Poverty Mapping in El Salvador 1 Poverty Mapping in El Salvador Poverty and Equity Global Practice Latin America Region Monica Robayo-Abril Britta Rude February 2023 2 © 2023 International Bank for Reconstruction and Development / The World Bank 1818 H Street NW Washington DC 20433 Telephone: 202-473-1000 Internet: www.worldbank.org This work is a product of the staff of The World Bank with external contributions. The findings, interpretations, and conclusions expressed in this work do not necessarily reflect the views of The World Bank, its Board of Executive Directors, or the governments they represent. The World Bank does not guarantee the accuracy of the data included in this work. The boundaries, colors, denominations, and other information shown on any map in this work do not imply any judgment on the part of The World Bank concerning the legal status of any territory or the endorsement or acceptance of such boundaries. Rights and Permissions The material in this work is subject to copyright. Because The World Bank encourages dissemination of its knowledge, this work may be reproduced, in whole or in part, for noncommercial purposes as long as full attribution to this work is given. Any queries on rights and licenses, including subsidiary rights, should be addressed to World Bank Publications, The World Bank Group, 1818 H Street NW, Washington, DC 20433, USA; fax: 202-522-2625; e-mail: pubrights@worldbank.org Translations—If you create a translation of this work, please add the following disclaimer along with the attribution: This translation was not created by The World Bank and should not be considered an official World Bank translation. The World Bank shall not be liable for any content or error in this translation. Adaptations—If you create an adaptation of this work, please add the following disclaimer along with the attribution: This is an adaptation of an original work by The World Bank. Views and opinions expressed in the adaptation are the sole responsibility of the author or authors of the adaptation and are not endorsed by The World Bank. Third-party content—The World Bank does not necessarily own each component of the content contained within the work. The World Bank therefore does not warrant that the use of any third-party-owned individual component or part contained in the work will not infringe on the rights of those third parties. The risk of claims resulting from such infringement rests solely with you. If you wish to re-use a component of the work, it is your responsibility to determine whether permission is needed for that re-use and to obtain permission from the copyright owner. Examples of components can include, but are not limited to, tables, Figures, or images. 3 Acknowledgments This report was produced under the World Bank El Salvador FY22 and FY23 Poverty and Equity Program and the El Salvador Systematic Country Diagnostic Update as one of the knowledge products that aim to contribute to a better understanding of the distribution of poverty at a subnational level in El Salvador. The poverty maps were produced as a collaboration between the World Bank and El Salvador's General Directorate of Statistics and Censuses (DIGESTYC). This work was prepared under the guidance of Michel Kerf (Country Director, ECCEU) and Ximena Del Carpio (Practice Manager, ELCPV). The analysis was conducted by a team including Monica Robayo-Abril (Senior Economist, TTL, ELCPV) and Britta Rude (Consultant, ELCPV). The team is grateful for the comments and advice from Paul Corral (Senior Economist, GGHVP). 4 Executive Summary Poverty mapping1 -- the spatial representation and analysis of human wellbeing and poverty indicators -- is becoming an increasingly important instrument for investigating and discussing socioeconomic issues, informing targeting efforts, and guiding the geographic allocation of resources. One approach to addressing poverty is the geographic approach. In the geographic approach, poor people are identified and targeted through poverty maps. Indeed, the geographical approach is one of the methods used worldwide for targeting anti- poverty programs to reduce the gaps in social protection coverage of poor and vulnerable groups, and it has been widely implemented in several countries around the world. In 2020, the Salvador's General Directorate of Statistics and Censuses (DIGESTYC) and the World Bank started working on the Project ‘Poverty mapping in El Salvador‘. The Project is part of the Government and International Bank for Reconstruction and Development (IBRD) Programme, which is performed by experts of the National Statistical Institute (NSI) and the World Bank (WB). The main objective is to calculate the shares of households living in moderate and extreme poverty at disaggregated territorial levels (municipalities). Poverty mapping enhances our understanding of the geographic distribution of people living in poverty This report presents poverty maps at the municipality level based on the Fay- Herriot model for small-area estimations. Direct estimates of poverty indicators at the municipality level rely on information generated from household surveys. Often, though, household surveys are not representative at disaggregated levels, such as municipalities. Consequently, small sample sizes limit their precision and estimates cannot be obtained for out-of-sample domains. Due to this, we resort to small-area estimation techniques, which rely on several data sources to improve the precision of survey-based direct estimates. For the case of El Salvador, we use data from the last available Population Census conducted in 2007 and the 2019 household survey (Encuesta de Hogares de Propositos Multiples, EHPM). We also draw from population projections at the municipality level, as El Salvador is subject to high emigration rates. Many methodologies for poverty mapping require that reference years of the data sources used as a basis for small area estimations are as close to each other as possible. Due to the fact that the last available census is from 2007, we decided to use small area 1Poverty maps rely on small area estimates of poverty. Small area estimates are based on statistical methods to improve the precision of survey estimates in geographical areas in which survey estimates lack sufficient precision. For a more detailed description of small area estimates, see Rao and Molina (2015). 5 estimation techniques based on the Fay-Herriot model, which is the most appropriate model in this case. Our results show that poverty varies at the municipality level in El Salvador. To measure poverty, we follow the national methodology defined by DIGESTYC. Additionally, we measure poverty at the household level. We generate poverty maps at the municipality for monetary poverty and multidimensional poverty. This report presents the results for the moderate and extreme poverty rate, poverty severity and poverty gaps, as well as multidimensional poverty. All poverty indicators analyzed in this report point towards the concentration of poverty in specific country areas. While there is a certain variation in the ranking of poverty across the different indicators investigated, the poorest municipalities are concentrated in the Northeast and West of El Salvador. The poverty maps are an important contribution to the country's agenda to eliminate poverty. The generated poverty maps can be used for geographic targeting programs. They can also be combined with complementing targeting mechanisms or additional data sources to design targeted policies. The methodology applied considers that the latest Census is from 2007 and might not be a good mirror of the current status quo in the country. The maps are, therefore, an important contribution to El Salvador's agenda to eliminate poverty in the country. 6 Contents Acknowledgments ............................................................................................................................................. 3 Executive Summary........................................................................................................................................... 5 1. Motivation and Scope .............................................................................................................................. 8 2. Estimating small-area poverty indicators for El Salvador ...............................................................11 Methodology ................................................................................................................................................14 Data ...............................................................................................................................................................16 Small-area Monetary Poverty Estimates and Comparisons with previous Poverty Maps ..............19 Multidimensional Poverty Maps ...............................................................................................................24 How could these updated maps be used to improve the targeting of social programs and inform the poverty eradication strategy? ..............................................................................................................29 Caveats and limitations ..............................................................................................................................31 3. Conclusion ...............................................................................................................................................31 References.........................................................................................................................................................33 Annex ................................................................................................................................................................36 Annex 1 – Choosing the best model among several model specifications........................................36 Annex 2 – FH Estimates of Extreme Poverty Indicators ....................................................................45 Annex 3 – FH Estimates of Poverty Gaps .............................................................................................48 Annex 4 – FH Estimates of Poverty Severity ........................................................................................50 Annex 5 – FH Estimates of Multidimensional Poverty .......................................................................52 Annex 6 – Comparison of municipality rankings of old and new maps (From highest to lowest poverty incidence) .......................................................................................................................................53 Annex 7 – Small area estimates of poverty (Fay-Herriot) ....................................................................62 Annex 8 – Poverty Maps using Poverty Headcount Ratios .................................................................70 7 1. Motivation and Scope For effective policymaking, it is often necessary to obtain information about poverty at disaggregated geographic levels, such as the municipality level. This information can be elevated from census data or administrative data. One challenge is that census data is often only collected roughly every ten years, and administrative data is often nonexistent in developing countries or protected by privacy regulations. Household data, which is gathered more frequently and is more accessible in developing countries, is usually not statistically representative at these disaggregated levels. This has led to a surge in small area poverty estimations. These estimations consider information from alternative data sources, such as census data or satellite data, or analyze the precision of information within an existing dataset, to generate small area estimations of income and income-related indicators. This document generates updated small-area poverty estimates for El Salvador and describes the underlying methodology and validity of the resulting estimators. Given the limitations of household survey to generate insights on poverty at disaggregated geographical levels, such as municipalities, as well as the absence of an updated census, this document applies a small area poverty estimation technique to municipalities in El Salvador. Specifically, we apply a well-established empirical methodology to generate poverty headcount ratios at the municipality level. We show that these estimates outperform poverty headcount ratios observed directly from survey data due to methodological improvements. Updated poverty maps are critical to improving the targeting of anti-poverty programs that use a geographical approach to reduce the gaps in social protection coverage of poor and vulnerable groups. Given the broad geographic disparities in poverty in some countries, some governments have shifted their approaches to fight poverty and social exclusion using this approach, which can also be combined with other targeting methods. Poverty maps can inform geographical targeting, which focuses on the most impoverished areas of the country. Additionally, poverty maps can be useful for the analysis of existing programs or resource allocation mechanisms. They can help to assess their effectiveness, too. The geographical approach has been applied in several Latin American countries. This approach has been used to extend social protection to the poor, indigenous populations, and ethnic minorities. One example is the "Red de Oportunidades Scheme" in Panama. The program is a cash transfer scheme designed to reduce extreme poverty, with a specific component for rural and indigenous areas. The scheme was initially rolled out in regions with a larger share of indigenous populations and was subsequently extended to indigenous 8 populations living