SOCIAL PROTECTION & JOBS DISCUSSION PAPER No. 2212 | OCTOBER 2022 Guiding Social Protection Targeting Through Satellite Data in São Tomé and Príncipe Peter Fisker, Jordi Gallego-Ayala, David Malmgren-Hansen, Thomas Pave Sohnesen, and Edmundo Murrugarra © 2022 International Bank for Reconstruction and Development / The World Bank 1818 H Street NW Washington DC 20433 Telephone: +1 (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: +1 (202) 522 2625; e-mail: pubrights@worldbank.org. Abstract retro geometric background: © iStock.com/marigold_88 Project 41595 Guiding Social Protection Targeting Through Satellite Data in São Tomé and Príncipe Peter Fisker; Jordi Gallego-Ayala; David Malmgren-Hansen; Thomas Pave Sohnesen; Edmundo Murrugarra 1 Abstract Social safety net programs focus on a subset of the population, usually the poorest and most vulnerable. However, in most developing countries there is no administrative data on relative wealth of the population to support the selection process of the potential beneficiaries of the social safety net programs. Hence, selection into programs is often multi-methodological approached and starts with geographical targeting for the selection of program implementation areas. To facilitate this stage of the targeting process in São Tomé and Príncipe, this working paper develops High Resolution Satellite Imagery (HRSI) poverty maps, providing both estimates of poverty incidence and program eligibility at a highly detailed resolution (110 m x 110 m). Furthermore, the analysis combines poverty incidence and population density to enable the geographical targeting process. This working paper shows that HRSI poverty maps can be used as key operational tools to facilitate the decision-making process of the geographical targeting and efficiently identify entry points for rapidly expanding social safety net programs. Unlike HRSI poverty maps based on census data, poverty maps based on satellite data and machine learning can be updated frequently at low cost supporting more adaptive social protection programs. JEL codes: O18, Q54, R11 Keywords: Social protection; Targeting; Machine learning; Satellite images. 1 Peter Fisker, University of Copenhagen, Department of Economics. pkf@econ.ku.dk; Jordi Gallego-Ayala, World Bank, Social Protection and Jobs Practice. jgallegoayala@worldbank.org; David Malmgren-Hansen, Independent consultant. david@datamines.dk; Thomas Pave Sohnesen, Independent consultant. thomas@sohnesen.dk; Edmundo Murrugarra, World Bank, Social Protection and Jobs Practice. emurrugarra@worldbank.org 1 CONTENTS Abstract .................................................................................................................................................. 1 Acknowledgments .................................................................................................................................. 4 Introduction............................................................................................................................................ 5 I.Context ................................................................................................................................................. 6 A. Country ....................................................................................................................................... 6 B. Poverty and social protection..................................................................................................... 7 II.Data ..................................................................................................................................................... 9 A. Welfare data ............................................................................................................................. 10 B. Data on buildings extracted from satellite images.................................................................... 11 C. Geospatial data..........................................................................................................................13 D. Linking the various data sources ...............................................................................................13 III.Predicting Neighborhood Poverty Levels ......................................................................................... 14 IV.Outcomes ......................................................................................................................................... 17 A. Operational use ........................................................................................................................ 20 V.Discussion and concluding remarks .................................................................................................. 27 VI.References........................................................................................................................................ 30 Annex A. Additional illustrations ......................................................................................................... 32 Annex B. Revised PMT model for STP .................................................................................................. 34 2 LIST OF FIGURES AND TABLES Figure 1 Study Overview......................................................................................................................... 9 Figure 2 Distribution of consumption per adult equivalent across districts ........................................ 10 Figure 3 Example of building footprint outlines on top of Google Maps imagery, Santo Antonio, Principe Island ...................................................................................................................................... 11 Figure 4 Distribution of normalized red, green and blue colors of roofs ............................................. 12 Figure 5 Units of analysis in central São Tomé ..................................................................................... 13 Figure 6 Predicted and Observed PMT Scores ..................................................................................... 16 Figure 7 Importance Scores for RF PMT Prediction Model .................................................................. 17 Figure 8 Average PMT Scores ............................................................................................................... 19 Figure 9 Concentration of poorest 40 percent and location of beneficiaries ...................................... 23 Figure 10 Overlap of bottom 40 percent, VFP beneficiaries and distribution of primary/secondary schools ................................................................................................................................................. 24 Figure 11 Exposure of poor households and cash transfer beneficiaries to severe drought and flash floods .................................................................................................................................................... 25 Table 1 Observed and predicted PMT Sccores ..................................................................................... 15 Table 2 Poverty distribution by sub-district in STP .............................................................................. 20 Table 3 Number and percentage of households and poor expose to flash floods and severe droughts .............................................................................................................................................................. 