg. Mg. Mg. Mg. Probit Effects Probit Effects Probit Effects Probit Effects Probit Effects Probit Effects Probit Effects TH index Flood 0.105 0.012 0.000 0.000 -0.032 -0.006 -0.025 -0.005 -0.036 -0.008 0.025 0.005 -0.018 -0.003 [0.056]* [0.006]* [0.042] [0.006] [0.042] [0.008] [0.017] [0.003] [0.032] [0.007] [0.013]* [0.003]* [0.023] [0.004] Number of Floods -0.014 -0.002 -0.086 -0.012 0.018 0.003 -0.089 -0.017 -0.039 -0.008 0.021 0.004 -0.009 -0.001 [0.005]*** [0.001]*** [0.084] [0.012] [0.022] [0.004] [0.031]*** [0.006]*** [0.019]** [0.004]** [0.008]** [0.002]** [0.005]* [0.001]* TH index Landslides -0.333 -0.038 0.063 0.009 -0.026 -0.005 -0.010 -0.002 0.027 0.006 -0.012 -0.002 -0.074 -0.012 [0.158]** [0.018]** [0.067] [0.009] [0.049] [0.009] [0.023] [0.004] [0.044] [0.010] [0.017] [0.003] [0.024]*** [0.004]*** Number of Landslides 0.231 0.026 0.171 0.032 -0.130 -0.025 -0.099 -0.019 0.047 0.008 [0.064]*** [0.007]*** [0.085]** [0.016]** [0.124] [0.024] [0.019]*** [0.004]*** [0.020]** [0.003]** TH index Earthquake 0.079 0.009 0.068 0.013 0.105 0.023 -0.010 -0.002 -0.022 -0.004 [0.092] [0.011] [0.074] [0.014] [0.123] [0.027] [0.023] [0.005] [0.040] [0.006] Number of Earthquakes -0.546 -0.062 0.878 0.125 -0.109 -0.021 0.765 0.146 0.019 0.004 -0.036 -0.007 -0.022 -0.003 [0.201]*** [0.023]*** [0.294]*** [0.042]*** [0.057]* [0.011]* [0.168]*** [0.032]*** [0.056] [0.012] [0.021]* [0.004]* [0.017] [0.003] TH index Cyclones 0.158 0.031 [0.029]*** [0.006]*** Number of Cyclones 0.059 0.008 0.300 0.057 0.039 0.009 0.003 0.001 [0.050] [0.007] [0.074]*** [0.014]*** [0.023]* [0.005]* [0.003] [0.001] TH index Drought 0.065 0.007 0.012 0.003 -0.049 -0.010 -0.002 -0.000 [0.083] [0.009] [0.145] [0.032] [0.022]** [0.004]** [0.024] [0.004] Number of Droughts 0.062 0.007 -0.028 -0.005 0.962 0.183 0.032 0.007 -0.023 -0.004 -0.030 -0.005 [0.020]*** [0.002]*** [0.094] [0.018] [0.212]*** [0.040]*** [0.030] [0.007] [0.034] [0.007] [0.045] [0.007] Controlling by exposure Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by coping mechanisms Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by region Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Note: *** p<0.01, ** p<0.05, * p<0.1. Standard errors clustered by districts in brackets (by states in Brazil and Ecuador). Source: authors' calculations based on ThinkHazard! EM-DAT, and SEDLAC database. 13 We also estimate Probit models for every natural hazard separately. Figure 3 presents the average marginal effects of each natural hazard of these estimations. Positive marginal effects indicate a positive relationship between household poverty and the hazard of the event. In Brazil, earthquakes and droughts show positive effects, ceteris paribus. In the Dominican Republic, the number of earthquakes and landslides' potential damage positively correlate to poverty. The number of landslides and the earthquake's potential damage in Ecuador positively correlate to poverty. The droughts and the number of landslides matter in El Salvador. In Honduras, mainly earthquakes are associated with household poverty. In Mexico, cyclones are the natural hazards associated with household poverty. In Peru, the number of landslides positively associates with poverty. Owing to the differences in natural hazard effects, below, when we present the results of vulnerability that come from the disasters separately models, we are going to present the results from the models that only include earthquakes in Honduras, cyclones in Mexico, floods in Peru, droughts in Brazil and El Salvador, and landslides in the Dominican Republic and Ecuador. Additionally, it is important to highlight that we do not count on the potential compounding effects of these hazards by incorporating their interactions into the models (for example, including the interaction between floods and landslides). This is an aspect that remains to be studied. Figure 3. Average marginal effects from the models with natural hazards separately. TH index Floods Number of Floods TH index Landslides Number of Landslides TH index Earthquakes Number of Earthquakes TH index Cyclones Number of Cyclones TH index Droughts Number of Droughts 0.026*** 0.027 0.024* 0.019 0.015** 0.015 0.014 0.014 0.011 0.008*** 0.009 0.007** 0.008 0.007 0.005** 0.004 0.002** 0.003 0.003 0.003 0.003 0.001** 0.002 0.001 0.001 -0.001 -0.002 -0.002 -0.003 -0.003 -0.004 -0.005** -0.006** -0.008 -0.008 -0.009** -0.010*** -0.011*** -0.013** -0.013*** -0.015 B R AZIL M E XIC O PERU DOM INIC AN EL HONDURAS ECUADOR REP. SALVADOR Note: *** p<0.01, ** p<0.05, * p<0.1. Each family color bar corresponds to a specification that includes one type of natural hazard controlled by exposure, coping mechanisms, and region. Source: authors' calculations based on ThinkHazard!, EM-DAT, and SEDLAC database. 14 The results show that Honduras has the highest vulnerability to poverty rate (41%), and the Dominican Republic has the lowest (7%). The vulnerability incidence in Peru is 24%, indicating that almost a quarter of the households have a predicted probability over 0.50 of falling into poverty. In Mexico, the vulnerability rate is 20%, El Salvador 17%, Brazil 16%, and Ecuador 14%. Figure 4. Vulnerability incidence (% of vulnerable households with a predicted probability greater than 0.50 including all hazards). Honduras 41% Peru 24% Mexico 20% El Salvador 17% Brazil 16% Ecuador 14% Dom. Rep. 7% 0% 10% 20% 30% 40% 50% Source: authors' calculations. In Figure 5, the maps show the percentage of vulnerable households in each state in the case of the regressions with only one natural hazard at a time. In the case of Brazil, droughts are the hazards included in the model. The results show higher vulnerability in the north in line with the fact that poverty, exposure, and natural hazards concentrate more in that region and that households in this area have fewer coping mechanisms than those in the south. The Dominican Republic shows that some administrative units present a vulnerability rate for landslides notably higher than the national average. The vulnerability rate of the provinces Independencia, Baoruco, and Pedernales is 36%, 25%, and 24%, respectively. Also, the map shows that Barahona and Puerto del Plata have districts with a higher vulnerability rate than the rest. In Ecuador, the vulnerable households are concentrated in the east, the region with more effective poverty rates. But where the TH index of landslides suggests more potential damage is in the central region, where some cantons also show a high vulnerability. El Salvador's vulnerability rate for droughts, except for San Salvador, is very similar among provinces. But some differences are found between districts. In particular, some districts of Morazan and Ahuachapan show a high rate in comparison to districts of other provinces. Honduras shows that the administrative unit at level 1 with the highest percentage of vulnerable households when earthquakes are considered is Lempira (78%). The map shows that many other districts present a sizeable rate. But, in Lempira's case, all their districts present a significant rate. Lempira is located in the west, corresponding to a region with significant potential earthquake damage. 15 With cyclones, Mexican households show a high vulnerability rate in Chiapas (59%), Oaxaca (56%), and Guerrero (46%), followed by Veracruz (39%) and Tabasco (37%). Furthermore, these provinces broadly suffer from a high amount of cyclone activity. From the results obtained for floods in Peru, the percentage of vulnerable households increases in the mountain regions in the center and tropical regions in the east. In the eastern is where floods are more potentially damaging. Even though Loreto province experiences more potential damage from floods, administrative units in the mountain regions with high poverty rates show the most significant vulnerability rates, such as Huancavelica, followed by Ayacucho and Cajamarca. It is worth noting that the coastal region is comparatively less vulnerable to floods mainly because is relatively less poor than other regions. However, it is important to consider the potential impact of coastal floods on this area, such as the El Niño Costero, which have not been addressed here. 16 Figure 5. Vulnerability incidence using a threshold of 0.50 by state. Vulnerability in Brazil Vulnerability in Dominican Vulnerability in Ecuador Vulnerability in El Salvador Vulnerability in Honduras Vulnerability in Mexico Vulnerability in Peru Droughts Republic Landslides Droughts Earthquakes Cyclones Floods % of vulnerable hh by Landslides % of vulnerable hh by % of vulnerable hh by state % of vulnerable hh by state % of vulnerable hh by state % of vulnerable hh by state state % of vulnerable hh by state state Maranhao Independencia Morona… Morazan Lempira Chiapas Huancavelica Piaui Baoruco Orellana Oaxaca Pedernales Ahuachapan Cajamarca Paraiba Pastaza La Paz Guerrero Alagoas Barahona Veracruz Ayacucho Hato Mayor Napo Cabanas Tabasco Para Intibuca Puno Elias Pina Sucumbios Puebla Acre Hua­nuco San Cristobal Esmeraldas Cuscatlan Olancho San Luis Potosi Amazonas San Pedro… Zacatecas Apurimac Sergipe Chimborazo Loreto Monte Plata La Union Comayagua Hidalgo Ceara Peravia Zamora… Durango Pasco Amapa San José de… Bolivar San Vicente Campeche Santa Barbara Tlaxcala Amazonas Pernambuco San Juan Loja Michoacan Cusco Rio… El Seibo Zona No… Sonsonate Ocotepeque Guanajuato San Martín Bahia Santo… Yucatan Manabi Usulutan Junin Roraima Distrito… Paraiso Morelos Duarte Cotopaxi Ancash Tocantins Mexico Azua Imbabura La Paz Nayarit Rondonia Copan Ucayali La Vega Santa Elena Queretaro Espirito… San Miguel Piura Maria… Carchi Choluteca Tamaulipas Minas… Espaillat Sinaloa La Libertad Mato Grosso Canar Chalatenango Coahuila Madre de Dios Valverde Yoro Rio De… Sanchez… Galapagos Quintana Roo Moquegua Mato… Santa Ana Colima Samana Los Rios Valle Lambayeque Sao Paulo Puerto Plata Chihuahua Azuay La Libertad Jalisco Tumbes Parana La Romana Colon Santo… Aguascalientes Tacna Goias Salcedo Sonora Guayas San Salvador Distrito… Dajabon Atlantida Arequipa Baja… Santiago Tungurahua Nuevo Leon Callao Rio… La Altagracia Pichincha 0% 50% Francisco Baja California Ica Santa… Morazan Santiago… El Oro Distrito Federal Lima 0% 50% Monte Cristi Cortes 0% 20% 40% 0% 50% 0% 50% 0% 20% 40% 60% 0% Source: authors' calculations. 