in urban areas and poor non-indigenous populations. The share of indigenous beneficiaries increased from 36 percent in 2007 to 58 percent in 2012 (Robles, 2009). Efficiency can augment through the geographic approach and leakage to the non-poor can be minimized by targeting smaller areas, such as municipalities. Box 2 presents an overview of countries applying poverty maps. Using geographic targeting in El Salvador over other methods of poverty alleviation has several advantages. First, it offers a decisive criterion for identifying target groups (i.e., targeting all households living in municipalities with higher- than-average extreme poverty rates or extreme poverty rates above a determined threshold). Second, it is possible to combine the criteria of geographic location (i.e., municipalities with high extreme poverty rates) with other socioeconomic characteristics of households and individuals (municipalities located in states with a significant lack of access to health and education). Additionally, geographic targeting involves local authorities in program monitoring, and can assist in the allocation of social welfare benefits and regional-development resources (Bigman and Fofack, 2000). Finally, it is administratively simple; it creates no labor disincentives, is unlikely to generate stigma effects, and is easy to combine with other targeting methods. To reduce the cost of poverty reduction programs further, geographical targeting should be combined with other targeting methods within areas, such as targeting based on individual or household characteristics associated with poverty. Targeting strategies might combine alternative methods and strategies (see Box 1). The literature suggests that this approach further increases targeting efficiency2. Proxy-means tests are one of the few methods available to target chronically poor households effectively, along with demographics, geographical and community-based targeting, and self-selection. Improvements are possible in the design of tools for proxy-means testing. The potential exists to enhance the performance of targeting by combining proxy- means tests with other targeting methods. For example, the "Orphans and Vulnerable Children Program" in Kenya and the "Prospera Program" in Mexico combine geographical targeting and proxy-means testing. Brazil's "Bolsa Família" relies on geographical targeting and means testing. To give another example, geographical targeting, combined with community-based targeting and proxy- means testing, is used in Tanzania. In a well-designed process, multiple methods can bring together complementary strengths to minimize errors of exclusion and inclusion. 2 Grosh, M., Del Ninno, C., Tesliuc, E., & Ouerghi, A. (2008). For protection and promotion: The design and implementation of effective safety nets. World Bank Publications. 9 Poverty maps need to be updated to reflect the changing welfare of households over time. This is particularly important in the case of El Salvador, where several social programs use geographic targeting as one of their criteria, and updated poverty maps can significantly improve targeting. One program using this approach is the "Rural Solidarity Communities" (RSC). The RSC is a cash transfer program based on public education and health services usage in households in the poorest 100 of the country's 262 municipalities, according to the 2004 Social Investment Fund for Local Development (FISDL) Poverty Map. Households are eligible if they meet several criteria, including geographic criteria, captured when the program starts in their community. In rural areas, all households in a municipality, that met the eligibility requirements at the time when the Population Census was conducted by the implementing agency (FISDL), were registered in the program. All eligible households entered the program in municipalities with "severe" extreme poverty in urban areas. However, a proxy-mean test is applied to selected beneficiaries in urban municipalities with "high" extreme poverty. The non-contributory "Universal Basic Pension" (Adulto Mayor) is a program for older adults in municipalities with "severe" and "high" extreme poverty, and also relies on the FISDL poverty maps. Another example is the "Temporary Income Support Program" (PATI). PATI was designed to protect the income of vulnerable households that face adverse situations of various kinds. The program is implemented in informal urban settlements (AUP) classified with extreme or high poverty levels according to the 2004 FISDL Poverty Maps. These examples show the importance of disposing of updated poverty maps in El Salvador. Box 1: Combining Geographic Targeting with alternative targeting methods. There are 6 different possibilities to target beneficiaries of policy programs: Means testing, proxy means testing, categorical testing, geographic targeting, self-targeting and community-based targeting (GIZ, 2019). Several social protection programs have used combined targeting methods to target potential beneficiaries. One example is the Oportunidades Program in Mexico, which combines geographic targeting and proxy means testing. So does Kenya’s Orphans and Vulnerable Children program. The Bolsa Familia program in Brazil uses geographic targeting and means testing. In Tanzania, the geographic targeting strategy is combined with community-based targeting and proxy means testing. The Bangladesh Rural Advancement Committee (BRAC) uses a combined targeting approach: First, the targeting process identifies geographic locations with a high concentration of ultra- poor households. Next, it applies a participatory wealth ranking of households. Lastly, program staff uses a questionnaire to determine the final selection of beneficiaries. 10 Going beyond targeting, the geographical approach can also help to inform sectoral interventions and inform subnational budget allocation. Poverty can be used as a criterion to identify target areas. A cutoff score can be used as a criterion to identify locations ("municipalities") eligible for a particular program; funding formulas can also be designed to vary benefit levels across the entire range of poverty scores. This method can also help target sectoral investments and transfers from the government budget, donor support, and other funds to those municipalities with the highest estimated level of poverty incidence and social exclusion. In El Salvador, subnational resource allocation from central to local governments ("FODES transfers") uses poverty at a municipal level as one of the criteria. Salvadoran municipalities receive their leading financial resource from central state grants. The largest transfer is allocated through the Fund for economic and social development of the municipalities of El Salvador (Fondo para el Desarrollo Económico y Social de las Municipalidades - FODES), which allocates 6% of the national budget to municipal governments, among which 80% are allocated for investment and 20% for operating expenses. The existing formula for FODES resource allocation established in the law is based on four criteria: population, poverty, equity, and land area with the following percentages: 50%, 25%, 20%, and land area 5%, respectively. Municipalities receive an increasing amount with the poverty rankings, as established by the 2004 Poverty maps, by decile. The report at hand is organized as follows. Section 2 documents the methodology used to produce recent small area estimates in El Salvador and the associated methodological challenges. The small area-estimates of poverty are presented, and the ranking of municipalities based on the new poverty maps and the old poverty maps are compared. It also presents a multidimensional poverty indicator for El Salvador. Section 3 concludes. 2. Estimating small-area poverty indicators for El Salvador The poverty maps presented in this report are produced as a result of a collaboration between the World Bank and El Salvador's General Directorate of Statistics and Censuses (DIGESTYC).3 Despite the usefulness of poverty maps for designing poverty targeting strategies, to the best of our knowledge, there are not any up-to-date poverty maps available for El Salvador. The following section describes the methodology used to produce such a poverty map at the 3Poverty maps rely on small area estimates of poverty. Small area estimates are based on statistical methods to improve the precision of survey estimates in geographical areas in which survey estimates lack sufficient precision. For a more detailed description of small area estimates, see Rao and Molina (2015). 11 municipality level for El Salvador from the 2019 official household survey (EHPM) and the Population Census from 2007. We also use information on population projections, provided by DIGESTYC. The resulting small area estimates, on which the maps are based, can be combined with alternative data sources to generate informed and data-based policymaking. Available data sources in El Salvador are not representative at the disaggregated level. In developing countries, auxiliary information is often not available at the unit level, and area-level models are therefore the better choice. This is the case in El Salvador, where the latest Population Census is from 2007. Yearly household surveys, on the other hand, are not only representative at the departmental level Figure 1 shows the results for average moderate poverty rates at the household level per department. The map reveals that there is significant variation in poverty at the departmental level, with Morazán having the highest moderate poverty rate (28.06 percent). The country’s data landscape leaves the country without updated poverty estimates at the municipality level. FIGURE 1: MODERATE POVERTY RATES AT THE DEPARTMENT LEVEL Notes: The graph plots poverty rates at the household level per department. Source: EHPM (2019). We apply a widely used methodology for Small Area Estimation (SAE), the Fay- Herriot Model. There are several different methodologies available to produce SAE. For an overview of the different methodologies, see Guadarrama et al. (2016).4 One can distinguish between unit-level models (such as work by Elbers et al. (2003), for example), and area-level models (such as the Fay-Herriot Model) (Eurostat, 2019). The difference is that unit-level models are estimated at the level 4Guadarrama, María, Isabel Molina, and J. N. K. Rao. "A comparison of small area estimation methods for poverty mapping." Statistics in Transition new series 1.17 (2016): 41-66. 12 of a single unit (such as a household) in a first step. In a second step, unit-level models apply model parameters estimated from household surveys to population census and generate a welfare vector for every single household in the census. From the simulated welfare vector, indicators of interest are obtained at the desired geographical level, while area-level models are directly estimated at the geographic level of interest. Given the current data sources available in El Salvador, an area-level model is therefore more appropriate. The Fay Herriot model is one of the most used area-level models. Instead of solely relying on past information from data sources with higher representation, such as the Population Census from 2007, the Fay-Herriot model also considers the varying level of precision of different domains present in national household surveys, such as the EHPM 2019. We estimate poverty maps in El Salvador for national poverty indicators, relying on national poverty lines at the household level. We produce maps for the following poverty indicators: The national poverty rate, defined as the share of households whose disposable income is below the national poverty threshold, as well as the national extreme poverty rate, defined as the share of households whose disposable income is below the national extreme poverty threshold. We additionally estimate maps for poverty severity and poverty gaps. Lastly, we produce small area poverty estimates of multidimensional poverty at the household level, following the national definition of multidimensional poverty. The maps are produced at the municipality level for the year 2019. Box 2: Poverty maps around the world. A variety of countries have produced poverty maps. In the LAC region those are Colombia, Guatemala, Honduras, Jamaica, Nicaragua, Peru and St. Lucia. Colombia has put the poverty maps at use to increase the targeting efficiency of social protection programs, as well as of private investments in social projects. The maps have also encouraged synergies between private and public agencies to reduce multidimensional poverty. In Nicaragua, poverty maps have informed fund allocations across municipalities along several sectors. Nicaragua’s Emergency Social Investment Fund (Fondo de Inversion Social de Emergencia, FISE) uses a poverty map to target the poor. Outside the LAC region, a variety of countries have published poverty maps, reaching from Bulgaria over Egypt to Nepal. The poverty maps form part of the Third National Development Plan in Uganda. In Zimbabwe, they are used to allocate resources geographically. In Burundi, they have informed the targeting of beneficiaries of a social safety net program. 