27 3 ACKNOWLEDGMENTS This report is co-authored by Peter Fisker (University of Copenhagen), Jordi Gallego-Ayala (Senior Social Protection Specialist HAES2), David Malmgren-Hansen (independent consultant), Thomas Sohnesen (independent consultant) and Edmundo Murrugarra (Senior Social Protection Economist HAES1). The report was prepared under the technical guidance of Jordi Gallego-Ayala and Edmundo Murrugarra. This report and its results were discussed with representatives of the Ministry of Labor, Solidarity, Family and Professional Training. Invaluable comments were received from colleagues at the World Bank, including: Scherezad Latif (Human Development Program Leader), Philippe Leite (Senior Social Protection Economist, HSASP) and Ugo Gentilini (Senior Social Protection Specialist, HJSDR). The team is also grateful from the support received from Consuella Andrianjakanava (Team Assistant, HAES2) during the preparation and editing of the report. This report also benefits from the data facilitated by the Directorate of Providence, Solidarity, Social Assistance and Family. The team would like to thank for the funding and support to the Rapid Social Response Trust Fund (ASA; task ID P176471) and the direct support received from the leadership of Jean-Christophe Carret (Country Director AECC2) and Paolo Belli (Practice Manager HAES2). 4 Introduction Poverty reduction is a key Social Development Goal, and one hampered by lack of current and updated data. Unfortunately, filling this data gap through traditional surveys has a high financial cost. The World Bank estimates that monitoring and tracking poverty in the poorest countries will cost USD 945 million between 2016 and 2030 (Kilic et al., 2017). Further, such surveys normally estimate vital statistics at urban/rural and regional levels, while they provide no guidance on planning or differences within specific urban/rural environments. However, the information provided by these datasets are fundamental for the implementation of social protection programs that targets the poorest and most vulnerable population in the selection process of program beneficiaries. Image recognition and machine learning methods hold the potential to reduce the data gaps and thereby pave the road for more efficient anti-poverty policies. Recent progress in the field includes Jean et al. (2016) who use nighttime light data and daytime images to estimate village poverty rates in five African countries. Engstrom et al. (2017) use object identification to identify number and density of buildings, a building height proxy, roof material, and car counts, which -when combined with geospatial data on roads and farmland- can predict municipal poverty rates in Sri Lanka. Similarly, object identification of water source, roof quality and lighting source are used to predict poverty at sub-district levels in Uttar Pradesh, India (Pandey et al., 2018). Object identification of roof quality, but not poverty itself, also provides the basis of allocation of anti-poverty transfers in villages in Uganda and Tanzania (Abelson et al., 2014). The social protection sector is potentially one of the end users of the new detailed poverty maps developed using satellite imagery. In fact, the use of big data and machine learning for the preparation of poverty maps is a window of opportunity to enhance the geographical targeting of the social safety net programs (Gentilini et al., 2021; Aiken et al., 2022; Grosh et al., 2022). This study, which focuses on the case of São Tomé and Príncipe (STP), contributes to the nascent literature by estimating poverty scores for locations of around 110 m x 110 m, which is an improvement on most previous studies that estimate poverty at municipal level or larger level, averaging over thousands of households. Importantly, this study employs methods that run automatically and can be applied at a large scale. Specifically, this working paper shows how High Resolution Satellite Imagery (HRSI) poverty maps developed for STP serve to assist policies and planning for the Government of São Tomé and Príncipe 5 in regards to geographical targeting for rapid expansion of social protection programs based on poverty incidence and population densities. The current geographical targeting in STP for social protection programs is based on poverty maps at district level and then follows a consultative process with local authorities to identify programs expansion entry points. The remainder of the working paper outlines the context of poverty reduction and social protection efforts in STP, a detailed account of the data and methods used in the study, and finally results and visualizations accompanied by notes on operational use and concluding remarks. I. Context A. Country São Tomé and Principe is a low-middle-income, small-island country that faces challenges typical of small states and it has recently been severely impacted by the COVID-19 pandemic. The country consists of two main islands in the Gulf of Guinea with a surface area of 1,001 sq. km. STP has a total population of approximately 200,000 people, with 42.6 percent of the population at or below the age of fourteen. In 2018, the country’s per capita Gross National Income (GNI) was estimated at USD 3,430 in purchasing power parity (PPP), and its per capita Gross Domestic Product (GDP) was USD 2,043. STP’s development challenges are typical of small island nations and include high fixed costs of public goods, limited investments in human development, as well as geographic isolation and small market size which constrain the growth of dynamic, competitive markets. STP was already facing an adverse macroeconomic situation before the COVID-19 crisis. In STP, 35 percent of the population is below the national poverty line and 25 percent of the population live on less than USD 1.9 PPP/day (international extreme poverty line). This extreme income poverty severely constrains chronically poor households from investing in nutrition, health and education. Human development outcomes in STP are low and were further negatively impacted by the COVID-19 pandemic. Trends since 2010 showed slight improvements, though school closings and a healthcare system struggling as a result of COVID-19, have most likely led to roll back those moderate gains registered before the pandemic. STP’s United Nations Human Development Index (HDI) value has increased from 0.561 to 0.625 between 2010 and 2019, placing it above the average for Sub-Saharan Africa (0.547), but below the average for other countries in the medium human development level group (0.631). Strong gains in the country’s HDI are largely attributable to an increase in average life expectancy, a reduction in infant mortality, and an increase in the average 6 years of schooling. From 2010 to 2017, life expectancy at birth increased from 65.9 years to 66.8 years, while infant mortality rates decreased from 33.5 to 25.2 (out of 1,000 live births). Women and youth are at high risk of poverty due to unemployment and low labor market participation. Extreme poverty is greater among children and in households headed by women: 40 percent of households in STP are headed by women, while accounting for 42 percent of the poorest 20 percent of households in the country and only 27 percent of the wealthiest 20 percent. Women are less likely to enter the labor market due to fewer work opportunities and a skills mismatch between the skills they possess and those needed by the labor market. The unemployment rate among women is three times higher than that of men (14.5 percent compared to five percent) (IOF 2017). Early pregnancies for girls and gender-based inequities in education and opportunities represent critical binding constraints behind lagging results in labor market outcomes. Furthermore, one in every three women has been a victim of physical domestic violence since the age of 15 (33 percent), a phenomenon particularly acute among low-income women, among which almost half (46 percent) have suffered physical domestic violence. B. Poverty and social protection The Government of STP (GoSTP) is committed to reducing poverty in the country by, among other interventions, taking steps to put in place a framework to strengthen the social protection system. The social protection legal framework in STP is based on the Social Protection Law approved in 2004 (Lei n.