17 Following the literature, we are using a threshold of 0.50, but we also perform the analysis for other thresholds obtaining vulnerability curves (Figure C.1 in Appendix C shows results with other thresholds). Although the levels are different, the pattern and conclusions are replicated in a similar manner. In the counterfactual exercise, the vulnerability in Mexico decreases from 20% to 12%, suggesting the cyclones contribute to vulnerability by 8 p.p. This difference represents around 11 million individuals who become vulnerable when the hazard from cyclones is considered. In Honduras, the simulation of earthquake hazard suggest that this hazard raises vulnerability from 33% to 41%; this difference of 8 p.p. represents around 600,000 (0.6 M) new vulnerable people. In the rest of the countries, the vulnerability increases 1 p.p. In Brazil, droughts increase vulnerability from 15% to 16% (nearly 2 million individuals becoming vulnerable). For landslides, the vulnerability rate in the Dominican Republic ranges from 6% to 7% and in Ecuador from 13% to 14% (0.1 M and 0.28 M individuals become vulnerable, respectively). In El Salvador, the vulnerability rate from droughts specification changes from 16% to 17% (almost 100,000 individuals becoming vulnerable). According to the World Bank (2022) report based on Jafino et al. (2020), climatic events could push between 2.4 million and 5.8 million people in the region into extreme poverty by 2030. In this study, from the simulated scenario, we estimate that natural hazards push around 14.2 million people to vulnerability in moderate poverty. Although this is a lower bound because we are including only one hazard per country, the number is in line with other simulations in macro frameworks. Figure 6. Vulnerability rates comparison between counterfactual scenario and actual scenario. Low harm Actual harm 45% 41% 41% 41% 39% 40% 37% 33% 35% 30% 25% 20% 17% 20% 14% 16% 16% 14% 13% 14% 15% 15% 12% 11% 12% 10% 6%7% 5% 0% Brazil Brazil Brazil Dom. Rep. Honduras Dom. Rep. Honduras Honduras Honduras Ecuador Mexico Ecuador Ecuador Mexico El Salvador Floods Landslides Earthquakes Cyclones Droughts Note: The results correspond to the specifications, including one type of natural hazard controlled by exposure, coping mechanisms, and region. Source: authors' calculations based on ThinkHazard!, EM- DAT, and SEDLAC database. Vulnerability profiles 18 We perform a profile assessment to understand the strategies of vulnerable households for coping or adapting in the face of natural disasters. For each country, we compute the percentage of vulnerable and non-vulnerable households looking whether the head of the household is a non-salaried worker, whether they work in only one sector, whether there is only one income earner in the household, whether the primary source of income is from agriculture, whether the dwelling has low-quality materials, whether the household receives benefits from social assistance, and whether the household does not receive are any remittances from abroad. We also compute the percentage of households with children under 5 years old, children under 12 years old, and households with all their members elders 65 years old or older. Table C.1 in Appendix C shows these percentages at the country level for the seven countries according to each specific natural hazard. In Figure 7, we show Mexico's profile from the model conditioned to cyclones hazard. The percentage of households with primary income from agriculture, operating in the informal sector and not having salaried household heads is higher among the vulnerable than the non- vulnerable. Risk assessments show that agricultural production and income from agriculture can be depressed for years as an additional impact of natural hazards (Hallegatte et al., 2017). Households highly dependent on agriculture and ecosystems are more exposed to cyclones. Thus, diversification of incomes, safety nets to smooth consumption, and climate-smart agriculture practices need to be on the country's agenda. In addition, physical infrastructure and housing with more climate and disaster strength to reduce damage to dwellings are also needed. The percentage of households living in fragile dwellings is significantly higher among the vulnerable households, showing that these dwellings can be particularly affected by cyclones. The rate of households with kids is higher among the vulnerable than the non- vulnerable, but this is not the case for elders. Therefore, vulnerability reduction measures need to include responses to avoid or diminish the permanent impacts of cyclones, especially on children's education and health. Figure 7. Mexican households' coping mechanisms and adaptation strategies. 80% Non Vulnerable Vulnerable 70% 60% 50% 40% 30% 20% 10% 0% HH head's main attending school Only one sector Walls' materials HH head Only elders (65 materials of low social assistance salaried worker Only one income or more) in HH informal HH head not a Presence of Roof's materials children <5 of low quality Benefits from income from of low quality children <12 earner in HH Presence of agriculture Dwelling's of activity reported quality Source: authors' calculations. 19 To check for differences between poor and non-poor households, we divided the sample by household's poor status. Figure C.1 in Appendix C shows that the vulnerability rate includes poor and non-poor households. This is due to the vulnerability model used in the analysis, which may include, among the vulnerable, households with a high probability of being poor (a probability over 0.50) even when they are not actually poor. Nonetheless, the percentage of non-poor households vulnerable to poverty is lower than that of poor households. Figure 8 shows that vulnerable households present a similar lack of coping mechanisms and adaptation strategies, whether poor or non-poor. Figure 8. Vulnerable Mexican households' coping mechanisms and adaptation strategies by poor status. Non poor Poor 80% 70% 60% 50% 40% 30% 20% 10% 0% not a informal no more one income children only children dwelling's roof's walls' not salaried than one income from younger elders (65 attending materials materials materials benefits worker sector of earner agriculture than 5 or more) school from activity social program Source: authors' calculations. The profiles at the national level of the other countries are in Table C.1 in Appendix C. The results suggest that in Brazil, strengthening vulnerable households to engage in multiple economic sectors is necessary. Vulnerable households report more working only in one sector of activity and having only one income earner. Also, the incidence of households with children is higher among vulnerable households. Children's incidence and lack of health insurance in the Dominican Republic are mainly distinguished among vulnerable from non- vulnerable households. In Ecuador, the prevalence of diversification from agricultural income and dwelling materials shows that fragile buildings and lack of income diversification mainly encompass vulnerable households. In El Salvador and Honduras, vulnerable households lack income diversification and occupation protection, and they also show a significant percentage of households with children whose human capital may be affected by natural hazards. In Peru, the lack of income and occupation diversification is noticeable, leaving households less able to cope and recover. Moreover, these profiles can be extrapolated or go as low as the survey allows it to analyze at the state or district level and see the degree of lack of adaptation and coping mechanisms of different regions. For example, the maps in Figure 9 present at the state level the potential hazard from cyclones, the percentage of households without coping mechanisms, and specific groups of households at significant risk in Mexico. The maps in green correspond to all households (vulnerable and non-vulnerable), and the maps in red correspond to only 20 vulnerable households showing the incidence of a specific characteristic. The maps display that the region where more cyclones have occurred is the region with more vulnerable households obtaining their main income from agriculture and having fragile buildings. Also, some states such as Guerrero, Colima, and Jalisco with great potential damage show a high percentage of vulnerable households without legal title over the dwelling, which discourages investing in their dwelling to reduce damage from natural hazards. Further, Michoacan and Jalisco show a high percentage of vulnerable households with children under 5, which can suffer more cyclone effects by reducing investments in their human capital. Figure 9. Percentage of Mexican households lacking coping mechanisms and adaptation strategies by geographic units. Source: authors' calculations. Figure 10 shows that in Brazil, a significant percentage of vulnerable households from Minas Gerais depend on agricultural income, an area also prone to drought. Likewise, Paraiba and Pernambuco states are prone to the effects of droughts, and access to water supply systems is not easy among vulnerable households in this area. The water scarcity likelihood of the northeast of Brazil makes water management a key factor for water security, food security, and productive activities in these states. 21 Figure 10. Percentage of Brazilian households lacking coping mechanisms and adaptation strategies by geographic units Source: authors' calculations. In Central America and the Caribbean, remittances are a source of household capacity to cope, increasing income diversification, and typically remain relatively stable after disasters (Hallegatte et al., 2017). Figure 11 shows for the Dominican Republic that although vulnerable households benefit from remittances, the regions suffering higher landslides potential damage are the regions where the percentage of vulnerable households receiving remittances is lower. Figure 11. Percentage of Dominican Republic households with remittances as a coping mechanism to landslide's potential damage by geographic units. Source: authors' calculations. 