13 Methodology One method of small area estimations is based on the Fay Herriot model, initially developed by Fay and Herriot (1979)5. This model has recently gained increased attention in academia and by statistical offices, as well as research institutions and international organizations, due to the model’s high precision. The Fay- Herriot model's underlying idea is that those small areas with low precision "borrow strength" from small areas with high precision.6 The Fay-Herriot model is a combination of a sampling and linking model, which we describe in detail in this section. The linking model refers to the part of the model, which approximates the relationship between auxiliary information and the outcome variable of interest . Only considering this part of the model results in the following equation, which creates a linear relationship between the outcome variable of interest and a number i of auxiliary variables at the domain level : = + , = 1, … , . , where are the fixed effects of auxiliary variables i, and are random effects. are assumed to be independent and identically distributed (iid) with mean zero and variance 2 . The underlying assumption is that the variance parameter is known. In practice, it can be estimated via likelihood-based methods, such 2 as the Restricted Maximum Likelihood (REML), or the Maximum Likelihood (ML). The downside is that these methods depend heavily on the distributional assumptions behind the sampling errors and the random effects . Importantly, the equation above cannot be estimated, as is not observed. In our empirical application, would be the true poverty rate at the municipality level, for example. For this reason, the Fay-Herriot model approximates these indicators by a sampling model. The sampling model refers to the part of the model, which relies on direct estimates at the area-level of interest from non-representative surveys. A sampling model occurs when only relying on direct estimates observed in the underlying data. A sampling model can be described by the following equation: = + , = 1, … , . In this case, there is no auxiliary information included, as the estimates are only based on information from the survey. is the sampling error, as is estimated inprecisely due to low sample size and non-representativeness at the geographic level of interest. This means that sampling errors arise, as the direct estimator from 5 Fay, R.E. and Herriot, R.A. (1979). Estimates of income for small places: an application of James- Stein procedures to census data. Journal of the American Statistical Association,74, 269-277. 6 Molina and Morales (2009). Small area estimation of poverty indicators. Boletín de Estadística e Investigación Operativa. Vol. 25, No. 3, Octubre 2009, pp. 218-22. 14 the survey is not equal to the true underlying variable of interest . The underlying assumption is that sampling errors are normally distributed with mean zero and variance 2 2 . can be directly estimated from the survey data at the geographic level of interest. The sampling variance likely differs at the domain level. Some domains might be subject to a more significant spread in the data than others. Combining the sampling with the linking model results in a linear mixed model. The resulting model is a linear mixed model of the following form: = + + , = 1, … , . , where is the estimator7 of the true mean of the variable of interest (e.g., the poverty rate) at the level of interest (e.g., the municipality), is a set of auxiliary variables linearly related to the outcome of interest at the area level of interest (e.g., the share of working population at the municipality level), are independent error terms with zero means and unknown constant variance, and are the sampling errors, which are independent with zero mean and heteroskedastic known variance. D is the number of domains (the respective areas of interest). In practice, the estimated variance of the direct estimators for is used frequently as the known error variance. 8 The Fay-Herriot model generates the best linear unbiased predictor (BLUP). When 2 is known, the Fay-Herriot model generates the BLUP of the true mean at the domain level of interest by applying a shrinkage factor , which gives higher weight to domains measured with higher precision. 9 The Mean Squared Error (MSE) of the BLUP is then always at least as efficient as the one of the direct estimator. This means that the Fay-Herriot model minimizes the MSE. The BLUPs gain in efficiency especially for areas with larger sampling variance. The BLUP is given by: ̃ , = 1, … , . = + ( 1 − ) , with the shrinkage factor being: 2 = 2 + 2 7 A direct estimator is an estimator based solely on the observed sample data in the corresponding domain (e.g., the mean income at the municipality level generated from sample data from this municipality). 8 You, Y., and B. Chapman. 2006. Small area estimation using area-level models and estimated sampling variances. Survey Methodology32: 97–103. 9 The shrinkage factor is the proportion of variance due to ud (accounting for between area variations). 15 The shrinkage factor depends on the error variance and the unexplained variation, accounting for precision and model strength. A detailed look at the composition of the shrinkage factor makes clear that it decreases with the error variance 2 . Consequently, the higher the error variance, the lower the share of BLUPs, which results from the direct estimates. This rationale is based on direct estimates with a higher variance being more imprecisely measured. Therefore, they receive less weights (importance) in the linear mixed model. At the same time, the estimates generated from the linking model part receive higher weights in these cases. Similarly, with growing unexplained variation 2 , the shrinkage factor increases. Therefore, the weaker the linear mixed model (the higher its unexplained variation), the higher the weights given to the direct estimates observed in the underlying survey data. To summarize, he shrinkage factor accounts for model strength and precision. and generates the Empirical BLUPs (EBLUPs). Using empirical estimators for The EBLUPs are more precise than the direct estimators always if the chosen model fits the underlying data well. This is an important caveat of the Fay-Herriot model. Evaluating the model fit is therefore a crucial step in the analysis. While the BLUP estimator requires normality, the EBLUP does not. This is highly beneficial in the context of development countries, as poverty is often highly concentrated, and normality might not be fulfilled. An important caveat is that the the Fay- Herriot model requires linearity in its linking model component. We estimate several model specifications of the Fay-Herriot model at the municipality level. Given that the Fay-Herriot model estimation takes place at the area-level of interest, it requires one line of data per area (see the next Section for a detailed overview of the underlying data used in this report). This approach requires aggregating data to the area-level of interest and running the empirical estimation at this level. Importantly, we run several different model specifications and choose the model with the lowest coefficient of variation (CV), which in our case is the Fay-Herriot model with ampl specification (for details, see Annex 1). The different model specifications we apply are: • Variance estimation using restricted maximum likelihood (REML) • Variance estimation using adjusted maximum-profile likelihood (AMPL) • Variance estimation using adjusted residual maximum-likelihood (ARYL) • Log-transformed estimation of direct estimators and corresponding variances, with and without out of sample predictions • Arcsin transformation, which confines the EBLUPs to a [0;1] interval. Data Given the current data landscape in El Salvador, the Fay-Herriot estimation is the appropriate small-area estimation method for this setting. There is currently no 16 representative, up-to-date information on poverty indicators available in El Salvador. The government conducted its last official population census in 2007. Since then, the country has been marked by significant structural changes as an outflow of nearly one-fourth of its population. By mid-2020, the IOM registers 1.6 million emigrants from El Salvador, compared to a total population of 6.5 million.10 Small-area estimation methods relying solely on information from the Census might therefore not be accurate as they do not present a true mirror of the country's current status quo anymore. The Fay-Herriot Model corrects for this shortcoming through the empirical approach described above, as it also considers more recent information through the sampling model. In our case, we combine information from the household survey (our direct estimates) with information from the Population Census (our auxiliary area-level information) by incorporating the area-level information from the census into the linkage model, and the direct estimates from the survey into the sampling model. We rely on several data sources, one being the official household survey from 2019. The 2019 official household survey used for the estimation of the poverty maps is representative at the urban and rural level, for the metropolitan area of San Salvador, the department level, and 50 self-representative municipalities. While El Salvador has conducted a continuous household survey every year since 1975 (EHPM), the survey is not representative at the municipality level. The household survey of 2019 consists of a total sample of 19,968 housing units, 21,326 households, and 74,435 individuals. The EHPM (2019) was conducted monthly, from January to December of 2019. There are 14 departments and 262 municipalities in El Salvador. Two hundred twenty-six municipalities are included in the EHPM (2019). In most survey rounds, including the 2019 round11, 50 of the 226 sampled municipalities are self-representative. The survey includes a primary sampling unit and stratification. We account for this design in our estimation of the poverty maps. Figure 2 shows the distribution of poverty rates at the municipality level from the 2019 EHPM. The only variable from the national household survey is our poverty estimate. Including independent variables from the household survey would bias our SAEs and lead to a higher random error.12 FIGURE 2: HISTOGRAM OF POVERTY RATES AT THE MUNICIPALITY LEVEL (2019) 10 Source: Migration Data Portal (2018). Link: https://migrationdataportal.org/data?i=inflow_work&t=2018&cm49=340 11 This is not the case for the 2020 household survey. 12Szymkowiak, Marcin, Andrzej Młodak, and Łukasz Wawrowski. "Mapping poverty at the level of subregions in Poland using indirect estimation." STATISTICS 609 (2017). 17 Notes: The graph plots a histogram of the national poverty headcount ratio at the municipality level relying on data from the national household survey from 2019. The x-axis reports the municipal national poverty rate and the y-axis the density. Source: EHPM (2019). We draw auxiliary information from the 2007 Population Census as well as population estimates provided by DIGESTYC. All explanatory variables included in the model are from the 2007 Census with the exception of the population estimates at the municipality level, provided by DIGESTYC. We include the following independent variables: • Household-level variables: If the household owns a car, house, radio, washing machine, has access to water, electricity, and sanitation, as well as the number of household members. • Condition of housing: The condition of the housing households live in. • Labor market and education variables: Labor market activity, the share of self-employed, the share of entrepreneurs, the share of public-sector workers, the share of the population with at least primary education, the share of children attending school, and the literacy rate. • Population characteristics: Population estimates from 2020 and the number of children in a municipality. We aggregate all datasets at the municipality level. Given that the Fay-Herriot model is a model performed at the area-level of interest, in our case municipalities, we aggregate all datasets at the municipality level. First, we aggregate the underlying household data at the municipality level. To do this, we consider the sampling design of the survey. Next, we aggregate our variables of interest from the census data at the municipality level. Lastly, we combine all data sources at the municipality level and run our model estimation at this same level of analysis. 18 Small-area Monetary Poverty Estimates and Comparisons with previous Poverty Maps This subsection estimates the Fay-Herriot model to construct small area estimates of monetary income poverty at the household level using the national poverty lines for El Salvador.13 We employ 5 different model specifications of the Fay-Herriot model and choose the best-performing model specification. We compare the performance of our 5 model specifications to each other based on two different criteria: the CVs and the MSEs. These comparisons indicate that the model specification with variance estimation using adjusted maximum-profile likelihood (AMPL) is the best performing model (for details, see Annex 1). We next report the results and poverty maps relying on this model specification. We apply the Fay-Herriot model specifications to estimate the extreme poverty rate, the poverty rate, as well as poverty severity, and poverty gap. Figure 2 plots the direct estimator of the poverty rate at the municipality level versus the EBLUP estimators of the extreme poverty rate. Table 1 gives an overview of the direct and Fay-Herriot estimator for the municipal level's national poverty rates. We then estimate a map of the poverty rate (figure 3), extreme poverty rate (figure 4), poverty severity (figure 5), and poverty gap (figure 6). Table 2 shows the results for the small-area estimation of the extreme poverty rate at the municipality level, table 3 for poverty severity, and table 4 for the poverty gap. FIGURE 3: SMALL AREA ESTIMATES VERSUS DIRECT ESTIMATES OF POVERTY RATES AT THE MUNICIPALITY LEVEL (FH MODEL WITH AMPL ESTIMATION) Notes: The graph shows a scatter plot of poverty estimates (EBLUPs) generated from a Fay-Herriot model (on the x-axis) and direct estimates relying only on household survey data (on the y-axis). 13 For the poverty maps using headcount ratios see the Annex 8. 