° 7/04, Lei de Enquadramento da Protecção Social) and the Social Protection Policy and Strategy (PENPS) approved in early 2014. The administration of the Social Protection system is managed by the Ministry of Labor, Solidarity, Family and Professional Training (MLSFPT) and the Directorate of Providence, Solidarity, Social Assistance and Family (DPSSF) under the guidance of MLSFTP implements the SP programs. STP has made significant progress in social protection in recent years and is guided by the National Social Protection Strategy. The strategy supports three social protection programs: (i) the social pension program; (ii) the Vulnerable Families Program (VFP); and (iii) the labor-intensive public works program. The GoSTP is also implementing the VFP-COVID-19 shock response program to provide support to households directly or indirectly affected by COVID-19. • The social pension program targets poor elderly 60 years and older (excluding those that receive a contributory pension) and all poor with disability or chronic disease who cannot 7 work. The program is covering 3,045 households and beneficiaries of the program receive a monthly benefit of STDB 150 (USD 7); however, the payment is delivered every three months. The payments of the program are delivered by DPSSF using a cash manual approach, but this program has suffered historically for delays in delivery the cash transfers due to financing constraints of the program that is fully funded by the Government. • The Vulnerable Families Program is a conditional cash transfer targeting poor households mainly headed by women that have at least one child below 18 years old who is attending school. The program has a coverage of 2,543 beneficiaries (97 percent of the beneficiaries are women and 87 percent are single mothers) with a monthly benefit level of STDB 600 (USD 30). Payments are delivered bimonthly and in September of every year beneficiaries of the program receive a top-up to the regular payment of additional STDB 600 to support the household with schools’ fees enrollment and other school related costs. The delivery of the payments is done through a payment service provider. The VFP has accompanying measures (Parent Education Program) to promote the development of vulnerable children by providing parents with tools and knowledge, nudge behaviors towards optimal child development and preventing gender and child violence in the home, among others. • The Vulnerable Families Program – COVID-19 program is an unconditional cash transfer rolled-out to provide support to poor and vulnerable households affected directly or indirectly by the COVID-19 crisis. The program has a coverage of 15,325 households (78.5 percent of the beneficiaries are women) with a monthly benefit of STDB 900 (USD 42). The payments of the program are delivered every two months by a payment service provider. The program has a duration of nine months. Significant progress has been made in building delivery systems, improving sector effectiveness, transparency and accountability. A fully functional Management Information System (MIS – Sistema de Informação Integrado de Protecção Social - SIIPS) serves as an entry point for intake, registration and enrollment of beneficiaries in the SP programs and allows interoperability with commercial banks to trigger automatic payments to beneficiaries. The SIIPS has a social registry with more than 15,500 households registered with a complete set of socioeconomic information. The formal payment mechanism through commercial banks allows timely and reliable payments to beneficiaries and increased transparency in the monitoring and reconciliation of funds allocated to social protection payments. Finally, regarding the targeting of the programs; DPSSF uses a two-step 8 methodology to target program beneficiaries. First, it uses geographical targeting and then it uses a Proxy Means Test (PMT) to verify beneficiaries’ eligibility for a given program. The geographical targeting uses poverty maps at district level, which does not provide enough granularity for the selection of municipalities and neighborhoods in each district where the programs should be implemented. This step currently relays on subjective criteria, based on government priorities and perceptions, which leads to large exclusion errors and under-coverage by the social protection programs of areas with high poverty incidence. Unfortunately, the budget allocation in STP for social protection does not provide sufficient coverage and generosity through the three above-mentioned safety net programs. In 2020, the country budgeted less than 0.25 percent of the GDP to social protection programs which is well below the African regional average of 1.2 percent of GDP. The social protection program coverage remains low, with permanent SP programs (excluding the Vulnerable Families Program – COVID-19 program) covering around 5,500 poor and vulnerable households, accounting for 25 percent of total poor households in the country. II. Data The poverty maps described in this working paper rely on a set of various data types from different sources. This section provides a short description of the data used to model and estimates poverty. Figure 1 shows an overview of the components of the study, and the remainder of this section describes the various data types represented by the green boxes on the left-hand side. Figure 1: Study Overview 9 A. Welfare data Consumption expenditure data is the basis of estimated welfare in the most recent household survey (Inquérito aos Orçamentos Familiares (IOF), 2017). Consumption expenditure data is also what is used to measure monetary poverty in STP. To estimate households’ welfare based on a limited number of questions, ultimately deciding whether households are eligible for public support through the social safety net programs, the consumption expenditure data is proxied through a Proxy-Means-Test (PMT) score. The IOF consumption data was recently re-calculated and based on this a new PMT model was constructed (see Annex 1). The mapped PMT scores are based on this new and revised PMT model. The survey data (IOF, 2017) itself shows that consumption is higher in Lobota and Lemba districts, but especially on the Autonomous Region of Principe (Figure 2). There is also notable variation within each region, with households among the poorest 25 percent in all seven regions. Unfortunately, the survey only covers a small part of the country, while the poverty map will further detail all locations that are generally poorer. Figure 2 Distribution of consumption per adult equivalent across districts 10 B. Data on buildings extracted from satellite images For this study, building footprints covering all STP have been obtained in on Shapefile with 79,294 features ("house structures") 2. A subset of these features can be seen in Figure 3. From this data, it has been calculated the average building area, number of buildings, and average perimeter length for each cell, or Unit of Analysis (UA, see Fig. 5). These variables are vital in predicting PMT scores as richer families tend to own larger houses with relatively more space around (depending on location). Figure 3 Example of building footprint outlines on top of Google Maps imagery, Santo Antonio, Principe Island To gain further information on the quality of the buildings underlying the footprints, the building footprints features were overlaid optical satellite images. Once combined, information on roof color and other image features derived using Convolution Neural Networks (CNNs) were extracted from the optical images for each building. 2 A building footprint is the perimeter outline of a structure seen from above, and typically extracted from satellite images. Building footprints for this study were obtained from Ecopia and other satellite image were obtained from Maxar. 