22 Households that own their dwellings are more likely to invest in and implement adaptive measures against natural hazards. Ownership provides a sense of stability and control, enabling residents to make long-term decisions and allocate resources towards protecting their homes and investing in adaptative strategies to potential hazards. Ownership of dwellings is recorded in household survey data in almost all the countries that have been included in this study. About 70% of the households in the sample report that they are the owners of the dwelling without finding a significant difference between vulnerable and non- vulnerable households in this regard in Brazil, El Salvador, Honduras, Mexico, and Peru. Except in the case of Ecuador, where around 23% of vulnerable households report owning the dwelling. In Ecuador, some states that face potential damage from landslides show a high percentage of households that do not own their dwellings. This is challenging to improve the houses to face landslides damage, without home ownership, investments in the dwellings towards protecting from natural hazards are significantly less. Figure 12. Percentage of Ecuadorian households that do not own their dwellings as a lack of dwelling's adaptative strategy to landslide's potential damage by geographic units. Source: authors' calculations. However, it is not always clear whether the household has legal documents of ownership. In some cases, homes are marked as owned even if they are still under payment or do not have legal title. The dwellings' improvement programs of risk mitigation may not have the expected response while the regulation and legal ownership do not change, and other adaptation and coping strategies to face potential damages to dwellings and infrastructure would be needed in these states. Surveys for Honduras, Peru, and Mexico include data on legal documentation. In Figure 13, we observe that in Peru, most of the states facing a higher potential for flood damage are those reporting the lowest percentages of vulnerable households without legal titles to their dwellings. Conversely, in the central region, a high percentage of vulnerable households lack legal property titles and may reside in districts with a high susceptibility to flooding. Also, the results suggest there is a high percentage of households that are vulnerable to poverty that are engaged in informal employment. These households are mainly located in the northern and southern regions of the central area, where informality can be a hindrance to coping, especially during floods in the north. 23 Figure 13. Percentage of Peruvian households that do not own legal title of their dwellings as a lack of dwelling's adaptative strategy to flood's potential damage by geographic units. % of vulnerable hh not own % of vulnerable hh legal title of the dwelling in informal employment by state by state HH with predicted probability > 0.50 HH with predicted probability > 0.50 Floods Floods (.83,.9] (.83,.89] (.77,.83] (.78,.83] (.68,.77] (.62,.78] [.5,.68] [.33,.62] Source: authors' calculations. Maps in Figure 14 show that some districts have a large percentage of vulnerable households living in dwellings with materials of low quality, which can be particularly affected by earthquakes. These Honduran districts can be targeted for future decisions on projects or policies that upgrade the infrastructure to reduce the vulnerability of their buildings. Figure 14. Percentage of Honduran households living in fragile dwellings to face earthquake's potential damage by geographic units. Source: authors' calculations. Robustness checks As a first robustness check, we verify whether the model accurately targets households that are actually poor. We can expect that the model estimates the probability of being effectively poor when the hazard and exposure terms are absent. Thus, we run this as a robustness check of our model to ensure that the predicted probability of being poor leads to similar poverty rates observed in the data. First, we use LASSO to select the covariates of the model that fit the poverty data more adequately. Then, we run the model excluding the hazard and exposure variables. The estimations suggest that coping and demographic variables are effective in 24 predicting poverty. The correlation between the predicted probability of being poor and the actual poverty status are significant compared to other studies. In addition, when we examined the headcount of households with a predicted probability of being poor larger than 0.50 and the headcount of poverty, we found that the ratio is close to 1 for most countries (for more details, see Appendix A). We also test for confounding of coping variables dropping hazard variables. We find that for example some coping variables are related to hazards but as we expect hazards are not related to coping. Even though we are not trying to estimate the causal effects of harm or coping variables, the results align with the fact that harm variables can affect the present household's characteristics/copings but not the other way around. The coping and characteristics of the household seem not to affect the variables of harm as is expected. We also estimate the model excluding ThinkHazard! indexes and including the number of occurrences from EM-DAT which is a proxy of frequency of hazards based on historical data. The estimations of EM- DAT coefficients do not show significant differences between both models (Table A.2 in the Appendix). Following Chaudhuri (2003), we model the household income per capita considering idiosyncratic shocks. These shocks enter directly into the vulnerability estimation through the standard deviation of the income model's error term making distributional assumptions in the error variance to draw the poverty rate. However, identifying which type of shocks and risks are operating in the disturbance term is puzzling, making this a handicap. The incidence under this model works as an upper bound as it considers other shocks and risks that fall in the disturbance term, which are not captured through the observables. The results highly correlate with the original vulnerability measures from the Probit model. The details and results of these robustness checks are in Appendix B. We also performed one more robust check constraining the hazard coefficients to positive values. Vulnerability to poverty using our approach might be higher or less than the actual poverty. The empirical strategy of estimating vulnerability to poverty conditioned to hazards, exposure, and coping factors may predict a high probability of being poor even if households do not experience considerable hazards. Further, it may not identify households with a high susceptibility to hazards as vulnerable when the potential damage of hazards correlates negatively with districts' poverty. The latter is a critical caveat of the model, and we explore the effects of this miss-allocation using a restricted non-linear model where we force hazard coefficients to be positive. We estimate a non-linear probability model, which we named the restricted model, where we set the hazard coefficients as positive. At first, it is not an appealing option to impose structure on the model forcing the sign of the coefficients, but it can be an appropriate robustness check exercise. The results of this exercise show a high correlation between the vulnerability rates computed with the restricted and Probit model (see Figure B.1 in Appendix B). 5. Conclusions We merge nationally representative household datasets of seven countries in the LAC region with aggregate natural disaster and area data to estimate the household vulnerability to poverty in the context of climate risk. We find considerable heterogeneities in terms of the vulnerability to poverty among events. For example, cyclones in Mexico and earthquakes in 25 Honduras imply a substantial increase in vulnerability to poverty. Brazilian and Salvadoran households present a higher vulnerability when drought-prone areas are considered. In Ecuador, landslides and earthquakes relate more to poverty. Regarding policy, many households in the region are unable to cope and adapt to these events, primarily through the lack of income diversification and dwelling material. The results draw attention to specific regions with higher vulnerability to poverty incidence. The methodology proposed in this paper can be an initial tool to explore households' coping strategies and the weakened capacity to respond to natural hazards, to identify priorities for action to reduce households' vulnerability. This paper examines a few mechanisms in the profiles and maps. But the analysis can be extended depending on the action and the country in evaluation. Safety nets generally increase the household's capacity to cope with events and risks. However, other risk management instruments such as insurance, land use, and building regulations can help households manage flood and landslide risks more. On the other hand, early warning systems can also reduce cyclones' impacts, and the awareness of being prepared can help for earthquakes. In our study, we acknowledge the role that several other characteristics can have in helping to cope with the adverse effects of natural hazards, such as the access to and utilization of financial services. However, it is important to note that not all countries incorporate some specific information in their surveys, leading to a lack of standardized variables within the databases in the region. For instance, while the survey in Peru addresses financial access, including account ownership and credit card usage, this detailed information is only uniformly available across some surveyed countries. Consequently, our ability to comprehensively analyze how financial services contribute to hazard coping in the region is constrained. Nonetheless, this type of specific coping mechanism could be incorporated in studies focused specifically on individual countries, which is beyond the scope of our current research. In our study, we employed a composite measure (e.g., TH index) for a systematic analysis of natural hazards, rather than relying on specific indexes for each hazard such as FATHOM for floods or SPEI for droughts. Although the latter approach offers more precision, our decision to use the composite measure was to facilitate comparison