19 The EBLUPs are generated from a model specification with variance estimation using adjusted maximum-profile likelihood (AMPL). Source: EHPM (2019) and Census (2007). TABLE 1: FH MODEL WITH AMPL ESTIMATION – POVERTY RATES (2019) Variable Obs Mean Std. Dev. Min Max Direct estimator 226 20.87 9.87 0 66.62 CV (Direct estimator) 225 1768.8 1722.97 0 9018.58 EBLUP estimator 262 21.09 8.87 0 66.62 CV (FH Model) 261 1878.94 1597.07 0 11243.48 MSE EBLUP 262 .22 .29 0 1.12 Notes: The table shows summary statistics for moderate poverty rates at the municipality level in El Salvador. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 municipalities in El Salvador. We employ a variance estimation using adjusted maximum-profile likelihood (AMPL). Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). TABLE 2: FH MODEL WITH AMPL ESTIMATION - EXTREME POVERTY RATE (2019) Variable Obs Mean Std. Dev. Min Max Direct estimator 226 6.54 6.6 0 38.46 CV (Direct estimator) 187 3720.77 2743.43 0 10385.96 EBLUP estimator 262 6.45 5.92 -1.67 38.46 CV (FH Model) 223 3749.25 10325.05 -119066.39 62830.85 MSE EBLUP 262 .08 .11 0 .4 Notes: The table shows summary statistics for extreme poverty rates at the municipality level in El Salvador. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 municipalities in El Salvador. We employ a variance estimation using adjusted maximum-profile likelihood (AMPL). Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). TABLE 3: FH MODEL WITH AMPL SPECIFICATION - POVERTY SEVERITY (2019) Variable Obs Mean Std. Dev. Min Max Direct estimator 226 2.14 2.01 0 11.2 CV (Direct estimator) 217 3280.47 2597.42 0 10906.5 EBLUP estimator 262 2.06 1.77 -.15 11.2 CV (FH Model) 253 3124.12 15445.86 -134209.7 150595.84 MSE EBLUP 262 .01 .01 0 .04 Notes: The table shows summary statistics for poverty severity at the municipality level in El Salvador. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty 20 estimates for all 262 municipalities in El Salvador. We employ a variance estimation using adjusted maximum-profile likelihood (AMPL). Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). TABLE 4: FH MODEL WITH AMPL SPECIFICATION - POVERTY GAPS (2019) Variable Obs Mean Std. Dev. Min Max Direct estimator 226 5.07 4.06 0 22.95 CV (Direct estimator) 217 2771.69 2379.64 0 10906.5 EBLUP estimator 262 5 3.62 0 22.95 CV (FH Model) 253 3769.34 7463.48 0 80769.03 MSE EBLUP 262 .03 .04 0 .17 Notes: The table shows summary statistics for poverty gaps at the municipality level in El Salvador. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 municipalities in El Salvador. We employ a variance estimation using adjusted maximum-profile likelihood (AMPL). Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). Figure 3 shows that there is considerable variation at the municipality level with respect to poverty rates. While none of the municipalities has a poverty rate larger than 66.6 percent, there are some agglomerations with significant poverty rates in the Northeast and Southwest of the country. Special attention should be paid to these municipalities as they are characterized by a high concentration of the poor. Comparing Figure 1 and 4 to each other reveals that there is significant variation at the municipality level within departments. Not all of the poorest municipalities are located in the poorest department, and vice versa. 21 FIGURE 4: SMALL AREA ESTIMATES OF THE NATIONAL POVERTY RATE AT THE MUNICIPALITY LEVEL (2019) Source: World Bank estimates based on EHPM (2019) and Population Census (2007). The poverty rate is measured at the household level and reported in percent. FIGURE 4: SMALL AREA ESTIMATES OF THE NATIONAL EXTREME POVERTY RATE AT THE MUNICIPALITY LEVEL (2019) Source: World Bank estimates based on EHPM (2019) and Population Census (2007). The extreme poverty rate is measured at the household level and reported in percent. 22 FIGURE 5: SMALL AREA ESTIMATES OF NATIONAL POVERTY SEVERITY AT THE MUNICIPALITY LEVEL (2019) Source: World Bank estimates based on EHPM (2019) and Population Census (2007). Poverty severity is measured at the household level and reported in percent. FIGURE 6: SMALL AREA ESTIMATES OF THE NATIONAL POVERTY GAP AT THE MUNICIPALITY LEVEL (2019) Source: World Bank estimates based on EHPM (2019) and Population Census (2007). The poverty gap is measured at the household level and reported in percent. All poverty indicators point towards a concentration of poverty in individual municipalities. The municipalities with the highest poverty concentration are in the Northeast and West of El Salvador. The municipalities with the highest poverty rate are: Potonico (66.6 %), Estanzuelas (57.2 %), Santo Domingo de Guzmán (50 23 %), and Meanguera del Golfo (50 %). The municipalities with the lowest poverty rate are Comalapa (0 %), San Rafael (3.7 %), Chalatenango (5.1 %) as well as Apaneca (5.3 %). Most of the poorer municipalities are in the following departments: Morazán and Ahuachapán. These are also the departments that concentrate many of the municipalities with the highest extreme poverty rate (see Figure 4). Multidimensional Poverty Maps Multidimensional Poverty Frameworks can help to gain a deeper understanding of the drivers of poverty and the non-monetary aspects of welfare. The Monitoring Global Poverty report by the World Bank in 2017 stresses that multidimensional poverty indexes (MPIs) should accompany the monitoring of global poverty.14 Multidimensional poverty frameworks can help to understand the underlying drivers of poverty better. They can be powerful tools to assess if certain countries, sub-regions, or demographic groups are more affected by some dimensions of poverty than others. They can serve as targeting mechanisms for sectoral interventions in the health, educational, or infrastructure sector. Recently, policymakers have made use of multidimensional poverty indexes to understand multidimensional poverty better. 15 In Colombia, for example, a multidimensional poverty map at the municipality level was used to improve the design and implementation of poverty reduction programs and policies. Similarly, Mexico has used the MPI to inform the creation of two large social protection strategies: the National Crusade Against Hunger as well as the Universal Pension System. In Buthan, the MPI is one of 5 criteria applied for the distribution of national resources to local government. Our multidimensional poverty index follows the methodology developed by the statistical office of El Salvador; it consists of 5 dimensions and 20 indicators. The Multidimensional Poverty Index in El Salvador consists of several indicators and dimensions. It was developed in 2015 together with various national and international advisors and is the result of work conducted since 2009. 16 The different dimensions and indicators of the index are presented in Table 5. It is important to note that the individual indicators are measured at the household level and not the individual level. This means that, in many cases, all household members in a particular household are deprived in a certain dimension as soon as there is one member affected by a particular deprivation (e.g., the entire household is affected by early care deprivation as soon as one child does not World Bank (2017). Monitoring Global Poverty: Report of the Commission on Global Poverty. 14 Washington, DC: World Bank. 15OPHI and BMZ (2015). Measuring Multidimensional Poverty:Insights from Around the World. Link: https://www.ophi.org.uk/wp-content/uploads/Informing-Policy-brochure-web-file.pdf 16 STPP y MINEC-DIGESTYC (2015). Medición multidimensional de la pobreza. El Salvador. San Salvador: Secretaría Técnica y de Planificación de la Presidencia y Ministerio de Economía, a través de la Dirección General de Estadística y Censos. 24 attend a nursery). In the case of El Salvador, there are five dimensions to multidimensional poverty and a total of 20 individual indicators. The national multidimensional poverty index measures the following dimensions of poverty: educational poverty, housing conditions, poverty-related to access to labor and social protection, health poverty, and the quality of the habitat. TABLE 5: MULTIDIMENSIONAL POVERTY INDEX INDICATOR DEPRIVED IF LIVING IN A HOUSEHOLD WHERE… Dimension 1: Education Inadequate early At least one child (1-3) does not attend a nursery care Non-attendance At least one child (4-17) does not attend school At least one child (10-17) is lagging behind in his/her Educational legging educational performance Low adult education At least one adult (+18) with low education Dimension 2: Housing conditions Inadequate material The roof of one's housing is made of inadequate – roof material Inadequate material The floor and wall of one's housing is made of – floor and walls inadequate material Overcrowding The ratio of rooms to members is smaller than one 0.34 Insecure tenure The land tenure is insecure Dimension 3: Labor and Social Protection Child labor The household is subject to child labor17 At least one household member (+16) is unemployed, Unemployment or is employed but without work for at least 1 month per year At least one household member (+16) works more than 40 hours a week and earns less than the minimum Under-employment wage, or is not a permanent salaried worker, or and job insecurity involuntarily works less than 40 hours a week or involuntarily conducts seasonal work/did not find work during a period of longer than 1 months a year Lack of access to social security and At least one household member (+16) has no health unemployment insurance or is no contributing member (only affiliated) benefits Dimension 4: Health 17The definition of child labor follows its legal form, but also considers caretaking. A household is affected by child labor if at least one child engages in labor due to the official age restriction, or engages in a dangerous form of child labor. Additionally, as soon as a child (5-13) dedicates more than 28 hours per week to unpaid care work, the household is subject to child labor. 25 Sum of food insecurity cases (by dimension and Food insecurity household member)18 A household member did not consult with a health Lack of access to professional or the public sector health infrastructure health services due to access constraints19 Lack of access to A household has no access to portable water water Lack of access to A household has no access to sanitation or only to a sanitation deprived from of sanitation20 Dimension 5: Quality of the habitat Lack of public Lack of soccer field, park, playgrounds, and spaces communal houses, or without activities/too far away21 At least one household member has fallen victim to Crime some form of crime22 At least one household member experiences Insecurity restrictions in their activities due to perceived insecurities in the neighborhood23 The dwelling has been affected by streams of water, Environmental risks causing damages; landslides or is exposed to and damages prohibited drainage systems24 18 The national household survey includes eight questions on food insecurity of adults and six questions on food insecurity of minors. The degree of food insecurity depends on whether the household includes a minor or not. A household with at least one minor is food-insecure as soon as at least one of the 14 different dimensions of food insecurity affects an adult or a minor. The degree of food insecurity depends on the number of food insecurity dimensions: In the case of households with minors, the food insecurity index is 2 if the household is affected by 1-5 dimensions, 3 if the household is affected by 6-10 and 4 if it is affected by 11-15 dimensions. In the case of households without minors, the food insecurity index is 2 if the household is affected by 1-3 dimensions, 3 if it is affected by 4-6 and 4 if it is affected by 7 to 8 dimensions. The final indicator is one for a food insecurity index higher than 2, and zero otherwise. 19A household is deprived if a household member consulted with healer, friend/family member or nobody and did not consult with the public health system due to access constraints (too expensive or too far away, lack of medicines or personnel, too sick or due to work reasons); a household member consulted with NGOs, pharmacies, a healer’s house or at home) and did not consult with the public health system due to access constraints; a household member consulted with the private health sector due to access constraints in the public health sector; a household would theoretically consult with one of the above in the case a household member gets sick in the future. 20 A deprived form of sanitation is everything but a toilet. 21 The household is only deprived if it is affected by all of the dimensions. 22As soon as one household member reports some form of victimization, the household is deprived in this dimension. 23As soon as one household member reports some form of restriction due to perceived insecurities, the household is deprived in this dimension. 24As soon as one household member reports some form of the above the household is deprived in this dimensions. 