11 The satellite images were originated from the GeoEye-1 satellite. The images contain 4 color bands, blue, green, red and near-infrared, in 0.5 m x 0.5 m pixel resolution. To cover STP the study relies on 41 different satellite images. Clouds are not masked out in these satellite images and can corrupt the color values in some cases. Cloud cover is not a concern as it is small and mostly present in locations with no or very limited structures. However, the images do not contain atmospherically corrected color values, so color and intensity vary with sunlight conditions. Hence, the color value of the same structure may differ from two different images as the sunlight conditions differ between the two. To overcome this challenge, the original coloring values were re-scaled to percentages of red, green and blue colors within each structure. Based on these, all structures were classified into the following roof colors: i) orange/red (associated with high quality roofs), ii) grey/silver (associated with medium-low quality roofs), iii) brown (associated with lower quality structures made of natural materials), and iv) blue. Figure 4 shows the full three-dimensional distribution along the red, blue and green dimensions. Figure 4 Distribution of normalized red, green and blue colors of roofs* *Notes: The color used in the graph is not the colors as observed on the ground, but merely a representation. 12 Second, a pre-trained Convolutional Neural Network (CNN) is used to extract characteristics of all UAs. The CNN looks for commonalities or trades, that might not per se be visible to the naked eye. A CNN trained to categorize everyday objects (Deng et al.,2009) using the architecture DenseNet (Huang et al., 2017) was used. Though not developed specifically to the challenges discussed in this paper, the network might still generate computer vision features that could be used to distinguish richer and poorer areas. C. Geospatial data Geospatial data on roads and distances are added for each unit of observation. These include the distance to the city center (separately for São Tomé and Príncipe islands), the total length of primary and residential roads as well as the distance from the center point of each UA to the nearest road; total road length is also included. Information on extent and classification of road network is derived from OpenStreetMaps. D. Linking the various data sources In order to generate a prediction model, that can predict welfare levels for the islands, the various data sources need to be linked. They are linked by location based on a grid of cells referred to as UAs. Each UA is approximately 110x110m large (0.001 degree longitude-latitude) as illustrated in Figure 5. Figure 5 Units of analysis in central São Tomé Notes: Red grid cells indicate the UAs, namely cells of 0.001 degree longitude-latitude (ca. 110 meter). Green polygons show building footprints, and turquoise lines show the road network. 13 III. Predicting Neighborhood Poverty Levels Poverty mapping has traditionally (since the early 2000’s) been based on combining census and survey data and provided estimates of poverty at city and district levels (Lanjouw et al., 2003). The approach in this working paper can be seen as a second generation of poverty mapping in that it relies on geospatial and satellite data as opposed to census data. The former is more frequently available and up to date, as opposed to census data that is available every 10 years on average. Hence, when using satellite data, the social protection targeting can be updated at low cost frequently as new images become available. Further, HRSI poverty maps also predict eligibility for program participation at a more local level as opposed to cities and districts level maps. The HRSI approach to poverty mapping is still recent, but similar examples exist in other countries, e.g. Mozambique, where such estimates were used to guide the expansion of an urban social protection program (Sohnesen et al., 2022). Other examples of satellite date used for poverty estimation include work using nighttime light data (Jean et al., 2016). Object identification of buildings and their quality as well as cars, combined with geospatial data on roads and farmland, were used to predict municipal poverty rates in Sri Lanka (Engstrom et al., 2017). Similarly, object identification of water source, roof quality and lighting source are used to predict poverty at sub- district levels in Uttar Pradesh, India (Pandey et al., 2018). Identification of roof quality, but not poverty itself, also provides the basis of allocation of anti-poverty transfers in some villages in Uganda and Tanzania (Abelson et al., 2014). Beyond poverty, local urban wealth, based on an asset index, has also been modeled based on satellite-derived land-use and cover (Georganos et al., 2019) and pockets of deprivation (Wang et al., 2019). This working paper estimates the average PMT Scores for each UA using a Random Forest (RF) prediction model. Application of RF for poverty predictions is at an early pioneering stage, though evaluation of the method compared to alternatives have been found favorable (McBride et al., 2016; Rusnita et al., 2020; Sohnesen et al., 2017). As illustrated in Figure 1 the RF models combines the various data sources into a joint model via the spatial overlap based on the UAs. The RF model is based on 1,042 UAs that have data on the PMT scores from the IOF survey. Each UA has between 1 and 14 household observations, with an average of 2.3 households observations. The RF model select the best predictors of PMT Scores among the geospatial and building characteristics outlined in the data section. 14 To evaluate accuracy of the prediction model, the 1,042 UAs are split into 90 percent of the sample used for training the RF model, leaving out a random 10 percent (within each of the two Islands) for evaluation of the prediction model. The prediction model has an out of sample R^2 of 0.38 and a rank correlation of 0.65. The notably higher rank correlation than R^2 indicate that the model does a better job at ranking households correctly than estimating their absolute Poverty Score level. Figure 6 shows the observed and predicted PMT scores, for both the with-in and out-of-sample observations. The concentration along the diagonal line in Figure 6 shows that the prediction model does a good job in replicating the ranking small areas (UAs) from rich to poor. This can be seen for both the observations used to generate the model (the light grey squares) as well as the out-of- sample observations (the black triangles). The model predicts Poverty Scores in a smaller range than what is observed in the survey (the difference between the X and Y axis). This could, in part, be a reflection of the standard errors in the original PMT scores that are based on a sample only. The HSRI poverty map add benefits by generating estimates of PMT Score at very local levels, something that was not available before where the household survey was the only source of data. Table 1 adds to Figure 6 and shows that the prediction models have same estimated means as the means observed in the survey. Hence, combined Figure 6 and Table 1 indicates that the predictions have good accuracy. Table 1 Observed and Predicted PMT Scores Mean PMT Score from survey Mean estimated PMT Scores HSRI model District (Previous targeting data) (New targeting data) Lobata 14.3 14.3 Mezochi 14.2 14.2 Cantagalo 14.3 14.3 Principe 14.5 14.5 Lemba 14.1 14.2 Aqua Grande 14.1 14.1 Caue 14.9 14.9 National 14.4 14.4 The importance of different predictors in RF models can be expressed by Importance scores. The RF models finds that location itself is very important, which is clearly seen by the high Importance scores 15 of districts. This also supports doing a spatial targeting of the program first. In addition to districts, there are