among many countries and hazards. We are aware that the precision of the indexes used is lower, and it can lead to certain limitations. Although, our approach allows including specific indexes for each hazard. We let this as a next step for future research towards improving the overall measures of vulnerability to poverty in the countries in the region. Finally, the methodology avoids several limitations of cross-section data using explicitly natural hazard potential damage data. However, we do not explore the compounding effects of these hazards, such as the interaction between floods and landslides, as well as interactions between hazards and specific household characteristics. This is something to address in upcoming steps. Also, some caveats to the analysis concerning unobserved household characteristics, actual natural hazard impact on vulnerability (causality), and the inter- relationship between hazards and other climate change impacts remain. Including climate change measures, such as extreme temperature and rains, as well as other types of disasters (e.g., pluvial and coastal floods), is the following apparent step. Also, in this study, the scarce availability of long-panel data limits the consistent estimates of vulnerability to poverty. Performing our approach using short synthetic panels to include dynamics in the model is a 26 forward step that could enhance the analysis. Synthetic panels based on matching individuals with the same time-invariant characteristics in consecutive cross-sections have been recently proposed as a substitute for such data (Moreno, Bourguignon, and Dang, 2021). Here we collapse the number of occurrences in a period but using short synthetic panels, we could include the time variability and explore observable households' strategies to cope with hazards across periods. References Acevedo, I., Castellani, F., Lotti, G., & Székely, M. (2023). Natural Disasters and Labor Market Outcomes in Mexico. IDB Working paper series No IDB - WP - 0 149 0. Inter - American Development Bank. http://dx.doi.org/10.18235/0005187 Aguilar-Gomez, S., Gutierrez, E., Heres, D., Jaume, D., & Tobal, M. (2022). Thermal stress and financial distress: Extreme temperatures and firms’ loan defaults in Mexico. Working paper, Available at SSRN 3934688. https://ssrn.com/abstract=3934688 or http://dx.doi.org/10.2139/ssrn.3934688 Albert, C., Bustos, P., & Ponticelli, J. (2023). The effects of climate change on labor and capital reallocation. NBER working papers, No. w28995. National Bureau of Economic Research. http://www.nber.org/papers/w28995 Alkire, S., & Santos, M. E. (2010). Acute Multidimensional Poverty: A New Index for Developing Countries. OPHI Working Paper No. 38 Bandyopadhyay, S., & Skoufias, E. (2015). Rainfall variability, occupational choice, and welfare in rural Bangladesh. Review of Economics of the Household, 13(3), 589-634. Banerjee, O., Cicowiez, M., Rios, A. R., & De Lima, C. Z. (2021). Climate change impacts on agriculture in Latin America and the Caribbean: an application of the Integrated Economic- Environmental Modeling (IEEM) Platform (No. IDB-WP-01289). IDB Working Paper Series. Bérgolo, M., Cruces, G., Gasparini, L., & Ham, A. (2010). Vulnerability to poverty in Latin America Empirical evidence from cross-sectional data and robustness analysis with panel data. Working paper, 170, Chronic Poverty Research Centre. Caruso, G. D. (2017). The legacy of natural disasters: The intergenerational impact of 100 years of disasters in Latin America. Journal of Development Economics, 127, 209-233. Chaudhuri, S., Jalan, J., & Suryahadi, A. (2002). Assessing household vulnerability to poverty from cross-sectional data: A methodology and estimates from Indonesia. Discussion Paper #:0102-52, Department of Economics, Columbia University, New York. Chaudhuri, S. (2003). Assessing vulnerability to poverty: concepts, empirical methods, and illustrative examples. Mimeo. Department of Economics, Columbia University, New York, 56. Cremen, G., Galasso, C., & McCloskey, J. (2022). Modelling and quantifying tomorrow's risks from natural hazards. Science of The Total Environment, 817, 152552. Dercon, S. (2006). Vulnerability: a micro perspective. Securing development in an unstable world, 30, 117-146. Figal Garone, L., León, S., Olarte Rodríguez, L., & Peña, X. (2019). Home durables and access to credit among the low-income in Latin America and the Caribbean: Trends and challenges. Development through the Private Sector Series TN No. 15, Washington DC: IDB Invest. 27 Gasparini, L., Marchionni, M., Olivieri, S., & Sosa Escudero, W. (2011). Multidimensional poverty in Latin America and the Caribbean. New evidence from the Gallup World Poll. Manuscript, CEDLAS. Guha-Sapir D., R. Below, & Ph. Hoyois (2022). EM-DAT: The CRED/OFDA International Disaster Database – www.emdat.be – Université Catholique de Louvain – Brussels – Belgium. Hallegatte, S. (2014). Trends in Hazards and the Role of Climate Change. In: Natural Disasters and Climate Change. Springer, Cham. https://doi.org/10.1007/978-3-319-08933- 1_4 Hallegatte, Stephane, Adrien Vogt-Schilb, Mook Bangalore, & Julie Rozenberg (2017). Unbreakable: Building the Resilience of the Poor in the Face of Natural Disasters. Climate Change and Development Series. Washington, DC: World Bank. doi:10.1596/978-1-4648- 1003-9. Hill, R. V., & Porter, C. (2017). Vulnerability to drought and food price shocks: evidence from Ethiopia. World Development, 96, 65-77. Jafino, B. A.; Walsh, B.; Rozenberg, J. & Hallegatte, S. (2020). Revised Estimates of the Impact of Climate Change on Extreme Poverty by 2030. Policy Research Working Paper; No. 9417. World Bank, Washington, DC. © World Bank. https://openknowledge.worldbank.org/handle/10986/34555 License: CC BY 3.0 IGO. IPCC (2014). Summary for policymakers. In: Climate Change 2014: Impacts, Adaptation, and Vulnerability. Part A: Global and Sectoral Aspects. Contribution of Working Group II to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change [Field, C.B., V.R. Barros, D.J. Dokken, K.J. Mach, M.D. Mastrandrea, T.E. Bilir, M. Chatterjee, K.L. Ebi, Y.O. Estrada, R.C. Genova, B. Girma, E.S. Kissel, A.N. Levy, S. MacCracken, P. R. Mastrandrea, and L. L. White (eds.)]. Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA, pp. 1-32. Lachaud, M. A., Bravo‐Ureta, B. E., & Ludena, C. E. (2022). Economic effects of climate change on agricultural production and productivity in Latin America and the Caribbean (LAC). Agricultural Economics, 53(2), 321-332. MODIS Landcover Collection (2018). MCD12Q1.006 MODIS Land Cover Type Yearly Global 500m from Earth Engine Data Catalog https://developers.google.com/earth- engine/datasets/catalog/MODIS_006_MCD12Q1#description, maintained by the NASA EOSDIS Land Processes Distributed Active Archive Center (LP DAAC) at the USGS Earth Resources Observation and Science (EROS) Center, Sioux Falls, South Dakota. Retrieved March 2022. Moreno, H.; Bourguignon, F. & Dang, H.A (2021). On Synthetic Income Panels. IZA Discussion Papers, No. 14236, Institute of Labor Economics (IZA), Bonn. Ravallion, M. (1988). Expected poverty under risk-induced welfare variability. The Economic Journal, 98(393), 1171-1182. Reyer, C. P., Adams, S., Albrecht, T., Baarsch, F., Boit, A., Canales Trujillo, N., ... & Thonicke, K. (2017). Climate change impacts Latin America and the Caribbean and their implications for development. Regional Environmental Change, 17, 1601-1621. Rude, B., & Robayo-Abril, M. (2023). Quantifying Vulnerability to Poverty in El Salvador. Policy Research Working Paper; No. 10289. World Bank, Washington, DC. © World Bank. 28 Skoufias, E., Vinha, K., & Beyene, B. M. (2021). Quantifying Vulnerability to Poverty in the Drought- Prone Lowlands of Ethiopia. Policy Research Working Paper; No. 9534. World Bank, Washington, DC. © World Bank. https://openknowledge.worldbank.org/handle/10986/35107 Sulla-Menashe, D., & Friedl, M. A. (2018). User guide to collection 6 MODIS land cover (MCD12Q1 and MCD12C1) products. USGS: Reston, VA, USA, 1, 18. Socio-Economic Database for Latin America and the Caribbean (2022). https://www.cedlas.econo.unlp.edu.ar/wp/en/estadisticas/sedlac/. CEDLAS and The World Bank. Retrieved March 2022. Solano-Rodríguez, B., Pye, S., Li, P. H., Ekins, P., Manzano, O., & Vogt-Schilb, A. (2021). Implications of climate targets on oil production and fiscal revenues in Latin America and the Caribbean. Energy and Climate Change, 2, 100037. https://doi.org/10.1016/j.egycc.2021.100037 Stampini, M., & Tornarolli, L. (2012). The growth of conditional cash transfers in Latin America and the Caribbean: did they go too far? IZA Policy Paper. No. 49 ThinkHazard! Platform (2022). https://thinkhazard.org/. Global Facility for Disaster Reduction and Recovery (GFDRR Labs). The World Bank. Retrieved March 2022. UNDP (2022). Social Protection in Times of Uncertainty. Technical Note: Recalibrating… Getting Back on Track Towards the SDGs. XIV Ministerial Forum for Development in Latin America and the Caribbean. Ward, P. J., Blauhut, V., Bloemendaal, N., Daniell, J. E., de Ruiter, M. C., Duncan, M. J., Emberson, R., Jenkins, S. F., Kirschbaum, D., Kunz, M., Mohr, S., Muis, S., Riddell, G. A., Schäfer, A., Stanley, T., Veldkamp, T. I. E. & Winsemius, H. C. (2020). Natural hazard risk assessments at the global scale. Natural Hazards and Earth System Sciences, 20(4), 1069-1096. World Bank (2022). Hoja de ruta para la acción climática en América Latina y el Caribe 2021-2025. Zhang, H., Zhao, Y., & Pedersen, J. (2020). Capital assets framework for analyzing household vulnerability during disasters. Disasters, 44(4), 687-707. 29 Appendix Appendix A: Data sources, database construction, and variables selected for the analysis The database for the analysis is the most recent wave of 2019 Household Surveys (2018 for Mexico) of the studied countries standardized under the SEDLAC's methodology. These datasets contain rich information on the households' coping mechanisms, adaptation strategies, and socio-demographic characteristics. We merged each household database with natural hazards and exposure aggregate data at the most disaggregate geographic unit level, which can correspond to a canton, a district, a municipality, or a province/state, depending on the country. In the case of Mexico and Peru, the units correspond to districts; in the Dominican Republic, El Salvador, and Honduras, they are municipalities; in Ecuador, they are cantons; and in Brazil, they are provinces (states). For aggregate data from natural hazards, we look at the five more common hazards in Latin America and the Caribbean (LAC): river floods, landslides, earthquakes, droughts, and cyclones. We use potential damage information from ThinkHazard! 