26 The multidimensional poverty indicator depends on the sum of all individual indicators and a predefined threshold. To define multidimensional poverty, one first creates individual indicators. Then, one aggregates all 20 indicators. A household is affected by multidimensional poverty if the sum of the indicators is larger than 7, which is the predefined multidimensional poverty threshold. In the case of El Salvador, 28.1 percent of households were multidimensionally poor in 2019. The table below presents the different indicators. TABLE 6: MPI EL SALVADOR – SHARE OF DEPRIVED HOUSEHOLDS BY INDICATOR Share of households Indicator deprived Non-attendance rate 10.2 Educational delays 1.7 Inadequate care of young children 14.2 Low adult education level 77.5 Inadequate materials (roof) 5.2 Inadequate materials (floor and walls) 18.3 Overcrowding 40.5 Insecurity of tenure 9.9 Vulnerable employment 61.3 Unemployment 14.2 Lack of Social Security 69.1 Child labor 4.8 Lack of health services 9.8 Lack of drinking water 19.6 Lack of sanitation 41.5 Food insecurity 16.0 Lack of public spaces 38.6 Crime 7.6 Insecurity 42.8 Environmental risk factors 5.2 MPI 28.1 Source: DIGESTYC estimates based on EHPM (2019). The figure below shows the small area estimates of multidimensional poverty. Multidimensional poverty varies between 0.0 and 77.6 percent. There is significant variation at the municipality level with respect to the share of households affected by multidimensional poverty. This is in line with what is observed for monetary poverty estimates. Like monetary poverty, there is a large concentration of multidimensional poverty in the Northeast and Southwest of the country. Contrary to the income-based measure of poverty, certain municipalities on the country's southern border are significantly affected by a high share of households being multidimensionally poor. The map below can 27 serve as a complementary measure to the poor income-based poverty maps and can draw additional insights into the underlying drivers. It can also generate a broader perspective on poverty without solely relying on monetary measures. FIGURE 7: SMALL-AREA ESTIMATES OF MULTIDIMENSIONAL POVERTY (2019) Notes: The map plots the multidimensional poverty indicator at the subnational level in El Salvador. We follow the national definition of multidimensional poverty. Source: World Bank estimates based on EHPM (2019) and Census (2007). The poverty rate is measured at the household level and reported in percent. There is considerable spatial variation in the five dimensions of the MPI. Figure 5 to 9 plot the five dimensions of the MPI. These graphs reveal that there is significant heterogeneity across municipalities in all dimensions. The spatial dimension of the overall MPI seems to be most aligned with the educational, housing, and health dimension, while the labor and living condition dimension reveal a slightly different spatial pattern. Investments in education, health and housing of the poorest might be most effective in decreasing multidimensional poverty. 28 FIGURE 5: MPI - EDUCATIONAL DIMENSION FIGURE 6: MPI - HOUSING DIMENSION FIGURE 7: MPI - LABOR DIMENSION FIGURE 8: MPI - HEALTH DIMENSION FIGURE 9: MPI - LIVING CONDITIONS DIMENSION Notes: The maps plot the 5 main dimensions of the MPI, following the national methodology. Source: World Bank estimates based on EHPM (2019) and Census (2007). The poverty rate is measured at the household level and reported in percent. How could these updated maps be used to improve the targeting of social programs and inform the poverty eradication strategy? The "Poverty Eradication Strategy" was established by the previous government in 2017, with the signature of the Executive Decree No. 28. The Strategy is defined as a set of programs and policies designed for the eradication of extreme poverty in the period 2017-2030. This is done through the social protection system and policies supporting skill development and income improvements among families living in extreme poverty in the 262 municipalities. The government of El Salvador currently identifies municipalities with high extreme poverty using information from the 2007 Population Census as well as the unique participant registration system (RUP). The RUP comes along with a 29 household questionnaire, allowing insights into each household's socioeconomic condition. The identification of poor municipalities and households is based on four different dimensions of this questionnaire: household wealth, access to public services, education, and each household's social capital. 25 Eligible municipalities are those part of strata 1 to 7, according to the RUP. This ranking of municipalities is currently the basis for geographic targeting mechanisms forming part of El Salvador's poverty eradication strategy 2020, such as the "Familias Sostenibles" program. In a second stage, the prioritization of households in each municipality is based on the score of the Quality-of-Life Index based on the Single Registry of Participants (IRUP). When comparing the ranking of municipalities based on the updated poverty maps to the ranking of municipalities currently used in the current poverty reduction strategy, there are significant differences. To evaluate this, we look at a number of example municipalities and compare their ranking from the RUP to the estimates of moderate monetary poverty from the Fay-Herriot model. The poorest municipality in the previous maps is San Isidro. In contrast, this municipality only ranks 49th when looking at the Fay-Herriot estimates. Similarly, the municipality with the lowest poverty rate on the previous maps (San Salvador) ranks at 219th in the Fay-Herriot maps. For a full comparison of the ranking of all municipalities, see Annex 6. This reranking suggests that there is scope for efficiency savings to reduce poverty by using more updated poverty information at a subnational level (Fay-Herriot Poverty maps) to identify poor municipalities. There are several possible reasons for the reranking of municipalities between the old and new poverty maps. First, there could be methodological reasons for the change in rankings. As detailed in the methodological section of this report, Fay-Herriot estimation techniques have important empirical advantages over alternative small-area poverty estimations. The Fay-Herriot model accounts for lack of precision and representativeness at disaggregated geographical levels. In addition, it allows to combine survey data with alternative data sources, drawing from additional information. Second, the reranking could be due to different data sources at use for the estimation of old and news maps. While our updated maps draw from a combination of updated household surveys, population estimates and the Census from 2007, the previous maps rely on data from the RUP Lastly, poverty stories could indeed have changed for municipalities. These changes could be due to the large emigration from El Salvador, urban-rural migration patterns, natural disasters, the development of criminal activity at the subnational level, or spatial patterns of corruption and fraud. We leave a detailed analysis of these potential drivers to future research. 25El Salvador 2020. Manual Operativo. Estrategia de Erradicación de la Pobreza. Familias Sostenibles.“ 30 Caveats and limitations Although the Fay-Herriot model has many advantages over alternative small- area poverty estimation techniques, it is subject to important caveats and limitations.26 The most important limitation is that estimates rely on a model, which relies on model assumptions. These assumptions, on the other hand, might be difficult to check or not aligned with the true underlying data distribution. 27 Another important caveat is that the model assumes that sampling variances are known, which is not true in empirical applications. Therefore, the model relies on estimates of these variances. This estimation could introduce potential errors for estimated MSEs. Lastly, the model relies on information gathered from sampled areas. Consequently, imprecision might still be an issue in area-level models. Given that the resulting poverty estimates rely on empirical estimation techniques, the municipal moderate poverty rate is not always larger than the municipal extreme poverty rate. When comparing the resulting EBLUPs of moderate and extreme poverty rates to each other, there are 7 municipalities, for which the model estimates a larger extreme poverty rate than a moderate poverty rate. This is the case for the municipalities of San Fernando, Delicias de Concepción, Santa Rita, Santiago de la Frontera, Arambala, Masahuat, and Comalapa. 3. Conclusion In this paper, we derive small-area poverty estimates at the municipality level by applying the Fay-Herriot model for small area estimations, using household data from 2019 and the Population Census from 2007. We estimate several model specifications and choose the one with the highest precision and lowest number of outbound predictions: the non-transformed model specification with an ampl correction of the variance. We find that poverty rates vary significantly at the municipality level in El Salvador. All poverty indicators shown in this report point towards the concentration of poverty in certain areas. While there are certain variations in the ranking of poverty across the different indicators at use, the poorest municipalities are concentrated in the Northeast and West of El Salvador. 26Corral, Paul; Molina, Isabel; Cojocaru, Alexandru; Segovia, Sandra. 2022. Guidelines to Small Area Estimation for Poverty Mapping. Washington, DC : World Bank. © World Bank. https://openknowledge.worldbank.org/handle/10986/37728 License: CC BY 3.0 IGO. These assumptions are the linearlity assumption as well as the normality assumption of the Fay- 27 Herriot model. 31 The poverty maps are an important contribution to El Salvador's agenda to eliminate poverty. The maps presented in this report can serve as an input for geographic targeting programs. The maps can also be combined with complementing targeting strategies in the design and application of public policies. The maps shed light on important poverty drivers in the country's development agenda. Small-area poverty estimates of multidimensional poverty and its dimensions are especially useful to detect investment needs for education, health, housing, or labor markets. Our methodology considers the data environment in El Salvador and therefore makes an important contribution to the country's agenda to eliminate poverty. 32 References Alkire, S., Kanagaratnam, U. and Suppa, N. (2020). 'The global Multidimensional Poverty Index (MPI): 2020 revision', OPHI MPI Methodological Note 49, Oxford Poverty and Human Development Initiative, University of Oxford. Alkire, S., Foster, J. E., Seth, S., Santos, M. E., Roche, J. M., & Fernandez, P. B. (2015). Multidimensional poverty measurement and analysis: chapter 5-the Alkire-Foster counting methodology. Besley, T., 1995. "Property Rights and Investment Incentives: Theory and Evidence from Ghana." Journal of Political Economy. 103(5): 903–937. Bondarenko M., Kerr D., Sorichetta A., and Tatem, A.J. 2020. Census/projection- disaggregated gridded population datasets for 189 countries in 2020 using Built- Settlement Growth Model (BSGM) outputs. WorldPop, University of Southampton, UK. doi:10.5258/SOTON/WP00684. Corral, Paul; Molina, Isabel; Cojocaru, Alexandru; Segovia, Sandra. 2022. Guidelines to Small Area Estimation for Poverty Mapping. Washington, DC : World Bank. © World Bank. https://openknowledge.worldbank.org/handle/10986/37728 License: CC BY 3.0 IGO. Elbers, C., Lanjouw, J. O., & Lanjouw, P. (2003). Micro-level estimation of poverty and inequality. Econometrica, 71(1), 355-364. Elbers, Chris; Fujii, Tomoki; Lanjouw, Peter; Ozler, Berk; and YIN, Wesley. Poverty Alleviation through Geographic Targeting. (2007). Journal of Development Economics. 83, (1), 198-213. Research Collection School of Economics. Available at: https://ink.library.smu.edu.sg/soe_research/269 El Salvador 2020. „Manual Operativo. Estrategia de Erradicación de la Pobreza. Familias Sostenibles.“ Eurostat, 2019. Guidelines on small area estimation for city statistics and other functional geographics. FRA, EU MIDIS-II, 2018. "Second European Union Minorities and Discrimination Survey. Roma - Selected Findings." Publications Office of the European Union. Fay, R.E. and Herriot, R.A. (1979). Estimates of income for small places: an application of James-Stein procedures to census data. Journal of the American Statistical Association,74, 269-277. Field, E. and M. Torero, 2003. "Do property Titles Increase Access to Credit? Evidence from Peru." Mimeo, Harvard University. 33 Field, E., 2005. "Property Rights and Investment in Urban Slums." Journal of the European Economic Association. 3: 279–290. Field, E., 2007. "Entitled to Work: Urban Property Rights and Labor Supply in Peru." The Quarterly Journal of Economics. 122(4): 1561–1602. Galiani, S. and E. Schargrodsky, 2011. "Land Property Rights and Resource Allocation." Journal of Law and Economics. 54(S4) S329 – S345. GIZ, 2019. "Poverty Targeting Primer. Concepts, Methods and Tools.". Link: https://www.giz.de/de/downloads/PovertyTargeting_Primer_FullVersion_2019.p df Grosh, M., Del Ninno, C., Tesliuc, E., & Ouerghi, A. (2008). For protection and promotion: The design and implementation of effective safety nets. World Bank Publications. Guadarrama, María, Isabel Molina, and J. N. K. Rao. "A comparison of small area estimation methods for poverty mapping." Statistics in Transition new series 1.17 (2016): 41-66. Halbmeier et al. (2019). The fayherriot command for estimating small-area indicators. The Stata Journal (2019). 19, Number 3, pp. 626–64. DOI: 10.1177/1536867X1987423. Human Development Initiative. (2018). Global Multidimensional Poverty Index 2018: The most detailed picture to date of the world's poorest people. University of Oxford, UK. Li, H., and P. Lahiri. 2010. An adjusted maximum likelihood method for solving small area estimation problems. Journal of Multivariate Analysis101: 882–892. Milan, T. and N. Burnet, 2013. "The Economic Cost of Out-of-School Children in 20 Countries." Results for Development. Molina and Morales (2009). Small area estimation of poverty indicators. Boletín de Estadística e Investigación Operativa. Vol. 25, No. 3, Octubre 2009, pp. 218- 22. Molina and Marhuenda. "Sae: An R Package for Small Area Estimation." Link: https://journal.r-project.org/archive/2015/RJ-2015-007/RJ-2015-007.pdf Neves, A., D. Silva, and S. Correa. 