contributions from the various types of spatial data, including the presence and distance to roads, the average size of structures, population in area, presence of vegetation captured by inferred light, as well as the share of structures dominated by red (orange) roofs, associated with high quality roofs (see Figure 7). The variable Feature 3 is a characteristic of the locations extracted from the CNN model. Figure 6 Predicted and Observed PMT Scores 16 Observed Poverty Score 13 14 12 15 14 14.5 15 15.5 Estimated Poverty Score RF Within sample Out-of-sample 16 Figure 7 Importance Scores for RF PMT Prediction Model IV. Outcomes Figure 8 shows the distribution of predicted PMT scores across all built-up areas of STP. Each pixel represents an UA of around 110m * 110m and a higher score (blue) is associated with richer neighborhoods, whereas lower scores (red) show poorer areas. White areas are UA where no building is located. In general, localities in the Autonomous Region of Príncipe and the central areas of the city of São Tomé (Agua Grande district) are estimated to be better off while rural and coastal areas outside the main city tend to be poorer. As is also evident from Table 1, the districts with the highest poverty rates include Caue, Lemba, and Cantagalo while the richest districts are Príncipe and Água Grande. Figure A1 in the Annex shows the distribution of poverty scores within the capital district of Agua Grande using an alternative scale to highlight the variation that exists within the city despite most of the UAs belonging to the higher end of the national distribution. Figure A2 illustrates how the outputs can be used in combination with Google Earth in order to zoom in on specific areas of interest and assesses the locations and characteristics of dwellings associated 17 with certain predictions. The highlighted excerpt shows how areas with higher estimated poverty scores in general have smaller houses, located more remotely from local infrastructure. Detailed HRSI poverty maps combining poverty incidence and population density were developed to support the operational use of the maps in the geographical targeting process. Departing from the distribution of predicted Poverty Scores, these maps show the location of the bottom quartile or bottom 40 percent of the population. All UAs represented on the maps have average Poverty Scores belonging to this low end of the distribution, while the colors on the map show the population density of the cell. This allows implementers to select rollout entry points with a high concentration of poor and vulnerable families as opposed to simply looking at the average poverty levels shown in Figure 8. Figure 9 shows HRSI maps combining poverty incidence and population density maps (bottom 40 percent) for São Tome island and for Agua Grande City; and Table 2 shows the poverty incidence by sub-district in STP. 18 Figure 8 Average PMT Scores (a) São Tomé (b) Príncipe 19 Table 2. Poverty distribution by sub-district in STP Households in Households in Total Poverty District Sub-district bottom 20 bottom 40 Households rate percent percent GUADALUPE 1744 61.9% 467 1080 CONDE 702 72.6% 166 510 LOBATA MICOLÓ 533 83.4% 327 444 SANTO AMARO 1967 56.3% 317 1107 NEVES 2809 87.9% 2051 2470 LEMBA SANTA CATARINA 695 89.6% 588 623 TRINDADE 4558 56.6% 1225 2580 BOMBOM 2344 38.9% 133 913 ME-ZOCHI MADALENA 787 73.0% 379 575 CAIXAO GRANDE 1780 40.9% 69 728 ALMAS 1318 43.2% 76 570 SÃO-TOMÉ 16498 6.3% 499 1042 AGUA GRANDE PANTUFO 839 28.4% 48 238 SANTANA 2766 76.8% 765 2126 CANTAGALO RIBEIRA AFONSO 1592 88.8% 729 1414 SÃO JOÃO DOS 1034 91.4% 699 945 CAUE ANGOLARES MALANZA 383 97.4% 339 373 RAP SANTO ANTÓNIO 1999 0.3% 2 7 Total 44348 8879 17744 A. Operational use From an operational point of view in the social protection sector, the maps can support four main areas of program implementation, namely: (i) geographical targeting; (ii) enrollment strategies; (iii) enrollment plans; and iv) budgeting. São Tome Island (see Figure 9b) and Agua Grande city maps are going to be used to illustrate the different operational options for the scale up of the Vulnerable Family Program. Figure 9a shows the Agua Grande city map for the bottom 40 percent. Geographical targeting HRSI poverty maps target areas of poverty incidence and population density and thus can be used to scale up the VFP program. The example of São Tome Island and Agua Grande are used to illustrate the use of the poverty maps for the geographical targeting for several reasons. First, the map shows the geographical location of the cells with high estimates of poverty --poorest quartile- and the darker cells show the ones with the highest number of households in the poorest quartile. Second, the maps 20 shows poverty pockets and how scattered the poverty pockets are distributed in São Tome Island and Agua Grande city, as well as the poverty asymmetries within specific neighborhoods. Thus, the decision is where to concentrate efforts for a rapid scale up of a given program, in this case the VFP. For these reasons, the geographical targeting should be focused on those areas with a high concentration of poverty and high population density. Needless to say, when the geo-location of actual social safety net programs’ beneficiaries is available, it should be an additional variable to consider in the enrollment entry points identification to prioritize areas with under-coverage; for the case study presented in this working paper the geographical location of the VFP beneficiaries is available. For São Tome Island (Figure 9b), the map shows that the VFP is well targeted in terms of geographical locations since focused mainly in those areas with high poverty incidence. However, in case the VFP needs to be scaled up at national level, the map allows to identify clearly the geographical areas in Me-Zochi District (Trinidade and Caixão Grande sub-districts) that can be targeted for the program expansion (see red circle in the map). For the case of Agua Grande city (Figure 9a) for example, three main geographic areas are identified as targeting areas in the maps (see red circles in the map). The targeting areas identified as priority -areas that extend across different neighborhoods- should be used as entry points to carry out beneficiaries’ enrollments and from those areas the program should expand. Using enrollment entry points in urban areas allows to move the operational conversation from a neighborhoods approach to one across neighborhoods focusing on areas that provide higher flexibility for programs. In previous use of urban poverty maps, a strict approach limited the geographical targeting to the administrative boundaries of the neighborhoods. As a result, the social safety net programs faced operational limitations in the field. In contrast, the use of the enrollment entry points provides much more flexibility to the programs and captures the dynamic expansion of poverty across neighborhoods. The use of a bottom 20 percent map allows to narrow the geographical targeting, but if there is a need to keep flexibility in the approach, the bottom 40 percent map can be used even if the entry points are the same. Urban and peri urban settings are very dynamic environments and depending on the data used to produce the map, new settlements can appear rapidly, and those are not captured in the map. Thus, flexibility in the approach must be applied when the enrollment teams are deployed in the field. 