's Index to measure households' susceptibility to hazards. This index classifies each natural hazard as High (potentially severe damage), Medium (potentially damaging effects), Low (potentially damaging events less likely to occur but still possible), and Very Low (potentially damaging effects unlikely to occur). Following ThinkHazard! 's classification, "… Hazard levels are derived from hazard maps presenting the spatial distribution of hazard intensity (e.g., flood depth, ground shaking) at a given frequency, or 'return period.' The timeframe for each hazard depends on the timescales over which the hazard causal processes operate and historical data available to assess long-term averages." The classification is made at the scale of the more disaggregated local administrative units following the ADM2 of the FAO Global Administrative Unit Layers (GAUL) division. In the case of Brazil, the Household Survey allows us to work at the province/state level and not at district level. Although, ThinkHazard! Platform provides an index at ADM 1 level; it is not used in Brazil because ThinkHazard follows the maximum method where the level of potential damage at ADM1 units is defined as the maximum hazard level by all its lower ADM 2 units. This aggregation process shows very little variability between provinces (for example, two provinces, one with only one district with potentially severe damage and the other with potentially severe damage in all its districts, would have the same ThinkHazard! Index at ADM 1). Therefore, we decided to aggregate to ADM 1 using the indexes at ADM 2 computing the ThinkHazard! 's index at province level as the average of the municipality’s indexes. We also include another measure of aggregate hazard related to observed frequency using the data from the Emergency Events Database (EM-DAT) of the Université Catholique de Louvain (Guha- Sapir et al., 2022). This database provides a collection of natural disasters in the studied countries from 1900 to the present. The database reports the impacted administrative units by the event since 2000. Therefore, we decided to work with the number of events from 2000 to 2021 that occurred in each country at the state level. Merging aggregate hazard data and household survey data is not straightforward in all countries. Even though we have information on natural hazards in most of the ADM 2 units, in some cases, we do not have the corresponding unit in the household survey database from SEDLAC. For example, in Ecuador, the canton Eloy Alfaro from Esmeraldas province is not in the SEDLAC database, and in Honduras, Gracias a Dios and Islas de Bahía states are not in the SEDLAC database. In these cases, we do not include these units in the final database. Meanwhile, in other cases, we have units in the SEDLAC database without natural hazard data. In these cases, we impute hazard data of their nearest and similar unit neighbor. We carried out this imputation in the Dominican Republic, Mexico, and Honduras. For example, in the Dominican Republic, we imputed the hazard level of Villa Hermosa by the one of La Romana, which is 14 km of distance; Villa Montellano by Puerto Plata (14 km); Puñal by Santiago (15 km); and Guayacanes by San Pedro de Macoris (19 km). 30 After the corresponding imputation, we obtain that in the Dominican Republic, all ADM 2 units (hereinafter "districts" for all countries) of the SEDLAC database have aggregate natural hazard information. In Ecuador, all districts in the SEDLAC database have natural hazard information too. In Honduras, we impute hazard data with the nearest neighbor method in four districts (Jutiapa by La Ceiba; Trojes by Danlí; San Francisco Opalaca by La Iguala; Nueva Frontera by Macuelizo), obtaining only 170 observations in the SEDLAC database without hazard information (0.7% of the sample). In Mexico, we re-coded Oaxaca and six San Luis de Potosi districts and imputed 33 districts of twelve states with the nearest neighbor method, reaching 0.9% of observations of the sample without hazard information distributed among 22 districts across nine states. We also consider the area of districts or, in Brazil's case, the area of provinces. We retrieved the area data from Google Earth Engine Data Catalog using the MODIS 500m Landcover Dataset of the NASA LP DAAC at the USGS EROS Center with a spatial resolution of 500 meters. The area data corresponds to the 2018 landcover images from LC type 1, which includes 17 classes of land use (the total area is computed as the sum of them). Based on these datasets, we use three variables of exposure, the natural logarithm of density (population/area in km2) as a proxy of the presence of people, the percentage of urbanization area as a proxy of infrastructure developed, and the percentage of cropland as a proxy of resources of the region. In the data, urban areas are those where at least 30% of the surface area is built-up, including building materials, asphalt, and vehicles, and the cropland areas are those where at least 60% of the surface area is cultivated cropland. One limitation of using 500 m is that in areas of the tropics where cropland field sizes tend to be much smaller than a MODIS pixel, agriculture is sometimes underrepresented and can be labeled as natural vegetation instead (Sulla- Menashe and Friedl, 2018). In addition, for the model, we select variables of coping mechanisms, adaptation, and demographics at the household level that are good enough to predict effective household poverty in the absence of natural hazards and exposure. For this selection of households' coping mechanisms and adaptation strategies to natural disasters, we follow the Unbreakable Report (Hallegatte et al., 2017). We choose the following variables as a proxy of income diversification and occupation: the household head is a salaried worker; the household head is informal; natural logarithm of the household's non-labor income per capita, labor income as a share of total income, dependency rate, and household head's principal income is from agriculture. Also, as a proxy of assistance, we choose a dummy indicating if the household receives benefits from social assistance. For housing conditions, we select low-quality materials in the dwelling, toilet facilities linked to sewer, access to electricity, overcrowding, assets ownership (including car, fridge, computer, washing machine), and access to the Internet. And as sociodemographic, we choose the years of education of the household head, the number of household members, the household head's age and squared age, and the gender of the household head. Next, we explained the selected coping and adaptation variables. It is well known that natural disaster events have socioeconomic consequences such as job losses and falling incomes. Hallegatte et al. (2017) illustrate from a survey of 200 households in Mumbai that floods cause problems with transport and fuel availability. Affected households lose workdays implying a loss of income, productivity, and sometimes jobs. A salaried job can provide minimal protection to lost workdays, and the probability of losing the job is lower than in another type of occupation, and there is the benefit of receiving a paycheck if the job is lost. Furthermore, informality is a primary concern, as households whose head is employed informally frequently do not have access to safety nets and protection when shocks hit. Diversifying income sources is a coping mechanism for households in any disaster event. Higher diversification leads to lower income losses. Therefore, we use labor income as a share of total income to measure the diversification of income sources. We also use the household's non-labor income per capita as a proxy of adaptation strategy. However, as Hallegatte et al. (2017) based on 31 Bandyopadhyay and Skoufias (2015) point out, diversifying incomes helps households at all income levels to cope, but mainly with small shocks and not larger ones. Further, households with high dependency ratios are more likely to be vulnerable (Chaudhuri, 2003). The number of income earners (the ratio's denominator) provides a proxy of the household's risk, especially if the household has only one income earner. The sectors of activity of households can also contribute to households' risks. For example, the income from agriculture can be significantly depressed by a natural disaster through agricultural productivity decline and loss of agricultural production. "… Poor people, especially in rural areas without functioning markets, are highly dependent on agricultural income and ecosystems, and they are therefore vulnerable to the impacts of natural disasters on yields and the health and functioning of ecosystems" (Hallegatte et al., 2017, pp. 47). On the other hand, social protection systems can protect poor people. Following Stampini and Tornarolli (2012), Conditional Cash Transfers (CCTs) have spread in the LAC region in the last years, reaching poor households. Social assistance programs' benefits contribute to household income diversification and reduce risk. In any natural disaster, these benefits can work as a buffer. In addition, they can operate as a channel to reach the casualties more quickly. Also, following Hallegatte et al. (2017) the poor and non-poor households' typical asset portfolios are very different. "… Poor people tend to have less diversified portfolios: they hold a larger percentage of their assets in material form and save "in kind." Moreover, … the quality of assets owned by poor people is lower". Households with few productive and financial assets are more likely to be exposed to adverse shocks and have limited earnings prospects and income-generating capacity (Chaudhuri, 2003). In addition, "… in absolute terms, wealthier people lose more assets or income from a natural