2013. Small domain estimation for the Brazilian service sector survey. Estadística 65: 13–37. Rao, J. N. K., and I. Molina. 2015. Small Area Estimation. 2nd ed. Hoboken, NJ: Wiley. 34 OPHI and BMZ (2015). Measuring Multidimensional Poverty: Insights from Around the World. Link: https://www.ophi.org.uk/wp-content/uploads/Informing-Policy- brochure-web-file.pdf Rao, J.N.K.. (2015). Empirical Best Linear Unbiased Prediction (EBLUP): Basic Area Level Model. 10.1002/9781118735855.ch6.https://www.researchgate.net/publication/3158340 99_Empirical_Best_Linear_Unbiased_Prediction_EBLUP_Basic_Area_Level_Model Robles, Claudia (2009). Pueblos indigenas y programas de transferencia con corresponsabilidad. Avances y desfios desde un enfoque étnico. Social Policy Series No 156, United Nations. Slud, E. V., and T. Maiti. 2006. Mean-squared error estimation in transformed Fay– Herriot models. Journal of the Royal Statistical Society, Series B 68: 239–257. Szymkowiak, Marcin, Andrzej Młodak, and Łukasz Wawrowski. "Mapping poverty at the level of subregions in Poland using indirect estimation." STATISTICS 609 (2017). World Bank. Beyond Monetary Poverty. Link: https://openknowledge.worldbank.org/bitstream/handle/10986/30418/9781464 813306_Ch04.pdf World Bank (2017). Monitoring Global Poverty: Report of the Commission on Global Poverty. Washington, DC: World Bank. You, Y., and B. Chapman. 2006. Small area estimation using area-level models and estimated sampling variances. Survey Methodology32: 97–103. 35 Annex Annex 1 – Choosing the best model among several model specifications To estimate the Fay-Herriot model, we apply the Stata ado command Fayherriot developed by Halbmeier et al. (2019). 28 The Fayherriot command is the most up- to-date command for an empirical analysis of the Fay-Herriot model and the most precise one. It allows: • To produce out-of-sample predictions. • Adjust non-positive random-effects variance estimates. • Deal with the violation of model assumption. The Fay-Herriot model is estimated at the municipality level and requires datasets at the domain level of interest with one observation per domain. We, therefore, first, aggregate the household data at our domain level of interest, the municipality level. Due to its mixed nature, the Fay-Herriot model requires a pre- specified estimation of the sampling error variance. We base our estimates of the sampling error variance on the direct estimates of the poverty indicators of interest. We also account for the survey design of the national household survey. We estimate several specifications of the Fay-Herriot model and then choose the one with the highest precision. A rule of thumb often applied by statistical offices is that the CV should not be larger than 20 percent. An additional selection criteria is the number of outbound estimates (e.g. the number of negative estimates). We implement the stata command fayherriot. Using this command allows for an out-of-sample estimation based on the in-sample observations. The first model specification is a simple linear mixed model, depending on direct estimates of the national poverty indicator at the household level and regressed on the municipality explanatory variables and the sampling error variance. The gamma option specifies the display of summary statistics of the shrinkage factor29, and nolog suppresses the iteration log of the optimization algorithm. The variance of the random effects is estimated through the reml estimation method. Table A1 shows the results of this specific Fay-Herriot model. While the CV and MSE of our estimator is low, all estimates are outbound (above 1). 28 Halbmeier et al. (2019). The fayherriot command for estimating small-area indicators. The Stata Journal (2019). 19, Number 3, pp. 626–64. DOI: 10.1177/1536867X1987423. 29 The shrinkage factor shows how direct estimates and model predictions are weighted when calculating the EBLUP. Large values off γd mean that a large weight is given to the direct estimate θd. 36 TABLE A1: FH MODEL - POVERTY RATE (2019) Variable Obs Mean Std. Dev. Min Max Direct estimator 226 20.87 9.87 0 66.62 CV (Direct estimator) 226 315.38 283.12 0 1479.23 EBLUP estimator 262 124.13 11.74 100 194.69 CV (FH Model) 262 342.75 265.22 0 962.88 MSE EBLUP 262 .37 .52 0 2.07 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using a simple linear mixed model. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 municipalities in El Salvador. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). To account for the occurrence of zero variances, we apply the ampl estimation. In this case, the random effect variance is estimated via the adjusted maximum- likelihood method (Li and Lahirini, 2010). 30 Table A2 shows the results. The mean squared error is lower for this model, and the CV is below 20 percent. There are also no outbound estimates. This model specification is therefore a good option for small-area estimations in El Salvador. TABLE A2: FH MODEL WITH AMPL ESTIMATION - POVERTY RATE (2019/20) Variable Obs Mean Std. Dev. Min Max Direct estimator 226 20.87 9.87 0 66.62 CV (Direct estimator) 225 1768.8 1722.97 0 9018.58 EBLUP estimator 262 21.09 8.87 0 66.62 CV (FH Model) 261 1878.94 1597.07 0 11243.48 MSE EBLUP 262 .22 .29 0 1.12 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using a ampl estimation technique for the variance. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 30Li, H., and P. Lahiri. 2010. An adjusted maximum likelihood method for solving small area estimation problems. Journal of Multivariate Analysis101: 882–892. 37 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 municipalities in El Salvador. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). An alternative estimation method to account for zero variances is the aryl estimation. Table A3 shows the result of this estimation. The CV is larger than the one in Table A2. TABLE A3: FH MODEL WITH ARYL ESTIMATION – POVERTY RATE (2019/20) Variable Obs Mean Std. Dev. Min Max Direct estimator 226 20.87 9.87 0 66.62 CV (Direct estimator) 225 1768.8 1722.97 0 9018.58 EBLUP estimator 262 21.09 8.9 0 66.62 CV (FH Model) 261 1920.76 1679.07 0 11865.86 MSE EBLUP 262 .24 .33 0 1.27 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using an aryl transformation. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 municipalities in El Salvador. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). Next, we estimate an arcsin transformation of the model. An arcsin transformation of the Fay-Herriot model is beneficial when the outcome of interest lies within a range of 0 and 1, as is the case for poverty rates. We first take the direct estimator's arcsin square root from the household survey to apply the arcsin transformation. We then estimate the arcsin variance from the effective sample size based on the actual sample size and the survey design's design effect (Results in Table A4). The arcsin method sets an upper- and lower bound for the estimated poverty rate (0 and 1) but does not report the Mean-Squared-Error or coefficients of variation. It is, therefore, difficult to compare the performance of this model to the other model specifications. TABLE A4: FH MODEL WITH ARCSIN TRANSFORMATION Variable Obs Mean Std. Dev. Min Max 38 Direct estimator 226 20.87 9.87 0 66.62 EBLUP estimator 262 20.4 6.26 5.23 38.41 Notes: The table presents results from a model specification of the Fay-Herriot model using an arcsin transformation. Under this specification, it is not possible to compute the coefficients of variation (CVs) nor mean-squared errors (MSEs). Source: EHPM (2019) and Census (2007) We then estimate a log transformation of the classical Fay-Herriot model for in- sample municipalities. The log transformation can help with analyses in which not all domains are sampled. We, therefore, log-transform equivalent incomes and the variances of the sampling error. For the back-transformation of the EBLUP and MSE to its original scale, we once applied the bias correction developed by Slud and Maiti (2006)31 and the crude bias correction by Neves et al. (2013) and Rao and Molina (2015)32. The results are shown in table A5 and table A6. Only under the crude bias correction method, out-of-sample predictions are possible. TABLE A5: FH LOG-TRANSFORMED MODEL - POVERTY RATE (2019) Variable Obs Mean Std. Dev. Min Max Direct estimator 226 20.87 9.87 0 66.62 CV (Direct estimator) 225 1768.8 1722.97 0 9018.58 EBLUP estimator 225 21.43 9.17 5.26 66.62 CV (FH Model) 225 1766.4 1530.39 0 6199.38 MSE EBLUP 225 .19 .27 0 1.53 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using a log transformation. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 municipalities in El Salvador. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). 31 Slud, E. V., and T. Maiti. 2006. Mean-squared error estimation in transformed Fay–Herriot models. Journal of the Royal Statistical Society, Series B 68: 239–257. 32 Neves, A., D. Silva, and S. Correa. 2013. Small domain estimation for the Brazilian service sector survey. Estadística 65: 13–37. Rao, J. N. K., and I. Molina. 2015. Small Area Estimation. 2nd ed. Hoboken, NJ: Wiley. 39 TABLE A8: FH LOG-TRANSFORMED MODEL WITH CRUDE BACK-TRANSFORMATION - POVERTY RATE (2019) Variable Obs Mean Std. Dev. Min Max Direct estimator 226 20.87 9.87 0 66.62 CV (Direct estimator) 225 1768.8 1722.97 0 9018.58 EBLUP estimator 262 21.63 8.68 5.26 66.62 CV (FH Model) 262 4272.51 7439.44 0 39927.3 MSE EBLUP 262 .24 .35 0 2.06 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using a log transformation with crude-bias-correction and out-of-sample predictions. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 municipalities in El Salvador. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). We choose the model version with the highest precision to estimate the poverty rate at the municipality level. Comparing the mean-squared errors and coefficients of variation among all model specifications shows that the model specification using an ampl estimation method for the random error variance is the best performing model. As an additional indication for the performance of our model we plot the EBLUPs against the direct estimators. Figure A1 shows the results. Under the ideal setting, we would see a symmetric allocation of the points around a imaginary diagonal straight line. The ampl specification fairs quite well when comparing the different model specifications, based on these images. FIGURE A1: THE PRECISION OF DIFFERENT MODEL SPECIFICATIONS Direct estimator Direct estimator with ampl estimation Direct estimator with aryl estimation Arcsin transformation 40 Log transformation Log transformation (out-of-sample) Notes: The figure shows scatter plots of the direct estimator observed from the household survey (on the y-axis) and the estimated small-area estimators from different Fay-Herriot model specifications (on the x-axis). Source: EHPM (2019) and Census (2007). The advantage of the national household survey in El Salvador is that it includes 50 municipalities, for which the data at hand is self-representative. We therefore restrict our evaluation of the goodness of fit to these 50 municipalities. Figure A2 plots the FH-estimators against the direct estimators of these municipalities. We also report our results in the tables to follow. These analyses are only robustness checks and not the main decision criteria, as we need to consider the full information feeding into our model to make a final decision. It therefore only serves to validate if our chosen model specification is completely off, which is not the case. The mean-squared error and coefficient of variation of our chosen model is low and there are no outliers in the EBLUPs. FIGURE A2: THE PRECISION OF DIFFERENT MODEL SPECIFICATIONS (50 SELF-REPRESENTATIVE MUNICIPALITIES) Direct estimator Direct estimator with ampl estimation 41 Direct estimator with aryl estimation Arcsin transformation Arcsin transformation with ampl Log transformation (out-of-sample) Source: EHPM (2019) and Census (2007) TABLE A9: FH MODEL - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 50 4.06 2.52 .35 10.41 CV (Direct estimator) 50 122.66 46.28 35.25 271.7 EBLUP estimator 50 103.97 2.31 100.42 110.11 CV (FH Model) 50 108.34 33.08 34.88 196.76 42 MSE EBLUP 50 .01 .01 0 .05 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using a simple linear mixed model. In this case, we restrict our sample to the 50 auto-representative municipalities. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). TABLE A10: FH MODEL WITH AMPL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 50 4.06 2.52 .35 10.41 CV (Direct 50 3768.1 1681.57 1764.71 10036.85 estimator) EBLUP estimator 50 3.68 2.05 .52 9.44 CV (FH Model) 50 3833.67 2497.08 1697.9 12217.02 MSE EBLUP 50 .01 .01 0 .03 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using an ampl estimation technique for the variance. In this case, we restrict our sample to the 50 auto- representative municipalities. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). TABLE A11: FH MODEL WITH ARYL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 50 4.06 2.52 .35 10.41 CV (Direct 50 3768.1 1681.57 1764.71 10036.85 estimator) EBLUP estimator 50 3.86 2.2 .41 9.62 CV (FH Model) 50 3509.64 1714.68 1705.14 8673.9 MSE EBLUP 50 .01 .01 0 .04 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using an aryl transformation. In this case, we restrict our sample to the 50 auto-representative municipalities. The first row reports the direct estimator (in %), which relies on direct observations 43 from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). TABLE A12: FH MODEL WITH ARCSIN TRANSFORMATION Variable Obs Mean Std. Dev. Min Max Direct estimator 50 4.06 2.52 .35 10.41 EBLUP estimator 50 4.04 2.46 .49 10.3 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using an arcsin transformation. In this case, we restrict our sample to the 50 auto-representative municipalities. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). TABLE A13: FH MODEL WITH ARCSIN TRANSFORMATION AND AMPL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 50 4.06 2.52 .35 10.41 EBLUP estimator 50 4.02 2.39 .68 10.19 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using an arcsin transformation and ampl estimation technique for the variance. In this case, we restrict our sample to the 50 auto-representative municipalities. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). TABLE A14: FH LOG-TRANSFORMED MODEL - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 50 4.06 2.52 .35 10.41 CV (Direct 50 3768.1 1681.57 1764.71 10036.85 estimator) EBLUP estimator 50 4.21 2.23 .78 10.64 44 CV (FH Model) 50 4026.75 603.73 3360.12 6481.88 MSE EBLUP 50 .03 .03 0 .15 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using a log transformation. In this case, we restrict our sample to the 50 auto-representative municipalities. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). TABLE A15: FH LOG-TRANSFORMED MODEL WITH CRUDE BACK-TRANSFORMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 50 4.06 2.52 .35 10.41 CV (Direct 50 3768.1 1681.57 1764.71 10036.85 estimator) EBLUP estimator 50 4.3 2.31 1.06 10.51 CV (FH Model) 50 2914.74 767.78 1764.17 5154.36 MSE EBLUP 50 .01 .01 0 .06 Notes: The table reports results from a Fay-Herriot estimation for moderate poverty rates using a log transformation with a crude-back-transformation and out-of-sample predictions. In this case, we restrict our sample to the 50 auto-representative municipalities. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. Source: EHPM (2019) and Census (2007). Annex 2 – FH Estimates of Extreme Poverty Indicators The tables below shows the results of the different model specification for the small-area estimation of extreme poverty. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 45 municipalities in El Salvador (with exception of the log-transformed model). Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. TABLE A16: FH MODEL - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 6.54 6.6 0 38.46 CV (Direct estimator) 226 169.76 196.67 0 1132.33 EBLUP estimator 262 106.92 6.7 98.56 146.9 CV (FH Model) 262 196.81 187.43 0 676.56 MSE EBLUP 262 .1 .15 0 .61 Source: EHPM (2019) and Census (2007) TABLE A17: FH MODEL WITH AMPL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 6.54 6.6 0 38.46 CV (Direct estimator) 187 3720.77 2743.43 0 10385.96 EBLUP estimator 262 6.45 5.92 -1.67 38.46 CV (FH Model) 223 3749.25 10325.05 -119066.39 62830.85 MSE EBLUP 262 .08 .11 0 .4 Source: EHPM (2019) and Census (2007) TABLE A18: FH MODEL WITH ARYL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 6.54 6.6 0 38.46 CV (Direct estimator) 187 3720.77 2743.43 0 10385.96 EBLUP estimator 262 6.47 5.94 -1.62 38.46 CV (FH Model) 262 3182.78 10654.79 -136246.7 65948.95 MSE EBLUP 262 .08 .12 0 .46 46 Source: EHPM (2019) and Census (2007) TABLE A19: FH MODEL WITH ACRSIN TRANSFORMATION Variable Obs Mean Std. Dev. Min Max Direct estimator 226 6.54 6.6 0 38.46 EBLUP estimator 262 7.76 4.79 .44 27.14 Source: EHPM (2019) and Census (2007) TABLE A20: FH MODEL WITH ACRSIN TRANSFORMATION AND AMPL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 6.54 6.6 0 38.46 EBLUP estimator 262 7.67 4.64 .46 26.66 Source: EHPM (2019) and Census (2007) TABLE A21: FH LOG-TRANSFORMED MODEL - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 6.54 6.6 0 38.46 CV (Direct estimator) 187 3720.77 2743.43 0 10385.96 EBLUP estimator 187 8.38 5.77 1.15 38.46 CV (FH Model) 187 3182.42 2003.28 0 6659.05 MSE EBLUP 187 .07 .11 0 .9 Source: EHPM (2019) and Census (2007) TABLE A22: FH LOG-TRANSFORMED MODEL WITH CRUDE BACK-TRANSFORMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 6.54 6.6 0 38.46 CV (Direct estimator) 187 3720.77 2743.43 0 10385.96 EBLUP estimator 262 9.25 6.01 1.28 38.46 47 CV (FH Model) 262 17586.62 27935.74 0 174622.73 MSE EBLUP 262 .13 .27 0 2.23 Source: EHPM (2019) and Census (2007) Annex 3 – FH Estimates of Poverty Gaps The tables below shows the results of the different model specification for the small-area estimation of moderate poverty gaps. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 municipalities in El Salvador (with exception of the log-transformed model). Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. TABLE A23: FH MODEL - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 5.07 4.06 0 22.95 CV (Direct estimator) 226 109.82 121.4 0 921.32 EBLUP estimator 262 105.22 3.95 100 125.8 CV (FH Model) 262 129.42 119.88 0 431.83 MSE EBLUP 262 .04 .06 0 .22 Source: EHPM (2019) and Census (2007) TABLE A24: FH MODEL WITH AMPL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 5.07 4.06 0 22.95 CV (Direct estimator) 217 2771.69 2379.64 0 10906.5 EBLUP estimator 262 5 3.62 0 22.95 CV (FH Model) 253 3769.34 7463.48 0 80769.03 48 MSE EBLUP 262 .03 .04 0 .17 Source: EHPM (2019) and Census (2007) TABLE A25: FH MODEL WITH ARYL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 5.07 4.06 0 22.95 CV (Direct estimator) 217 2771.69 2379.64 0 10906.5 EBLUP estimator 262 5.01 3.63 0 22.95 CV (FH Model) 253 3834.65 7598.93 0 81750.99 MSE EBLUP 262 .03 .05 0 .19 Source: EHPM (2019) and Census (2007) TABLE 26: FH MODEL WITH ACRSIN TRANSFORMATION Variable Obs Mean Std. Dev. Min Max Direct estimator 226 5.07 4.06 0 22.95 EBLUP estimator 262 4.91 2.34 .87 14.33 Source: EHPM (2019) and Census (2007) TABLE A27: FH MODEL WITH ACRSIN TRANSFORMATION AND AMPL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 5.07 4.06 0 22.95 EBLUP estimator 262 4.85 2.22 .92 13.55 Source: EHPM (2019) and Census (2007) TABLE A28: FH LOG-TRANSFORMED MODEL - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 5.07 4.06 0 22.95 CV (Direct estimator) 217 2771.69 2379.64 0 10906.5 49 EBLUP estimator 217 5.39 3.74 .01 22.95 CV (FH Model) 217 4237.61 4100.38 0 20765.56 MSE EBLUP 217 .05 .08 0 .68 Source: EHPM (2019) and Census (2007) TABLE A29: FH LOG-TRANSFORMED MODEL WITH CRUDE BACK-TRANSFORMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 5.07 4.06 0 22.95 CV (Direct estimator) 217 2771.69 2379.64 0 10906.5 EBLUP estimator 262 5.82 4.04 .01 24.2 CV (FH Model) 262 25791.81 60792.3 0 350094.63 MSE EBLUP 262 .06 .2 0 2.1 Source: EHPM (2019) and Census (2007) Annex 4 – FH Estimates of Poverty Severity The tables below shows the results of the different model specification for the small-area estimation of moderate poverty severity. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 municipalities in El Salvador (with exception of the log-transformed model). Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. TABLE A30: FH MODEL - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 2.14 2.01 0 11.2 50 CV (Direct estimator) 226 54.21 69.25 0 602.43 EBLUP estimator 262 102.11 1.84 99.89 111.85 CV (FH Model) 262 62.27 60.58 0 208.59 MSE EBLUP 262 .01 .01 0 .05 Source: EHPM (2019) and Census (2007) TABLE A31: FH MODEL WITH AMPL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 2.14 2.01 0 11.2 CV (Direct estimator) 217 3280.47 2597.42 0 10906.5 EBLUP estimator 262 2.06 1.77 -.15 11.2 CV (FH Model) 253 3124.12 15445.86 -134209.7 150595.84 MSE EBLUP 262 .01 .01 0 .04 Source: EHPM (2019) and Census (2007) TABLE A32: FH MODEL WITH ARYL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 2.14 2.01 0 11.2 CV (Direct estimator) 217 3280.47 2597.42 0 10906.5 EBLUP estimator 262 2.07 1.77 -.13 11.2 CV (FH Model) 253 2942.09 17462.22 -164864.92 148287.67 MSE EBLUP 262 .01 .01 0 .04 Source: EHPM (2019) and Census (2007) TABLE A33: FH MODEL WITH ACRSIN TRANSFORMATION Variable Obs Mean Std. Dev. Min Max Direct estimator 226 2.14 2.01 0 11.2 EBLUP estimator 262 1.94 1.01 .4 5.87 51 Source: EHPM (2019) and Census (2007) TABLE A34: FH MODEL WITH ACRSIN TRANSFORMATION AND AMPL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 2.14 2.01 0 11.2 EBLUP estimator 262 1.91 .96 .38 5.47 Source: EHPM (2019) and Census (2007) TABLE A35: FH LOG-TRANSFORMED MODEL - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 2.14 2.01 0 11.2 CV (Direct estimator) 217 3280.47 2597.42 0 10906.5 EBLUP estimator 217 2.28 1.93 0 11.2 CV (FH Model) 214 14746.64 22601.62 0 178691.44 MSE EBLUP 217 .07 .13 -.01 .87 Source: EHPM (2019) and Census (2007) TABLE A36: FH LOG-TRANSFORMED MODEL WITH CRUDE BACK-TRANSFORMATION – DIRECT ESTIMATOR – 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 2.14 2.01 0 11.2 CV (Direct estimator) 217 3280.47 2597.42 0 10906.5 EBLUP estimator 262 2.89 3.74 0 44.59 CV (FH Model) 262 82685.79 229534.48 0 1455918.3 MSE EBLUP 262 .05 .33 0 5.09 Source: EHPM (2019) and Census (2007) Annex 5 – FH Estimates of Multidimensional Poverty 52 The tables below shows the results of the different model specification for the small-area estimation of multidimensional poverty. The first row reports the direct estimator (in %), which relies on direct observations from survey data. The indicator is only reported for 226 municipalities included in the survey. The second row presents the coefficient of variation (CV) multiplied by 100 of the direct estimator. Row 3 reports the empirical best linear unbiased predictor (EBLUP) (in %) at the municipality level generated from a Fay-Herriot model. In this case, we perform out-of-sample predictions and estimate poverty estimates for all 262 municipalities in El Salvador. Row 4 presents the respective CV of the EBLUPs as well as its mean squared error (MSE) multiplied by 100. TABLE A37: FH MODEL WITH AMPL ESTIMATION - DIRECT ESTIMATOR - 2019 Variable Obs Mean Std. Dev. Min Max Direct estimator 226 32.9 16.8 0 81.1 CV (Direct estimator) 223 166.8 81.1 48.3 525.8 EBLUP estimator 262 25.6 16 -22 77.6 CV (FH Model) 259 125.2 546.1 -523.5 6626.9 MSE EBLUP 262 204.2 149.8 0 1411.8 Source: EHPM (2019) and Census (2007) Annex 6 – Comparison of municipality rankings of old and new maps (From highest to lowest poverty incidence) TABLE A38: COMPARISON OF OLD AND NEW MUNICIPALITY RANKING (FROM POOREST TO LEAST POOR) Old Ranking33 New Ranking San Isidro Potonico San Antonio Estanzuelas 33 We do not dispose of the old poverty estimates at the municipality level. Alternative measures of previous estimates can be found here: https://esri- sv.maps.arcgis.com/home/item.html?id=0bc142780f3e44f39d8bbdf7ed9f9116 53 Santo Domingo De Cualococti Guzman Cuisnahiat Meanguera Del Golfo Guaymango Berlin San Simon Jucuapa Nuevo Eden De San Torola Juan Lislique Gualococti Cacaopera Mercedes La Ceiba Cancasque Santo Domingo Monte San Juan San Simon San Fernando Cacaopera Guatajiagua San Antonio Yamabal Yamabal Jucuaran San Emigdio San Fernando Santa Isabel Ishuatan Santa Catarina Jutiapa Masahuat Jicapala El Rosario San Francisco Javier Corinto El Rosario Guaymango Joateca Ojos De Agua Nuevo Eden San Juan San Jorge San Cristobal Jicalapa San Pedro Puxtla Guatajiagua San Antonio de la Cruz Paraiso De Osorio Caluco Candelaria Tacuva San Julian Santa Isabel Ishuatan Jujutla Teotepeque Santa Elena Santa Clara San Pedro Puxtla San Dionisio Lislique San Francisco Chilango Menendez Arambala El Triunfo Jujutla San Lorenzo Sensembra Yoloaiquin 54 El Carmen Chilanga San Francisco Morazan San Pedro Masahuat Alegria San Cristobal Mercedes la Ceiba San Fernando Santa Catarina Masahuat Caluco San Fransico Chinameca Zaragoza San Lorengo San Ramon Santon Domingo de El Rosario Guzaman San Miguel De Chiltiupan Mercedes Nueva Granda Cuisnahuat Victoria Nahulingo Carolina San Gerardo Perquin Mercedes Umaña Huizucar San Isidro Lolotiquillo Poloros San Idefonso Tacuba San Antonio Los Comasagua Ranchos Oratorio de Concepción El Congo San Gerardo Perquin Cinquera El Paraiso Tecoluca San Luis De La Reina Masahuat Tepecoyo Corinto San Luis Del Carmen Jocoatique Nueva Guadalupe San Emigdio Alegria Mercedes Umaña Santa Cruz Michapa El Rosario Nombre De Jesus Lolotique Cinquera Sesori Santa Cruz Analquito San Pedro Nonualco Concepcion Batres El Porvernir San Lorenzo Santa Maria Ostuma Bolivar Nahuizalco Tejutla Dolores Santo Tomas 55 San Jorge San Francisco Morazan Santa Cruz Analquito Sacacoyo Tapaluaca Ilobasco Nueva Trinidad San Buena Ventura Estanzuelas Sociedad San Micuel Tepezontes San Rafael Cedros Rosario de Mora Verapaz San Rafael Oriente Yucuaiquin San Francisco Jiquilisco Chinameca Paraiso de Osorio San Agustin San Antonio Masahuat Ahuachapan San Pedro Masahuat Santa Maria Ostuma Osicala Tenancingo Ozatlan Torola Apastepeque El Divisadero Tecapan Jerusalén San Juan Tepezaontes San Antonio Masahuat Yucuaiquin Dolores Jerusalen Las Flores Las vueltas Salcoatitan Suchitoto Juayua Ojos de Agua San Ildefonso La Laguna San Luis La Herradura Sociedad Joateca San Ramon Ciudad Barrios San Julian Nahuizalco San Matias San Antonio Pajonal San Pedro Perulapan Concepcion De Ataco San Fransicco Menendez Ozatlan Chinameca Colon Delicias de Concepcion Tamanique Moncagua Nueva Trinidad Ciudad Barrios San Antonio Del Monte Concepcion Matres Santiago Texacuangos Comalapa Suchitoto 56 El Carrizal Jiquilisco Santiago Nonualco Coatepeque Candelaria San Juan Tepezontes Meanguera San Fernando San Carlos Cancasque Tenanciango Jocoaitique Cuirilagua Nueva Esparta Yoloaiquin Zacatecoluca San Luis la Herradura California San Agustin Carolina San Jose San Sebastian San Antonio Los Tepetitan Ranchos San Luis la Reina Citala Yayantique Tecoluca Panchimalco San Pablo Tacachico Tejuteque El Carrizal Arcatao San Francisco Javier Santa Elena Intipuca San Luis del Carmen Lolotique Anamoros El Transito Santa Rosa Guachipilin San Jose Guayabal Concepción de Ataco Ereguayquin Berlin San Alejo San Buena Ventura La Libertad Potonico San Isidro Tamanique Nueva Granada San Esteban Catarina Moncagua Izalco San Luis Talpa San Luis