21 Enrollment strategies, plans and costing Various enrollment strategies could be applied depending on poverty distribution within the enrollment entry points. Government can then decide on a strategy, such as to cover the bottom 20 percent or the bottom 40 percent of the population. Once the strategy is set, enrollment can be done. First, in enrollment areas with high poverty concentration, the social assistance institutions can rely on local structures to rapidly enroll potential beneficiaries. Second, for enrollment entry points with a high dispersion of the poor households within the enrollment area, social assistance institutions can piggyback on referral systems from other government institutions. For instance, Ministries of Education can provide a list of parents of school-aged children in those selected areas to prepare list of potential beneficiaries to be enrolled. Figure 10 shows the overlap of HRSI with VFP beneficiaries distribution and primary/secondary schools that can support this additional enrollment process. Once enrollment strategies are identified, the enrollment plans can be prepared for different program coverage strategies. The identification of the enrollment entry points allows to calculate the potential number of beneficiaries to be covered by area. Hence, since the potential number of beneficiaries is known by targeting area, the enrollment plans can be prepared. For example, enrollment plans can be developed to cover the bottom 20 percent or the bottom 40 percent of the population. To illustrate this, the Agua Grande city example is used in this regard. Agua Grande city presents six neighborhoods in the identified enrollment areas. Thus, the program scale up can range from 150 to 350 households depending on whether the strategy is to cover, respectively, the 20 or 40 percent poorer segment of the population. Based on the alternatives for the program implementation in terms of coverage the operational alternatives and their costs can be drafted. 22 Figure 9 Concentration of poorest 40 percent and location of beneficiaries a) Within Agua Grande b) São Tomé 23 Figure 10. Overlap of bottom 40 percent, VFP beneficiaries and distribution of primary/secondary schools Adaptive social protection3 HRSI combined with natural hazards maps allows to prepare detailed exposure maps that can support the Adaptive Social Protection (ASP) agenda in the country. In fact, the geographical targeting for ASP is fundamental since will facilitate a rapid scale up of the social protection programs in shock response scenarios as well as during the planning, shock preparedness and preparation of contingency plans. Thus, the exposure HRSI maps are a fundamental tool to identify prone areas to shocks and plan future responses. Figure 11 shows poor households and current VFP beneficiaries exposure maps to severe drought and flash floods. 3 This sub-section is based on the report prepared by Beazley, Gallego-Ayala, Marzi, and Cereda (2022). “São Tome and Principe - Adaptive Social Protection Readiness Assessment and Policy Options.” 24 Figure 11. Exposure of poor households and cash transfer beneficiaries to severe drought* and flash floods** *The intensity is defined as the magnitude*duration of a drought, which in turn is defined as a range of values in the Standardized Precipitation-Evapotranspiration Index (SPEI). 2.1 is the value of the upper quintile of drought intensity STP according to the data available. ** The estimation considers the number of households located in areas where a 1-in-100 years pluvial flood results in a water depth above 0.5 meters. 25 The HRSI exposure maps allows to identify the areas where poor and vulnerable households and VFP beneficiaries are extremely exposed to natural disasters. The use of the HRSI exposure maps in STP would provide operational flexibility to the social protection sector to response to shocks from two different angles; taking into account, that currently the VFP doesn’t consider the households’ exposure to multiple-climate related hazards: First, it can enhance the geographical targeting of the VFP focusing on areas prone to natural disasters to enhance the resilience of the households against climate change; thus, further scale-ups of the VFP can prioritize areas highly exposed to natural disasters. Second, they support the preparedness to shocks by identifying areas to carry out pre- registration activities of potential beneficiaries in highly vulnerable areas to shocks. This action will allow a pool of potential beneficiaries to be pre-identified in the Management Information System, that can be rapidly derived to the shock response program when is deployed in a territorial area already covered by the pre-registration exercise. The HRSI exposure maps could also be used in the contingency planning exercise since it is possible to estimate the number of poor and vulnerable households exposed to droughts and flash floods. Estimates based on data from IOF 2017, shows that more than 33,000 households (see Table 3) is located in areas exposed to drought, of which slightly more than 10,000 households are poor. Floods, by contrast, affect fewer households; estimates shows that more than 2,000 households are exposed to flash floods of above 0.5 meters; of these households, approximately 600 are poor. In addition, the beneficiaries of the cash transfer program VFP are also greatly exposed to droughts, and less so to floods. 26 Table 3. Number and percentage of households and poor expose to flash floods and severe droughts Exposure to droughts (intensity > 2.1 Exposure to floods (>0.5m) SPEI) District % of % of N poor % of total N total N poor % of total N total HHs total total HHs population HHs HHs population poor poor CANTAGALO 397 354 9.1% 10.0% 4037 3231 92.6% 91.3% CAUÉ 55 55 3.9% 4.2% 96 96 6.8% 7.3% LEMBÁ 44 44 1.3% 1.4% 0 0 0.0% 0.0% LOBATA 124 94 2.5% 3.0% 3758 2285 76.0% 72.8% MÉ ZÓCHI 66 32 0.6% 0.6% 8119 3384 75.3% 63.1% PRÍNCIPE 454 0 22.7% 0.0% 0 0 0.0% 0.0% ÁGUA GRANDE 1033 23 6.0% 1.8% 17337 1281 100.0% 100.0% Total 2174 601 4.9% 3.4% 33347 10277 75.2% 57.9% V. Discussion and concluding remarks This study contributes to more accurate and efficient targeting of social protection programs by providing current and detailed maps of poverty across STP. It is achieved through use of survey and GIS data, optical satellite images, as well as methods commonly applied in data science and in the field of image recognition. Survey data, and partly GIS data, are commonly used in various aspects of poverty analysis, while the use of optical satellite images analyzed through image recognition, and the use of prediction methods from data science are not. This study showcases an operational proposal of geographical targeting for rapid expansion of social safety net programs based on poverty incidence and population densities using as example the HRSI poverty maps developed for STP. The HRSI poverty maps presented in this study show the geospatial location of the areas (cells) with high estimates of poverty and the areas with the highest number of families in the poorest percentiles. These maps can be produced using cells with e.g. the bottom 20, 40 or 50 percent to provide more flexibility to the policy dialogue. The poverty incidence and population density binomial allows the identification of the enrollment entry points, as well as estimate the number of households to be enroll in a given social safety net program per area. The methodology used also allow quick and cheap updates to the targeting since the models developed can be replicated using new images. This is a significant comparative advantage to poverty maps based on census data, for instance. Four main conclusions from operational point of view can be drawn: 27 • Geographical targeting using HRSI poverty maps can be used to support rapid scale-up of programs; this can be a critical element specifically in shock response scenarios were social safety net programs needs to be roll-out rapidly i.e. during the COVID-19 pandemic. • HRSI poverty maps allow to identify program enrollment entry points based on poverty incidence and population density; other relevant variables can be added as needed or available to identify those enrollment entry points i.e. social safety net beneficiaries’ distribution, exposure to hazards etc. • HRSI urban