disaster, which is expected because they have more assets and higher incomes. In relative terms, however, poor people always lose more than the non-poor from floods and storms" (Hallegatte et al., 2017, pp.42). Furthermore, Zhang, Zhao, and Pedersen (2019) show that the ownership of physical assets is crucial for households' post‐disaster recovery. Nonetheless, according to Chaudhuri (2003), in the absence of sufficient assets or insurance to smooth consumption, a shock may lead to irreversible losses, such as permanent damage or a distressed sale of productive assets, which does not allow for a fast post- disaster recovery. The sale of these assets is a costly coping strategy, particularly for poor households. From the above considerations, we choose a variety of home durable goods as household assets to reflect the well-being and coping of the households. Following Alkire and Santos (2013), we include car and refrigerator ownership in the household assets. We also include washing machines and computers because Figal Garone et al. (2019) show that LAC has high heterogeneity in owning these household assets within countries, especially washing machines. Though high-income segments exhibit asset ownership close to developed countries, low-income segments have considerably lower access to these assets. Most poor people live in fragile buildings or dwellings with low-quality materials (Hallegatte et al., 2017). These materials make areas riskier in the event of natural hazards. Also, natural disasters, especially floods, can cause problems with the power supply, drinking water, household drains, and sewers that worsen with a weak infrastructure. Moreover, "… besides losses of private income and assets, natural disasters cause significant disruption to public infrastructure. Even though all people depend to some extent on electricity, working roads, and running water, poorer people tend to be less able to protect themselves from the consequences of disruptions in infrastructure services" (Hallegatte et al., 2017, pp. 46). 32 Finally, following Chaudhuri (2003), a household is more likely to be exposed to adverse shocks and have limited earnings prospects and income-generating capacity if it has low levels of human capital, know-how, and access to information. Also, poor households have little access to early warnings (Hallegatte et al., 2017). Hence, we incorporate household head education and Internet access as coping mechanisms. The estimations of the model in Table A.1 show that the variables of coping and demographics selected are good enough to predict effective poverty. The correlation between the predicted probability of being poor and the effective poverty status goes from 0.59 to 0.77, depending on the country (Brazil presents a correlation coefficient of 0.767, Peru 0.734, Honduras 0.695, Mexico 0.641, El Salvador 0.621, the Dominican Republic 0.583, and Ecuador 0.590). These correlation coefficients are relatively significant compared with other studies, such as Alkire and Santos (2010), which find a correlation of 0.870 for 104 developing countries, and Gasparini et al. (2011), which estimate a correlation of 0.428 for LAC countries. Moreover, when we explore the headcount of households with a predicted probability of being poor larger than 0.50 and the headcount of poverty by using a poverty line of 6.85 USD PPP 2017, we find a ratio for most countries close to 1. The ratio in Brazil is 0.871, the Dominican Republic 0.916, Ecuador 0.652, El Salvador 0.717, Honduras 1.002, Mexico 0.836, and Peru 1.031. Table A. 1. Estimations of the expected poverty model (unconditional natural hazards and exposure). Dominican El Brazil Mexico Peru Honduras Ecuador Republic Salvador Probit Probit Probit Probit Probit Probit Probit Variables HH head is salaried in the main occupation -0.227 -0.211 -0.163 0.159 -0.311 -0.315 -0.217 [0.030]*** [0.022]*** [0.037]*** [0.082]* [0.040]*** [0.064]*** [0.055]*** HH head is informal in the main occupation 0.320 0.317 0.191 0.159 0.165 0.339 0.352 [0.019]*** [0.019]*** [0.041]*** [0.079]** [0.033]*** [0.053]*** [0.044]*** Ln. non-labor income per capita PPP 2017 -0.409 -0.287 -0.674 -0.165 -0.263 -0.311 -0.291 [0.022]*** [0.011]*** [0.032]*** [0.028]*** [0.017]*** [0.026]*** [0.023]*** Labor income as a share of total income (monetary) -2.113 -2.244 -3.662 -2.003 -1.928 -2.319 -2.162 [0.064]*** [0.057]*** [0.174]*** [0.160]*** [0.100]*** [0.147]*** [0.085]*** Dependency rate (members/ number of income earners) 0.692 0.294 0.365 0.304 0.495 0.348 0.233 [0.011]*** [0.012]*** [0.019]*** [0.040]*** [0.023]*** [0.023]*** [0.018]*** HH head's principal income is from agriculture 0.107 0.249 0.308 0.318 0.324 0.411 0.188 [0.040]*** [0.026]*** [0.037]*** [0.099]*** [0.042]*** [0.061]*** [0.054]*** Receives benefits from a social assistance program 1.382 0.679 0.469 0.208 0.649 0.821 0.454 [0.028]*** [0.024]*** [0.036]*** [0.061]*** [0.104]*** [0.100]*** [0.028]*** Dwelling's materials of low quality 0.148 0.135 0.075 0.378 0.205 0.184 0.248 [0.050]*** [0.023]*** [0.058] [0.123]*** [0.029]*** [0.086]** [0.038]*** Toilet facilities linked to the sewer -0.082 -0.207 -0.137 -0.005 -0.158 -0.170 -0.176 [0.024]*** [0.023]*** [0.040]*** [0.077] [0.058]*** [0.059]*** [0.049]*** Dwelling has access to electricity -0.269 -0.091 -0.138 -0.183 -0.045 -0.066 -0.174 33 [0.121]** [0.104] [0.049]*** [0.212] [0.073] [0.111] [0.078]** Number of hh members over the number of bedrooms > 3 0.247 0.235 0.144 0.237 0.149 0.226 0.240 [0.037]*** [0.028]*** [0.054]*** [0.101]** [0.052]*** [0.072]*** [0.040]*** Has a car -0.584 -0.395 -0.468 -0.639 -0.589 -0.408 -0.022 [0.020]*** [0.021]*** [0.079]*** [0.125]*** [0.053]*** [0.093]*** [0.043] Has a fridge -0.190 -0.278 -0.355 -0.242 -0.356 -0.422 -0.348 [0.044]*** [0.027]*** [0.033]*** [0.097]** [0.027]*** [0.058]*** [0.014]*** Has a computer -0.126 -0.253 -0.216 -0.490 -0.416 -0.230 -0.224 [0.044]*** [0.030]*** [0.058]*** [0.074]*** [0.094]*** [0.085]*** [0.029]*** Has a washing machine -0.371 -0.208 -0.303 -0.166 -0.302 0.009 [0.024]*** [0.019]*** [0.052]*** [0.089]* [0.051]*** [0.026] Access to the Internet in the hh -0.370 -0.349 -0.377 -0.319 -0.270 -0.386 -0.316 [0.037]*** [0.024]*** [0.051]*** [0.079]*** [0.076]*** [0.075]*** [0.027]*** Years of education of hh head -0.054 -0.041 -0.035 -0.042 -0.044 -0.058 -0.033 [0.002]*** [0.003]*** [0.004]*** [0.009]*** [0.004]*** [0.009]*** [0.004]*** Controlling by demographics Yes Yes Yes Yes Yes Yes Yes Controlling by region Yes Yes Yes Yes Yes Yes Yes Observations 69,362,803 33,425,611 8,801,118 3,144,980 1,937,964 1,966,380 4,527,881 Note: hh is household. *** p<0.01, ** p<0.05, * p<0.1. Standard errors clustered by districts in brackets (by states in Brazil and Ecuador). Source: Authors' calculations based on SEDLAC surveys, 2019. Table A.2 presents the models' estimations used to estimate vulnerability, including all natural hazards at once and individually. 34 Table A. 2. Probit model estimations including and excluding ThinkHazard indexes. Brazil Probit 1 2 3 4 5 6 7 8 9 10 Number of Floods 0.000 0.000 -0.019 -0.014 [0.005] [0.005] [0.004]*** [0.005]*** TH index Flood -0.019 0.105 [0.069] [0.056]* Number of Landslides 0.042 0.043 0.142 0.231 [0.018]** [0.017]** [0.059]** [0.064]*** TH index Landslides -0.004 -0.333 [0.056] [0.158]** Number of Earthquakes 0.133 0.133 -0.248 -0.546 [0.065]** [0.066]** [0.113]** [0.201]*** TH index Earthquake 0.029 0.079 [0.037] [0.092] Number of Cyclones TH index Cyclones Number of Droughts 0.022 0.020 0.066 0.062 [0.010]* [0.008] * ** [0.020]*** [0.020]*** TH index Drought 0.060 0.065 [0.025] ** [0.083] Controlling by exposure Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by coping mechanisms Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by region Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Dominican Republic Probit 1 2 3 4 5 6 7 8 9 10 Number of Floods 0.009 0.013 -0.082 -0.086 [0.043] [0.045] [0.076] [0.084] TH index Flood -0.027 0.000 [0.041] [0.042] Number of Landslides - - - TH index Landslides 0.065 0.063 [0.065] [0.067] Number of Earthquakes 0.133 0.133 0.827 0.878 [0.116] [0.116] [0.260]*** [0.294]*** TH index Earthquake - Number of Cyclones -0.023 -0.023 0.049 0.059 [0.030] [0.030] [0.049] [0.050] TH index Cyclones - Number of Droughts - TH index Drought 35 Controlling by exposure Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by coping mechanisms Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by region Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Ecuador Probit 1 2 3 4 5 6 7 8 9 10 Number of Floods 0.017 0.018 0.014 0.018 [0.021] [0.021] [0.023] [0.022] TH index Flood 0.002 -0.032 [0.042] [0.042] Number of Landslides 0.120 0.129 0.161 0.171 [0.067]* [0.075]* [0.071]** [0.085]** TH index Landslides -0.007 -0.026 [0.047] [0.049] Number of Earthquakes -0.025 -0.044 -0.085 -0.109 [0.045] [0.054] [0.048]* [0.057]* TH index Earthquake 0.080 0.068 [0.096] [0.074] Number of Cyclones TH index Cyclones Number of Droughts -0.077 -0.077 -0.024 -0.028 [0.115] [0.115] [0.095] [0.094] TH index Drought - - Controlling by exposure Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by coping mechanisms Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by region Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes El Salvador Probit 1 2 3 4 5 6 7 8 9 10 11 12 Number of Floods 0.000 0.003 -0.077 -0.089 [0.011] [0.012] [0.030]** [0.031]*** TH index Flood -0.016 -0.025 [0.016] [0.017] Number of Landslides 0.039 0.035 -0.143 -0.130 [0.078] [0.078] [0.122] [0.124] TH index Landslides -0.013 -0.010 [0.025] [0.023] Number of Earthquakes -0.007 -0.007 0.696 0.765 [0.030] [0.030] [0.165]*** [0.168]*** TH index Earthquake - Number of Cyclones -0.002 -0.002 0.276 0.300 [0.019] [0.019] [0.073]*** [0.074]*** TH index Cyclones - Number of Droughts 0.056 0.056 0.826 0.962 36 [0.049] [0.049] [0.206]*** [0.212]*** TH index Drought - Controlling by exposure Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by coping mechanisms Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by region Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Honduras Probit 1 2 3 4 5 6 7 8 9 10 11 12 Number of Floods -0.025 -0.027 -0.043 -0.039 [0.012]** [0.013]** [0.018]** [0.019]** TH index Flood -0.027 -0.036 [0.031] [0.032] Number of Landslides - - - - TH index Landslides 0.062 0.027 [0.043] [0.044] Number of Earthquakes 0.059 0.064 0.020 0.019 [0.046] [0.046] [0.051] [0.056] TH index Earthquake 0.125 0.105 [0.117] [0.123] Number of Cyclones 0.018 0.018 0.033 0.039 [0.018] [0.018] [0.024] [0.023]* TH index Cyclones - - Number of Droughts -0.010 -0.010 0.040 0.032 [0.019] [0.019] [0.029] [0.030] TH index Drought 0.036 0.012 [0.140] [0.145] Controlling by exposure Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by coping mechanisms Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by region Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Mexico Probit 1 2 3 4 5 6 7 8 9 10 11 12 Number of Floods 0.003 0.002 0.018 0.021 [0.007] [0.007] [0.009]** [0.008]** TH index Flood 0.014 0.025 [0.014] [0.013]* Number of Landslides -0.056 -0.054 -0.090 -0.099 [0.016]** [0.017]** * * [0.020]*** [0.019]*** TH index Landslides 0.004 -0.012 [0.017] [0.017] Number of Earthquakes -0.047 -0.047 -0.055 -0.036 [0.020]** [0.022]** [0.020]*** [0.021]* TH index Earthquake 0.008 -0.010 [0.025] [0.023] Number of Cyclones 0.007 0.006 0.007 0.003 [0.003]* [0.003] * ** [0.003]** [0.003] 37 TH index Cyclones 0.133 0.158 [0.029] *** [0.029]*** Number of Droughts 0.001 0.006 -0.003 -0.023 [0.034] [0.034] [0.036] [0.034] TH index Drought -0.068 -0.049 [0.020]*** [0.022]** Controlling by exposure Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by coping mechanisms Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by region Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Peru Probit 1 2 3 4 5 6 7 8 9 10 Number of Floods -0.003 -0.004 -0.005 -0.009 [0.004] [0.004] [0.004] [0.005]* TH index Flood 0.017 -0.018 [0.021] [0.023] Number of Landslides 0.037 0.049 0.031 0.047 [0.017]** [0.017]** * [0.018]* [0.020]** TH index Landslides -0.060 -0.074 [0.021]** * [0.024]*** Number of Earthquakes -0.036 -0.029 -0.030 -0.022 [0.013]** * [0.014]** [0.015]* [0.017] TH index Earthquake -0.048 -0.022 [0.038] [0.040] Number of Cyclones TH index Cyclones Number of Droughts -0.085 -0.083 -0.017 -0.030 [0.039]* [0.041] * ** [0.044] [0.045] TH index Drought 0.005 -0.002 [0.025] [0.024] Controlling by exposure Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by coping mechanisms Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Controlling by region Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Note: hh is household. *** p<0.01, ** p<0.05, * p<0.1. Standard errors clustered by districts in brackets (by states in Brazil and Ecuador). Source: Authors' calculations based on SEDLAC surveys, 2019. 38 In addition, some variables not included in the model are relevant to explore households' coping mechanisms and adaptation. Therefore, we examine them in the profiles of vulnerability. In the profiles, we explore the incidence of vulnerable households that report only one sector of activity, only one income earner, and that receive benefits from remittances from abroad and assistance mechanisms as income diversification. We also explore the percentage of vulnerable households living in dwellings with legal ownership. This latter is particularly important because legal ownership allows a more feasible improvement of the infrastructure of the dwelling (Hallegatte et al., 2017). Also, insurance markets play a fundamental role in the event of a natural disaster. Unfortunately, the databases do not provide much information on different types of insurance. But there is information on household health insurance that is relevant as natural events' health impacts can be pervasive, especially when the public health system is deficient, pushing households to poor health and catastrophic health expenditures. Likewise, specific demographic groups can face more risks from a natural disaster. Children are often the main ones affected, with permanent impacts on health and education. Hence, we compute the percentage of vulnerable households with children under five years old and children under 12 years old attending school. We also identify vulnerable households where all their members are older than 65. Older adults' coping and adaptation mechanisms can be insufficient if their losses are significant. The elderly usually retire from the labor market and have already made investments during their active lifetime. Thus, a negative shock can make them permanently disabled from generating new incomes and coping with losses. 39 Appendix B: Robustness exercises We simulate a counterfactual scenario where we set minimum hazard to all households. We give each household the minimum number of occurrences of events among districts (states in the case of Brazil) and a value of ThinkHazard! Index of very low. The results suggest that in Brazil, the hazard that households can face in the future increases the vulnerability rate for the threshold of 0.50 by 1 p.p., in Ecuador by 2.5 p.p., in El Salvador by 5 p.p., and in Mexico and Honduras around 7 p.p. Peru and the Dominican Republic, contrary to what is expected in this counterfactual scenario show an increase of their vulnerability. In Peru, inconsistent households represent 5.6% of the sample, and in the Dominican Republic represent 1% of the sample. These inconsistencies derive from the caveats of the methodology when the estimated coefficients associated with the hazard variables of the primary model are negative (hazard variables correlate negatively with poverty). Table B. 1. Vulnerability rates comparison between counterfactual scenario and actual scenario. Vulnerability incidence Vulnerability incidence Vulnerability incidence Floods Landslides Earthquakes Country Actual Scenario Actual Scenario Actual Scenario predicted prob. > 0.50 Brazil 16.17% 16.47% 16.20% 15.94% 16.19% 16.01% Dom. Rep. 7.12% 7.52% 7.13% 6.24% 7.16% 7.00% Ecuador 13.80% 11.68% 13.89% 13.13% 13.85% 11.32% El Salvador 16.54% 16.58% 16.72% 17.28% 16.74% 16.91% Honduras 40.85% 48.18% 41.16% 37.38% 41.12% 33.36% Mexico 20.29% 19.67% 20.34% 20.71% 20.07% 21.17% Peru 23.70% 24.10% 23.71% 25.50% 23.80% 26.56% Vulnerability incidence Vulnerability incidence Vulnerability incidence Cyclones Droughts Total hazards Country Actual Scenario Actual Scenario Actual Scenario predicted prob. > 0.50 Brazil n.a. n.a. 16.18% 15.34% 16.11% 14.86% Dom. Rep. 7.16% 8.17% n.a. n.a. 7.22% 7.58% Ecuador n.a. n.a. 13.86% 14.35% 13.89% 11.41% El Salvador 16.74% 16.81% 16.80% 15.56% 16.75% 11.70% Honduras 40.95% 38.58% 41.04% 40.77% 40.84% 34.23% Mexico 20.14% 12.13% 20.13% 22.33% 20.45% 12.81% Peru n.a. n.a. 23.76% 24.02% 23.73% 29.36% Source: Authors' calculations. Another robustness check comes from Chaudhuri's procedure. Following Chaudhuri (2003), we estimate vulnerability contemplating unexplained idiosyncratic shocks. The procedure uses the expected mean of income and the variance of the unexplained part of the income process. We assume the following simple, functional form for the household income per capita: ln ℎ = ℎ + ℎ where ℎ is the household income per capita of the household ℎ (to simplify, we drop the geographic subindexes), ℎ = (, , ℎ , ℎ ) is a vector of variables of natural hazards, exposure, coping, and demographics used in the primary model and ℎ is the disturbance term that we allow that depends on the observable variables as follows 40 2 ,ℎ = ℎ . We estimate and by the three-step feasible generalized least squares (FGLS) method (Amemiya, 1977). See the estimation's details in Chaudhuri (2003). Then we estimate the expected natural logarithm of household income per capita and its variance: � (ln ̂ �ℎ | ℎ ) = ℎ � (ln �ℎ | ℎ ) = 2 � � ,ℎ = ℎ . Then, the estimated conditional probability of being vulnerable is given by: ̂ ln − ℎ �ℎ < | ℎ ) = Φ � � (ln � � �ℎ where z is the poverty line of 6.85 USD (adjusted by PPP 2017), and Φ denotes the cumulative density function of the standard normal. Chaudhuri vulnerability estimates offer an upper bound rate because it includes unobservable idiosyncratic shocks that contribute to different per capita household incomes. Nevertheless, in � can take negative values, which is inconsistent with the fact that practice, the term ℎ 2 � ,ℎ is positive. Chaudhuri (2003) finds that few observations fall into this issue and drops them. Here, for the Dominican Republic and El Salvador, this is a sizeable problem, with a large number of households having ℎ � under zero. Thus, we do not provide Chaudhuri estimation in these countries in Table B. 2. Estimations indicate that the Chaudhuri procedure would increase vulnerability in the case of the threshold of 0.50 by 1 p.p. in Mexico, 3 p.p. in Peru, 4 p.p. in Honduras, and 14 p.p. in Ecuador. In Brazil, the Chaudhuri estimation is more prominent than in other countries. Brazil's estimations are computed at the province level; thus, the measurement errors and unobserved income components can be more extensive and lead to a greater inaccuracy than at the district level. Table B. 2. Vulnerability rates by Chaudhuri procedure when the specification includes all hazards together. P(y/H,E,C,D) Chaudhuri Conditional to hazards and Var(e) conditional to H, E, events C, D % HH with Mean SD Mean SD Brazil predicted prob. > 0.25 25% 0.431 95% 22% predicted prob. > 0.50 16% 0.368 72% 45% predicted prob. > 0.75 10% 0.299 30% 46% Mexico predicted prob. > 0.25 38% 0.485 51% 0.500 predicted prob. > 0.50 20% 0.403 21% 0.406 predicted prob. > 0.75 9% 0.287 5% 0.208 Peru predicted prob. > 0.25 37% 0.482 53% 0.499 predicted prob. > 0.50 24% 0.425 27% 0.443 predicted prob. > 0.75 14% 0.342 7% 0.253 41 Dominican Rep. predicted prob. > 0.25 18% 0.386 n.a. n.a. predicted prob. > 0.50 7% 0.259 n.a. n.a. predicted prob. > 0.75 3% 0.163 n.a. n.a. El Salvador predicted prob. > 0.25 34% 0.474 n.a. n.a. predicted prob. > 0.50 17% 0.373 n.a. n.a. predicted prob. > 0.75 7% 0.255 n.a. n.a. Honduras predicted prob. > 0.25 60% 0.491 77% 0.420 predicted prob. > 0.50 41% 0.492 45% 0.498 predicted prob. > 0.75 25% 0.431 17% 0.376 Ecuador predicted prob. > 0.25 31% 0.461 58% 0.494 predicted prob. > 0.50 14% 0.346 28% 0.448 predicted prob. > 0.75 5% 0.228 10% 0.305 Source: Authors' calculations. We performed another robustness check estimating a restricted model where we set the hazards’ coefficients as positive. The restricted non-linear model for variables of hazards is 1 ℎ,, = + 1 , + ⋯ + , + , + ℎ,, + ℎℎ,, + + ℎ,, , where the positive coefficients are 1 , … , . Table B.3 presents the vulnerability rates for the Probit and restricted models. Table B. 3. Vulnerability rate from different models considering the specification with all hazards together. Vulnerability Probit model Restricted model incidence % HH % HH Country Mean SD Mean SD predicted prob. > 0.50 Brazil 16% 0.368 14% 0.345 Mexico 20% 0.403 18% 0.384 Peru 24% 0.425 22% 0.412 Dom. Rep. 7% 0.259 4% 0.198 El Salvador 17% 0.373 13% 0.340 Honduras 41% 0.492 39% 0.487 Ecuador 14% 0.346 12% 0.321 Source: Authors' calculations. 