Talpa Sesori San Jose Guyabal El Porvenir El Carmen San Vicente Poloros San Juan Nonualco San Cayetano El Paisanl Istepeque Coatepeque Acajutla 57 Santiango de la Frontera Guadalupe El Transito El Paisnal San Isidro Turin Conchagua Chiltiupan La Reina Victoria Agua Caliente Texistepeque Nueva Esparta Oscicala Ereguyquin Azacualpa San Pablo Tacachico Guazapa Apaneca Izalco San Cayetano Istepque Puerto El Triunfo El Dividadero Santa Clara Comacaran Huizucar Talnique Arcatao Puerto el Triunfo San Pedro Perulapan Intipuca Panchimalco Quelepa San Rafael Obrajuelo San Ignacio Talnique El Triunfo Conchagua Candelaria De La Uluazapa Frontera Tepetitan Lolotiquillo San Alejo San Ignacio Santa Cruz Michapa Chalchuapa Ilobasco Santa Rosa Guachipilin Santo Domingo Jutiapa Guadalupe Monte San Juan Guacotecti Concepcion De Oriente San Sebastian Chapeltique Salitrillo Atiquizaya San Juan Opico La Palma La Palma Salcoatitan San Martin Santa Rita Las Vueltas Las Flores Comacaran Verapaz Ciudad Arce 58 Citala Jucuaran Tepecoyo Apopa El Paraiso San Juan Talpa nueva Concepcion Chinameca Sensutepeque Yayantique Nombre de Jesus Santiago Nonualco San Isidrio Labrador San Isidro Labrador El Sauce Santa Maria Ahuachapan Aguilares Meanquera del golfo Meanguera San Jose Villanueva San Rafael Oriente Texistepeque Chirilagua Tejutal Cojutepeque Bolivar San Francisco Gotera San Rafael Cedros Atiquizaya Zacatecoluca Sensuntepeque Nejapa Santa Ana Jucuapa Nueva Concepcion Candearia de la Frontera Comasagua Acajutal Anamoros San Rafael Cedros Metapan San Lorenzo Olocuilta Concepcion de Oriente Ciudad Delgado San Sebastian Sonsonate Dulce Nombre de Maria Rosario De Mora Concepcion Armenia Quezaltepeque El Congo Nejapa La Libertad Chapeltique Cuyultitlan Quelepa Juayua Jayaque Concepcion La Laguna Quezaltepeque Guazapa Apastepeque San Juan Nonualco San Dionisio Turin Tecapan 59 San Miguel de Mercedes La Reina San Juan Opico Usulutan El Rosario San Francisco Lempa Santa Maria La Union Jocoro Cuscatancingo Ciudad Arce Ilopango San Antonio Pajonal Tonacatepeque Olocuitla El Sauce Nahuilingo San Pedro Nonualco Jayaque San Carlos San Bartolome Quezaltepeque Azacualpa San Marcos Pasaquina San Salvador San Rafel Obrajuelo Cuyultitan San Juan Talpa Santa Rosa De Lima El Refugio Agua Caliente San Antonio De La Sacacoyo Cruz Santiango Texacuangos El Carmen Usulutan Tejutepeque Metapan El Rosario Chalchupa San Jose Villanueva Santo Tomas El Refugio La Union Armenia Nueva Gadaupe Soyapango San Vicente El Carmen Sonsonate Santa Tecla Oratorio De Santa Rosa de Lima Concepcion Santiango de Maria San Miguel Nuevo Cuscatlan Santiago De Maria San Francisco Gotera Guacotecti Delicias De Quezalpeque Concepcion San Antonio del Monte San Esteban Catarina Cojuteque San Jose California San Miguel Tepezontes 60 Zaragoza Santa Rita Chalatenango Mejicanos San Miguel Uluazapa Aquilares Nuevo Cuscatlan San Francisco Lempa Pasaquina Santa Ana Jocoro Santiago De La San Sebastian Salitrillo Frontera San Martin San Matias Colon Teotepeque Cuidad Delgado Tapalhuaca Sonzacate Arambala San Bartolome Tonacateque Perulapia Dulce Nombre De Ayutuxtepeque Maria Apopa Sensembra San Marcos Ayutuxtepeque San Tecla Sonzacate Ilopango Antiguo Cuscatlan Cuzcatancingo Masahuat Mejicanos Apaneca Antiguo Cuscatlan Chalatenango Soyapango San Rafael San Salvador Comalapa Notes: The table reports the ranking of municipalities on national poverty rates for old and new poverty maps. The highest rank (1) is the poorest municipality Source: World Bank estimates based on EHPM 2019 and the Population Census, and El Salvador 2020 Manual operativo Estrategia de Erradicacion de Pobreza 61 FIGURE A4: SCATTER PLOT OF NEW AND OLD MUNICIPALITY RANKINGS Notes: The graph plots the old municipality ranking in municipal poverty estimates against the new ranking. The orange line represents the fitted line. Source: World Bank estimates based on EHPM 2019 and the Population Census, and El Salvador 2020 Manual operativo Estrategia de Erradicacion de Pobreza Annex 7 – Small area estimates of poverty (Fay-Herriot) The below table presents small-area poverty estimates at the municipality level. We consider the national poverty line for moderate poverty and household estimates. Column 1 presents the name of the respective municipality, Column 2 poverty estimates from the Fay-Herriot model, Column 3 the related MSE (multiplied by 100) and Column 4 the related CV (multiplied by 100). TABLE A39: SMALL-AREA POVERTY ESTIMATES (FAY-HERRIOT) Moderate poverty Municipality MSE CV estimates (%) Potonico 66.62 0 0 Estanzuelas 57.18 0 0 Santo Domingo De 50.00 Guzman 0 0 Meanguera Del Golfo 50.00 0 0 Berlin 49.98 0 0 Jucuapa 46.46 0.06 538.49 62 Nuevo Eden De San 44.44 Juan 0 0 Gualococti 42.86 0 0 Mercedes La Ceiba 41.13 0 0 Santo Domingo 38.46 0 0 San Simon 38.31 0.11 851.56 Cacaopera 36.93 0.12 943.52 San Antonio 35.71 0 0 Yamabal 35.44 0.36 1697.56 San Emigdio 34.90 0.19 1253.88 Santa Isabel Ishuatan 33.39 0.93 2891.31 Santa Catarina 33.33 Masahuat 0 0 El Rosario 33.33 0 0 Corinto 33.17 0.17 1228.64 Guaymango 32.59 0.05 660.45 Ojos De Agua 31.80 0 183.61 San Jorge 31.43 0.4 2010.22 Jicalapa 31.30 0 0 Guatajiagua 31.18 0.06 756.22 Paraiso De Osorio 31.06 0.87 2995.19 Candelaria 30.93 0.27 1664.83 San Julian 30.66 0.09 986.62 Jujutla 30.00 0.17 1390.04 Santa Elena 29.72 0 230.38 San Pedro Puxtla 29.41 0 0 Lislique 29.29 0.09 997.17 San Francisco 29.05 Menendez 0.12 1194.37 El Triunfo 28.60 0 0 San Lorenzo 28.57 0 0 Yoloaiquin 28.57 0 0 Chilanga 28.38 0.22 1639.04 San Pedro Masahuat 28.21 0.18 1522.45 San Cristobal 28.04 0.08 1037.53 San Fernando 27.58 1.01 3635.03 Caluco 27.30 0.8 3285.41 63 Zaragoza 27.26 0.09 1069.47 San Ramon 27.15 0.82 3331.39 El Rosario 27.13 0.46 2511.01 San Miguel De 27.09 Mercedes 0.92 3536.05 Cuisnahuat 26.84 0.56 2782.51 Nahulingo 26.67 0 0 San Gerardo 26.67 0 0 Mercedes Umaña 26.62 0.07 1010.14 San Isidro 26.41 0.9 3598.76 Poloros 26.24 0.2 1720.13 Tacuba 26.24 0.37 2324.98 San Antonio Los 26.09 Ranchos 0.81 3453.11 El Congo 26.01 0.15 1506.73 Perquin 26.00 0.69 3184.48 El Paraiso 25.97 0.33 2200.28 San Luis De La Reina 25.76 0.59 2979.83 Tepecoyo 25.67 0 218.9 San Luis Del Carmen 25.55 0.88 3681.56 Nueva Guadalupe 25.49 0.04 746.14 Alegria 25.46 0.06 961.09 Santa Cruz Michapa 25.24 0.16 1569.17 Nombre De Jesus 25.05 0.17 1633.39 Cinquera 25.02 0.08 1143.79 Santa Cruz Analquito 25.00 0 0 Concepcion Batres 25.00 0 0 San Lorenzo 24.94 0.66 3251.92 Bolivar 24.88 0.02 596.23 Tejutla 24.86 0 0 Santo Tomas 24.81 0.09 1225.43 San Francisco Morazan 24.81 0.85 3712.89 Sacacoyo 24.81 0.15 1586.4 Ilobasco 24.64 0.05 898.87 San Buena Ventura 24.51 0.81 3661.53 Sociedad 24.35 0.07 1073.54 64 San Rafael Cedros 24.33 0.22 1938.82 Verapaz 24.32 0.02 541.36 Yucuaiquin 24.22 0 268.4 San Francisco 24.19 Chinameca 0.8 3700.44 San Agustin 24.17 0.81 3724.2 Ahuachapan 24.16 0.09 1271.73 Santa Maria Ostuma 24.13 0.39 2600.39 Tenancingo 24.12 0.79 3681.28 Torola 23.97 0.07 1138.74 El Divisadero 23.93 0.04 784.61 Jerusalen 23.90 0.85 3860.43 San Antonio Masahuat 23.65 0.84 3872.35 Dolores 23.47 0.16 1709.73 Las Flores 23.41 0.83 3880.89 Salcoatitan 23.29 0.06 1057.2 Juayua 23.27 0.43 2814.58 San Ildefonso 23.25 0.46 2907.78 San Luis La Herradura 23.25 0.25 2129.57 Joateca 23.08 0 0 Ciudad Barrios 23.07 0.13 1549.91 Nahuizalco 22.92 0.05 947.47 San Antonio Pajonal 22.84 0.47 3009.31 Concepcion De Ataco 22.83 0.41 2791.82 Ozatlan 22.80 0.79 3905.35 Colon 22.77 0.06 1039.59 Tamanique 22.72 0.79 3918.33 Nueva Trinidad 22.66 0.84 4035.22 San Antonio Del Monte 22.54 0.12 1560.96 Santiago Texacuangos 22.53 0.21 2044.23 Suchitoto 22.52 0.1 1400.03 Jiquilisco 22.41 0.05 974.12 Coatepeque 22.28 0.06 1072.29 San Juan Tepezontes 22.25 0.88 4213.72 San Fernando 22.22 0 0 Cancasque 22.18 0.95 4391.26 65 Jocoaitique 22.16 0.59 3468.21 Nueva Esparta 21.91 0.37 2784.65 Zacatecoluca 21.87 0.05 1057.49 California 21.83 1.11 4820.92 Carolina 21.77 0.06 1161.72 San Sebastian 21.75 0.22 2167.56 Tepetitan 21.74 0.09 1412.57 Citala 21.70 0.02 697.56 Tecoluca 21.66 0.22 2157.34 San Pablo Tacachico 21.52 0.14 1719.26 El Carrizal 21.51 0 0 San Francisco Javier 21.43 0 0 Intipuca 21.43 0 0 Lolotique 21.32 0.24 2298.25 El Transito 21.22 0.16 1871.22 San Jose Guayabal 21.20 0.32 2674.21 Ereguayquin 21.14 0.79 4217.07 San Alejo 21.12 0.13 1721.35 La Libertad 21.06 0.04 967.35 San Isidro 21.01 0.1 1526.91 Nueva Granada 20.95 0.8 4282.06 Moncagua 20.85 0.19 2096.52 San Luis Talpa 20.71 0.09 1430.77 Sesori 20.63 0.21 2243.83 El Porvenir 20.62 0.04 919.42 San Vicente 20.59 0.05 1056.03 San Juan Nonualco 20.51 0.47 3347.53 San Cayetano 20.25 Istepeque 0.03 812.24 Acajutla 20.22 0.12 1734.46 Guadalupe 20.12 0.15 1935.05 El Paisnal 20.04 0.2 2227.36 Turin 20.00 0 0 Chiltiupan 20.00 0 0 Victoria 19.97 0.15 1922.88 Texistepeque 19.89 0.02 698.64 66 Oscicala 19.86 0.17 2058.64 Azacualpa 19.84 0.9 4776.82 Guazapa 19.77 0.08 1435.85 Izalco 19.71 0.18 2135.38 Puerto El Triunfo 19.68 0.18 2148.04 Santa Clara 19.65 0.87 4758.2 Huizucar 19.64 0.79 4525.09 Arcatao 19.59 1.04 5212.5 San Pedro Perulapan 19.53 0.24 2525.96 Panchimalco 19.51 0.06 1210.18 San Rafael Obrajuelo 19.35 0.78 4560.99 Talnique 19.15 0.86 4833.91 Conchagua 19.13 0.04 1092.79 Candelaria De La 19.10 Frontera 0.03 872.61 Lolotiquillo 18.83 0.24 2596.62 San Ignacio 18.80 0.22 2502.43 Chalchuapa 18.79 0.07 1402.31 Santa Rosa Guachipilin 18.71 0.17 2175.34 Jutiapa 18.49 0.28 2864.67 Monte San Juan 18.45 0.44 3585.69 Concepcion De Oriente 18.41 0.32 3073.36 San Sebastian Salitrillo 18.28 0.24 2674.73 San Juan Opico 18.27 0.07 1494.3 La Palma 18.23 0.14 2064.86 San Martin 18.21 0.06 1328.26 Las Vueltas 18.18 0 0 Comacaran 18.18 0 0 Ciudad Arce 18.13 0.07 1494.89 Jucuaran 18.07 0.19 2386.75 Apopa 18.05 0.07 1435.17 San Juan Talpa 18.03 0.35 3291.5 Chinameca 17.98 0.14 2079.15 Yayantique 17.89 0.03 909.83 Santiago Nonualco 17.85 0.05 1189.06 San Isidro Labrador 17.75 1.05 5781.15 67 Santa Maria 17.73 0.17 2338.84 Aguilares 17.72 0.02 866.65 Meanguera 17.57 0.34 3341.13 San Rafael Oriente 17.56 0.23 2702.57 Chirilagua 17.48 0.29 3107.79 Cojutepeque 17.44 0.07 1498.15 San Francisco Gotera 17.27 0.05 1337.43 Atiquizaya 17.05 0.06 1462.68 Sensuntepeque 16.96 0.05 1299.91 Santa Ana 16.71 0.07 1540.87 Nueva Concepcion 16.66 0.09 1806.93 Comasagua 16.59 0.32 3429.01 Anamoros 16.33 0.13 2233.02 Metapan 16.21 0.02 966.2 Olocuilta 16.11 0.06 1490.57 Ciudad Delgado 15.90 0.11 2127.08 Sonsonate 15.72 0.07 1678.74 Rosario De Mora 15.66 0.27 3294.92 Concepcion 15.61 Quezaltepeque 0.02 909.98 Nejapa 15.59 0.13 2300.98 Chapeltique 15.48 0.24 3172.32 Quelepa 15.39 0.83 5920.96 Jayaque 15.38 0 0 La Laguna 15.38 0 0 Apastepeque 15.37 0.21 3006.43 San Dionisio 15.36 0 0 Tecapan 15.34 0 0 La Reina 15.29 0 413.2 Usulutan 15.24 0.07 1763.51 San Francisco Lempa 15.22 1.12 6954.45 La Union 14.91 0.05 1531.11 Cuscatancingo 14.88 0.1 2083.72 Ilopango 14.74 0.05 1580.26 Tonacatepeque 14.65 0.07 1783.97 El Sauce 14.58 0.31 3801.39 68 San Pedro Nonualco 14.29 0 0 San Carlos 13.97 0.05 1625.19 Quezaltepeque 13.70 0.04 1514.01 San Marcos 13.70 0.04 1513.15 San Salvador 13.69 0.12 2544.27 Cuyultitan 13.48 0.81 6686.2 Santa Rosa De Lima 13.47 0.06 1887.03 Agua Caliente 13.44 0.25 3713.86 San Antonio De La Cruz 13.43 0 494.04 El Carmen 13.26 0.34 4429.84 Tejutepeque 12.73 0.2 3504.05 El Rosario 12.22 0.33 4695.48 San Jose Villanueva 11.76 0 0 El Refugio 11.55 0.12 2982.17 Armenia 11.44 0.03 1621.58 Soyapango 11.41 0.07 2294.01 El Carmen 11.15 0.07 2373.6 Santa Tecla 10.46 0.06 2367.25 Oratorio De Concepcion 10.35 0.05 2102.69 San Miguel 10.33 0.03 1727.88 Santiago De Maria 10.14 0.25 4942.99 Guacotecti 10.10 0.11 3248.53 Delicias De Concepcion 10.06 0.03 1816.73 San Esteban Catarina 10.00 0 0 San Jose 10.00 0 0 San Miguel Tepezontes 9.11 0 0 Santa Rita 9.09 0 0 Mejicanos 8.37 0.04 2409.41 Uluazapa 8.33 0 0 Nuevo Cuscatlan 8.18 0.85 11243.48 Pasaquina 8.13 0.09 3738.2 Jocoro 7.75 0.07 3476.52 Santiago De La 7.69 Frontera 0 0 San Matias 7.69 0 0 Teotepeque 7.69 0 0 69 Tapalhuaca 7.69 0 0 Arambala 7.69 0 0 San Bartolome 7.17 Perulapia 0 0 Dulce Nombre De Maria 7.16 0 0 Sensembra 7.14 0 0 Ayutuxtepeque 6.97 0.05 3150.42 Sonzacate 6.44 0.04 2999.35 Antiguo Cuscatlan 6.06 0.05 3724.27 Masahuat 5.56 0 0 Apaneca 5.26 0 0 Chalatenango 5.09 0.02 2436.89 San Rafael 3.66 0.05 5989.92 Comalapa 0.00 0 0 Notes: The table presents results from a Fay-Herriot model using an ampl estimation technique ofr variances. Column 1 reports the estimated moderate poverty rate per municipality, Column 2 the related mean-squared errors (MSE) and Column 3 the coefficients of variation (CVs). Source: World Bank estimates based on EHPM (2019). Source: Census (2007) and EHPM (2019).n nndmann@worldbank.orgsbrodmann@worldbank.org) Annex 8 – Poverty Maps using Poverty Headcount Ratios The following presents municipal poverty maps relying on poverty headcount ratios. While the previous maps rely on household estimates, these maps rely on population estimates. 70 FIGURE A5: SMALL AREA ESTIMATES OF THE MODERATE POVERTY HEADCOUNT RATE AT THE MUNICIPALITY LEVEL (2019) Source: EHPM (2019) and Census (2007). Poverty is measured at the population level and using national poverty lines. FIGURE A6: SMALL AREA ESTIMATES OF THE EXTREME POVERTY HEADCOUNT RATE AT THE MUNICIPALITY LEVEL (2019) Source: EHPM (2019) and Census (2007). Extreme poverty is measured at the population level and using national poverty lines. 71 FIGURE A7: SMALL AREA ESTIMATES OF THE POVERTY SEVERITY AT THE MUNICIPALITY LEVEL (2019) Source: EHPM (2019) and Census (2007). Poverty severity is measured at the population level and using national poverty lines. FIGURE A8: SMALL AREA ESTIMATES OF THE POVERTY GAP AT THE MUNICIPALITY LEVEL (2019) Source: EHPM (2019) and Census (2007). Poverty gaps are measured at the population level and using national poverty lines. 72