poverty maps support enrollment plans since it is possible to estimate HHs to be covered in each enrollment area. • The maps allowed to move the operational conversation in social protection to one across neighborhoods focusing on areas (cell clusters) which provide more flexibility to the policy dialogue. The previous lack of accurate poverty maps in STP has resulted in the social protection programs under-coverage in areas with high poverty concentration. Thus, the HRSI poverty maps allow the Government of STP to gain efficiency in the programs geographical targeting and to have a transparent and objective criterion for the selection of the municipalities and neighborhoods that could be part of those programs. Moreover, the HSRI poverty maps in STP could enhance the efficiency of the geographical targeting process of the social safety net programs and support the Government in developing the programs’ coverage expansion strategies. The layering approach for the geographical targeting presented in this study would allow the Government of STP to concentrate their efforts for rapid scale up of the programs in those specific enrollment areas. Looking down the road, the possibility to combine the HRSI poverty maps with poverty population exposure to hazards opens the opportunity to sharpen the ability of social protection programs to contribute to a climate adaptation strategy, by facilitating the identification of areas with high poverty incidence and highly exposed to shocks that can be prioritized in the expansion of the social safety net programs. This methodology can be used not only for SP programs but for other social policies as well, such as deployment and allocation of social services like schools of health care centers where poverty drives the type and size of facility. The HRSI poverty maps also has a potential use in areas of high mobility or uneven/very vulnerable infrastructure (e.g. displacement, post war) where geospatial maps can help identify population concentrations and quality of infrastructure to 28 allocate resources (the 2000 Kosovo Household Survey used geospatial information to create a sample frame since the latest census was outdated given the conflict in the area). In African countries with high exposure to the effects of the climate change and conflict related displacement this type of technology could be very valuable to guide the deployment of social safety nets programs. Finally, to roll-out the methodology presented in this paper there is a need to support the Governments in enhancing their capacity to manage and use this approach. In fact, the methodology presented is valuable for policy making to the extent that the government is comfortable and in control of the tool. This poses further efforts in increasing government capacities to do it. 29 VI. References Abelson, B., Varshney, KR. and Sun, J. (2014) "Targeting direct cash transfers to the extremely poor", 20th ACM SIGKDD international conference on Knowledge discovery and data mining, New York, New York, USA. Aiken, E., Bellue, S., Karlan, D; Udry, C. and Blumenstock, J. (2022) Machine learning and phone data can improve targeting of humanitarian aid. Nature 603, 864–870, 2022. Deng, J., Wei, D., Socher, R., Li, L., Li, K., and Fei-Fei, L. (2009) “Ima- genet: A large-scale hierarchical image database”. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Elbers, C., Lanjouw, J.O. and Lanjouw, P. (2003) “Micro-Level Estimation of Poverty and Inequality.” Econometrica 71, no. 1, 2003: 355–64. Huang, G., Liu, Z., Van Der Maaten, L. and Weinberger, K. (2017) Densely connected convolutional networks. In Proceedings of the IEEE confer- ence on computer vision and pattern recognition, pages 4700–4708, 2017. Gentilini, U., Saksham, K. and Almenfi, M. (2021) Cash in the City : Emerging Lessons from Implementing Cash Transfers in Urban Africa. Social Protection and Jobs Discussion Papers 2101. The World Bank, Washington, DC. Georganos, S., Gadiaga, A.N., Linard, C., Grippa, T., Vanhuysse, S., Mboga, N., Wollf, E., Dujardin, S. and Lennert, L. (2018) "Modelling the Wealth Index of Demographic and Health Surveys within Cities Using Very High-Resolution", Remotely Sensed Information Remote Sensing. Grosh, M., Leite, P., Wai-Poi, M. and Tesliuc, E. (2022) Revisiting targeting in social assistance: a new look at old dilemas. The World Bank, Washington, DC. Jean, N., Burke, M., Xie, M., Davis, W.M., Lobell, D.B and Ermon, S. (2016) "Combining satellite imagery and machine learning to predict poverty", Science, 353, 790-794, 2016. Kilic, T., Serajuddin, U., Uematsu H. and Yoshida, N. (2017) Costing Household Surveys for Monitoring Progress Toward Ending Extreme Poverty and Boosting Shared Prosperity. Policy Research Working Paper 7951. The World Bank, Washington, DC. Mcbride, L. and Nichols, A. (2016) Retooling poverty targeting using out-of- sample validation and machine learning. The World Bank Economic Review. 30 Hamdan, R.I., Bakar, A. and Sani, A. (2020) Does artificial intelli- gence prevail in poverty measurement? Journal of Physics: Conference Series, 1529:042082, 04 2020. doi: 10.1088/1742-6596/1529/4/042082. Sohnesen, T. and Stender, N. (2017) Is random forest a superior method- ology for predicting poverty? an empirical assessment. Poverty & Public Policy, 9(1):118–133, 2017. Sohnesen, T., Fisker, P. and Malmgren-Hansen, D. (2022) Using satellite data to guide urban poverty reduction. Review of Income and Wealth,. doi: https://doi.org/10.1111/roiw.12552. URL https:// onlinelibrary.wiley.com/doi/abs/10.1111/roiw.12552. Wang, J., Kuffer, M., Roy D. and Pfeffer, K. (2019) Deprivation pockets through the lens of convolutional neural networks. Remote Sensing of Environment, 234, 2019. 31 Annex A: Additional illustrations Figure A1: PMT scores in Agua Grande (alternative scale): 32 Figure A2: Example using Google Earth 33 Annex B Revised PMT model for STP Background In 2015, under the Rapid Social Response (RSR) Trust Fund - Building Blocks for the Social Protection System in São Tomé and Principe (P149534) a PMT model for São Tomé and Príncipe was developed. The model was based on the 2009/2010 Inquérito aos Orçamentos Familiares (IOF) representative household survey. In 2020, this model was updated based on the newer IOF 2017 (Fisker, 2020). However, further detailed analysis of the underlying consumption data revealed some quality issues that lead to a reduction of the number of observations of sufficient quality for poverty analysis. The measurement of poverty had also changed from per capita to per adult equivalent. Reflecting both these changes its necessary to reassess and revise the PMT model updated on 2020. Reassessment of the PMT model To reassess the PMT model three alternatives have been tested: 1. Use same existing PMT model with new coefficient based on the newly cleaned consumption data measured in per adult equivalent. 2. Revise the PMT model, reflecting the new consumption data, but allowing a different model based on the data currently being collected for the PMT model. 