42 Figure B.1 permits visualization of the high correlation between the vulnerability rates at the state/district level from both models. Figure B.1. Vulnerability rates at the unit level from different models. Note: the x-axis indicates vulnerability rates from the restricted model, and the y-axis indicates the vulnerability rates from the Probit model. In Brazil and Ecuador, the units correspond to states, and the rest of the countries correspond to districts. Source: Authors' calculations. 43 Appendix C: Vulnerability profiles Figure C.1. Vulnerability curves based on different predicted probability thresholds by poor status. 1.00 1.00 Non-poor Brazil Non-poor Dom. Rep. 0.80 Poor Droughts 0.80 Landslides Poor Percentage of HH Percentage of HH 0.60 0.60 0.40 0.40 0.20 0.20 0.00 0.00 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 Probability predicted of being a vulnerable HH Probability predicted of being a vulnerable HH 1.00 1.00 Non-poor Ecuador Non-poor El Salvador 0.80 Poor Landslides 0.80 Poor Droughts Percentage of HH Percentage of HH 0.60 0.60 0.40 0.40 0.20 0.20 0.00 0.00 1 6 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 Probability predicted of being a vulnerable HH Probability predicted of being a vulnerable HH 1.00 1.00 Non-poor Honduras Non-poor Mexico 0.80 Poor Earthquakes 0.80 Poor Cyclones Percentage of HH Percentage of HH 0.60 0.60 0.40 0.40 0.20 0.20 0.00 0.00 1 6 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 Probability predicted of being a vulnerable HH Probability predicted of being a vulnerable HH 1.00 Non-poor Peru 0.80 Poor Floods Percentage of HH 0.60 0.40 0.20 0.00 1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 Probability predicted of being a vulnerable HH Source: Authors' calculations. 44 Table C.1. Percentage of households lacking coping mechanisms and adaptation strategies. Non- Vulnerable HH with prob. > Vulnerable 0.50 vulnerable All households HH with Countries HH with prob. > 0.50 Non-poor Poor prob. <= 0.50 Mean SD Mean SD Mean SD Mean SD Mean SD Brazil - Droughts HH head is not a salaried worker 61% 0.488 58% 0.494 74% 0.438 72% 0.447 74% 0.436 Only one sector of activity reported 69% 0.464 65% 0.478 86% 0.350 77% 0.422 88% 0.328 Only one income earner in HH 36% 0.481 34% 0.474 52% 0.500 49% 0.500 53% 0.499 HH head's main income from agriculture 6% 0.245 5% 0.220 14% 0.343 14% 0.347 14% 0.342 Presence of children <5 21% 0.406 16% 0.363 49% 0.500 42% 0.494 50% 0.500 Only elders (65 or more) in HH 8% 0.273 10% 0.296 0% 0.029 0% 0.052 0% 0.021 Presence of children <12 attending school 26% 0.437 20% 0.400 56% 0.496 51% 0.500 58% 0.494 Dwelling's materials of low quality 3% 0.177 2% 0.154 7% 0.257 5% 0.224 8% 0.264 Roof materials of low quality 3% 0.160 2% 0.149 4% 0.206 4% 0.198 5% 0.208 Walls' materials of low quality 5% 0.225 4% 0.204 10% 0.301 8% 0.273 11% 0.307 Benefits from social assistance 19% 0.394 10% 0.305 67% 0.469 57% 0.496 70% 0.459 No remittances from abroad - - - - - - - - - - At least one member without health insurance - - - - - - - - - - Dominican Republic - Landslides HH head is not a salaried worker 62% 0.486 63% 0.484 77% 0.422 70% 0.462 80% 0.404 Only one sector of activity reported 68% 0.468 67% 0.470 83% 0.379 72% 0.449 86% 0.346 Only one income earner in HH 38% 0.484 37% 0.483 50% 0.501 53% 0.501 49% 0.501 HH head's main income from agriculture 8% 0.278 8% 0.277 12% 0.329 11% 0.311 13% 0.335 Presence of children <5 24% 0.428 21% 0.405 64% 0.481 62% 0.489 65% 0.479 Only elders (65 or more) in HH 8% 0.273 9% 0.284 0% 0.000 0% 0.000 0% 0.000 Presence of children <12 attending school 30% 0.459 27% 0.443 67% 0.472 66% 0.477 67% 0.471 Dwelling's materials of low quality 6% 0.236 5% 0.218 17% 0.377 19% 0.391 17% 0.373 Roof materials of low quality 52% 0.499 49% 0.500 69% 0.465 62% 0.488 71% 0.455 Walls' materials of low quality 22% 0.413 19% 0.390 35% 0.479 30% 0.460 37% 0.484 Benefits from social assistance 26% 0.440 25% 0.430 33% 0.470 38% 0.487 31% 0.463 No remittances from abroad 89% 0.309 89% 0.312 89% 0.312 85% 0.362 91% 0.291 At least one member without health insurance 42% 0.494 39% 0.487 71% 0.453 72% 0.452 71% 0.453 Ecuador - Landslides HH head is not a salaried worker 57% 0.495 55% 0.498 68% 0.468 64% 0.481 69% 0.462 Only one sector of activity reported 58% 0.494 56% 0.497 70% 0.460 56% 0.496 75% 0.433 Only one income earner in HH 13% 0.333 13% 0.333 12% 0.330 13% 0.339 12% 0.326 HH head's main income from agriculture 24% 0.426 19% 0.390 55% 0.498 48% 0.500 58% 0.494 Presence of children <5 30% 0.457 25% 0.435 57% 0.495 50% 0.500 60% 0.490 Only elders (65 or more) in HH 9% 0.283 10% 0.296 3% 0.164 6% 0.238 1% 0.120 Presence of children <12 attending school 43% 0.495 39% 0.487 68% 0.465 53% 0.499 74% 0.437 Dwelling's materials of low quality 20% 0.398 15% 0.358 48% 0.500 40% 0.490 51% 0.500 Roof materials of low quality 62% 0.485 59% 0.491 80% 0.402 75% 0.431 81% 0.389 Walls' materials of low quality 16% 0.365 12% 0.325 39% 0.489 32% 0.465 42% 0.494 45 Benefits from social assistance 19% 0.389 13% 0.337 53% 0.499 42% 0.494 57% 0.496 No remittances from abroad 97% 0.170 97% 0.174 98% 0.146 97% 0.184 98% 0.128 At least one member without health insurance 82% 0.388 80% 0.402 92% 0.266 91% 0.281 93% 0.260 El Salvador- Droughts HH head is not a salaried worker 59% 0.491 58% 0.494 67% 0.469 69% 0.462 67% 0.472 Only one sector of activity reported 67% 0.470 64% 0.480 81% 0.391 73% 0.445 84% 0.364 Only one income earner in HH 43% 0.495 39% 0.487 65% 0.476 65% 0.477 66% 0.475 HH head's main income from agriculture 14% 0.351 10% 0.301 36% 0.479 33% 0.472 37% 0.482 Presence of children <5 26% 0.440 21% 0.405 54% 0.498 48% 0.500 57% 0.495 Only elders (65 or more) in HH 8% 0.268 9% 0.285 2% 0.136 3% 0.165 2% 0.124 Presence of children <12 attending school 34% 0.472 28% 0.450 61% 0.489 56% 0.496 62% 0.485 Dwelling's materials of low quality 26% 0.436 20% 0.397 55% 0.498 48% 0.500 57% 0.494 Roof materials of low quality 76% 0.425 78% 0.417 70% 0.459 73% 0.442 69% 0.464 Walls' materials of low quality 23% 0.420 18% 0.380 49% 0.500 43% 0.496 52% 0.500 Benefits from social assistance 2% 0.136 1% 0.100 6% 0.241 5% 0.210 7% 0.251 No remittances from abroad 78% 0.412 77% 0.419 83% 0.372 80% 0.398 85% 0.361 At least one member without health insurance 87% 0.333 85% 0.355 98% 0.146 95% 0.207 99% 0.113 Honduras - Earthquakes HH head is not a salaried worker 64% 0.479 59% 0.492 73% 0.443 74% 0.441 73% 0.444 Only one sector of activity reported 65% 0.478 57% 0.496 73% 0.443 63% 0.482 76% 0.429 Only one income earner in HH 32% 0.467 28% 0.449 40% 0.490 43% 0.496 39% 0.488 HH head's main income from agriculture 24% 0.426 9% 0.281 47% 0.499 32% 0.468 51% 0.500 Presence of children <5 36% 0.479 26% 0.440 51% 0.500 47% 0.500 52% 0.500 Only elders (65 or more) in HH 5% 0.209 3% 0.181 3% 0.163 4% 0.203 2% 0.151 Presence of children <12 attending school 43% 0.496 34% 0.474 59% 0.492 52% 0.500 61% 0.488 Dwelling's materials of low quality 11% 0.318 3% 0.160 24% 0.425 15% 0.359 26% 0.438 Roof materials of low quality 78% 0.414 82% 0.381 72% 0.451 74% 0.438 71% 0.454 Walls' materials of low quality 38% 0.486 19% 0.392 64% 0.481 50% 0.501 67% 0.469 Benefits from social assistance 12% 0.327 5% 0.219 23% 0.422 17% 0.372 25% 0.432 No remittances from abroad 83% 0.373 80% 0.402 87% 0.337 82% 0.381 88% 0.325 At least one member without health insurance - - - - - - - - - - Mexico - Cyclones HH head is not a salaried worker 43% 0.495 41% 0.492 55% 0.497 53% 0.499 56% 0.496 Only one sector of activity reported 62% 0.486 61% 0.489 67% 0.470 57% 0.496 71% 0.454 Only one income earner in HH 30% 0.457 30% 0.457 30% 0.459 33% 0.469 29% 0.455 HH head's main income from agriculture 13% 0.334 7% 0.255 36% 0.480 29% 0.452 38% 0.486 Presence of children <5 27% 0.442 21% 0.405 50% 0.500 47% 0.499 51% 0.500 Only elders (65 or more) in HH 7% 0.256 7% 0.254 8% 0.270 8% 0.278 8% 0.266 Presence of children <12 attending school 39% 0.489 33% 0.471 63% 0.482 57% 0.495 65% 0.476 Dwelling's materials of low quality 26% 0.437 18% 0.381 58% 0.494 50% 0.500 61% 0.489 Roof materials of low quality 22% 0.415 15% 0.354 51% 0.500 43% 0.495 54% 0.498 Walls' materials of low quality 10% 0.305 7% 0.249 25% 0.430 19% 0.394 26% 0.441 Benefits from social assistance 29% 0.453 21% 0.406 62% 0.486 54% 0.498 64% 0.480 No remittances from abroad 95% 0.210 96% 0.202 94% 0.245 91% 0.285 95% 0.228 46 At least one member without health insurance 34% 0.473 34% 0.473 35% 0.477 39% 0.487 33% 0.472 Peru - floods HH head is not a salaried worker 66% 0.473 61% 0.488 86% 0.348 80% 0.402 88% 0.330 Only one sector of activity reported 57% 0.495 52% 0.500 75% 0.432 65% 0.479 78% 0.414 Only one income earner in HH 32% 0.468 29% 0.454 43% 0.496 47% 0.499 42% 0.494 HH head's main income from agriculture 27% 0.442 15% 0.358 67% 0.471 54% 0.499 70% 0.457 Presence of children <5 25% 0.435 21% 0.408 36% 0.481 35% 0.478 36% 0.481 Only elders (65 or more) in HH 11% 0.312 10% 0.303 14% 0.351 14% 0.342 15% 0.354 Presence of children <12 attending school 30% 0.458 26% 0.439 41% 0.492 40% 0.490 42% 0.493 Dwelling's materials of low quality 13% 0.332 11% 0.309 20% 0.401 18% 0.384 21% 0.405 Roof materials of low quality 9% 0.290 8% 0.268 8% 0.277 8% 0.269 9% 0.279 Walls' materials of low quality 45% 0.497 33% 0.471 82% 0.385 69% 0.463 85% 0.354 Benefits from social assistance 47% 0.499 36% 0.479 88% 0.322 83% 0.373 90% 0.306 No remittances from abroad 98% 0.136 98% 0.149 99% 0.075 99% 0.117 100% 0.060 At least one member without health insurance 42% 0.493 45% 0.498 28% 0.449 39% 0.487 25% 0.435 Source: authors' calculations. 47