3. Expand the revised PMT model with three extra durable assets that are available in IOF 2017, but not currently part of the data collection for the existing PMT model. The three alternatives PMT models can be seen in Table A.1, and key observations to notice are: • Updating the consumption data make a large difference in coefficients of the model as well as the overall fit, capturing quality, of the model. The correlation between consumption and the explanatory variables, as indicated by R-square, increase from 0.27 to 0.46. • Revising the model, by changing the definitions of data already collected also improves the model, leading to a further increase in R-square from 0.47 to 0.50. That is, there is no new data used, in fact the model is more parsimonious than the existing, only how variables are defined changes. The changes are in the demographic sections, where the revised model allows for different coefficient for different household sizes, as well as the inclusion of the dependency ratio 4 capturing the relationship between work-able family members and those that depend on support from the family. Allowing different households sizes to have different coefficients as opposed to having one coefficient for each additional household member allows more flexibility in the relationship between consumption and household size. Further, the dependency ratio also seems to be important. The revised model also excludes most of the location variables. This is done for two reasons: first, few of them are significant. Secondly, location variables are time-invariant, which is an unfortunate quality for a PMT model. For instance, if a specific location has a surge in wealth, then ownership of durable assets will increase for households living there reflecting the improved welfare. However, the coefficient in the PMT model for location will not change and therefore not reflect the improvements in 4 Defined as those aged 15-65 divided by the household size. 34 wealth (because its time-invariant). • The IOF 2017 have data on some durable assets that are not included in the existing model. Adding these additional variables to the model, increase r-square to 0.52, compared to 0.50. Table A.1 PMT Alternatives Old model before Old model after New model after New model after data clean data clean data cleaning data cleaning added questions (1) (2) (3) (4) Log consumption Log consumption Log consumption Log consumption Household head Female -0.068* -0.086*** No education -0.093 -0.092** -0.081* -0.069 Secondary 0.144*** 0.115*** Higher education 0.367*** 0.149** 0.256*** 0.199*** Self employed 0.198*** 0.113*** 0.091*** 0.095*** Demographics Household size 0.012 -0.188*** Number children 0.077*** -0.024 Household size=2 -0.541*** -0.543*** Household size=3 -0.790*** -0.793*** Household size=4 -0.946*** -0.953*** Household size=5 -1.095*** -1.111*** Household size=6 -1.226*** -1.239*** Household size=7 -1.270*** -1.276*** Household size>7 -1.442*** -1.448*** Dependency ratio -0.303*** -0.264*** Assets Number of rooms 0.087*** 0.074*** 0.066*** 0.058*** Piped water 0.03 0.034 Wood for cooking -0.190*** -0.152*** -0.136*** -0.115*** Septic tank toilet 0.064 0.143*** 0.143*** 0.132*** Radio 0.175*** 0.169*** 0.182*** 0.159*** Tv 0.196*** 0.088** 0.135*** 0.104*** Car 0.489*** 0.589*** 0.597*** 0.528*** Computer 0.165*** Cellphone 0.189*** Location Lobata 0.039 0.032 Lemba -0.012 0.029 Agua_grande 0.035 0.143*** Caue 0.172** -0.015 Principe 0.272*** 0.479*** 0.407*** 0.393*** Other # primary education -0.004 -0.009 # primary education 0.084** 0.036* # public employees 0.123*** 0.067*** _cons 17.312*** 14.664*** 14.907*** 14.840*** N 3037 2729 2729 2729 r2 0.266 0.457 0.504 0.521 * p<0.10, ** p<0.05, *** p<0.01 35 Evaluating the PMT models Following the common practice, the PMT model is evaluated according to their inclusion, exclusion and targeting errors defined as follows: a) Exclusion errors – The share of truly poor not identified as poor according to the PMT model. b) Inclusion errors – The share of poor according to PMT model that are not truly poor. c) Targeting errors – The combination of Exclusion and Inclusion errors, i.e the proportion erroneously identified as either poor or non-poor according to the PMT model. Hence, inclusion errors are the share of non-poor households that joins a poverty targeted program, while exclusion errors is share of poor households that did not enter to the program, though qualifying for the program. Combined the two errors are the overall targeting error. Table A.2 shows the inclusion, exclusion and overall targeting error in IOF 2017, if the PMT model was used to target the 20 or 40 percent poorest population. Reflecting the better fit of the predictions models as observed in Table A.1, the revised model with additional variables is best at targeting the poorest (as seen by the lowest targeting error). However, the gains are not very large, so one could argue for keeping the current model and just update the coefficients. However, the revised model based on exactly same data can be implemented without any change in data collection and might have some better qualities over time (which cannot be evaluated without new rounds of survey data) as it does not rely on time-invariant variables. Figure A.1 further illustrates the relationship between predicted and actual consumption. The better performance of the new data models is illustrated by the tighter fit around the 45-degree line. Based on the evaluation, New model I, the same model with updated coefficients, was selected as the new PMT model. This model has a good precision and requires limited change to model predictions and no changes to data collections instruments and implementation. Table A.2 Inclusion, exclusion and targeting errors of PMT models 20th percentile 40th percentile Exclusion Inclusion Targeting Exclusion Inclusion Targeting error error error error error error New model I: Some model, 62% 11% 29% 40% 15% 29% new coefficients New model II: New model, same data collection 60% 10% 27% 41% 16% 30% instrument New model III: New model, new data collection 60% 9% 27% 40% 15% 28% instrument Notes: Inclusion, exclusion errors are evaluated based on the 20 and 40 percentile of the actual consumption distribution and the predicted, respectively. 36 Figure A.1 Relationship between prediction and actual consumption value 18 New model I New Model II New model III 18 18 17 17 17 Consumption per adult equivalent Consumption per adult equivalent Consumption per adult equivalent 16 16 16 15 15 15 14 14 14 13 13 13 12 12 12 12 13 14 15 16 17 18 12 13 14 15 16 17 18 12 13 14 15 16 17 18 Fitted values Fitted values Fitted values 37 Social Protection & Jobs Discussion Paper Series Titles 2020-2022 No. Title 2212 Guiding Social Protection Targeting Through Satellite Data in São Tomé and Príncipe by Peter Fisker, Jordi Gallego-Ayala, David Malmgren-Hansen, Thomas Pave Sohnesen, and Edmundo Murrugarra October 2022 2211 Tracking Global Social Protection Responses to Price Shocks (Living paper v.3) by Ugo Gentilini, Mohamed Almenfi, Hrishikesh TMM Iyengar, Yuko Okamura, Emilio Raul Urteaga, Giorgia Valleriani, and Sheraz Aziz September 2022 2210 Tracking Global Social Protection Responses to Price Shocks (Living paper v.2) by Ugo Gentilini, Mohamed Almenfi, Hrishikesh TMM Iyengar, Yuko Okamura, Emilio Raul Urteaga, Giorgia Valleriani, Jimmy Vulembera Muhindo, and Sheraz Aziz July 2022 2209 Tracking Social Protection Responses to Displacement in Ukraine and Other Countries by Ugo Gentilini, Mohamed Almenfi, Hrishikesh TMM Iyengar, Yuko Okamura, Emilio Raul Urteaga, Giorgia Valleriani, Jimmy Vulembera Muhindo, and Sheraz Aziz June 2022 2208 Tracking Global Social Protection Responses to Price Shocks (Version 1) by Ugo Gentilini, Mohamed Almenfi, Hrishikesh TMM Iyengar, Yuko Okamura, Emilio Raul Urteaga, Giorgia Valleriani, Jimmy Vulembera Muhindo and Sheraz Aziz April 2022 2207 Adapting Social Protection to FCV and Insecurity - The Case of the Democratic Republic of Congo by Silvia Fuselli, Mira Saidi, Afrah Alawi Al-Ahmadi April 2022 2206 Voluntary Savings Schemes to Protect Informal Workers in Jordan by Friederike Rother, Carole Chartouni, Javier Sanchez-Reaza, Ernesto Brodersohn, Montserrat Pallares- Miralles April 2022 2205 Enhancing Workers’ Protection in Jordan by Friederike Rother, Carole Chartouni, Javier Sanchez-Reaza, Gustavo Paez Salamanca, Belal Fallah April 2022 2204 Humanitarian and Social Protection Linkages with Examples from South Asia by Steen Lau Jorgensen and Maria Virginia Ceretti April 2022 2203 Cash in the City: The Case of Port-au-Prince by Olivia D’Aoust, Julius Gunneman, Karishma V. Patel and Caroline Tassot March 2022 2202 Tracing Labor Market Outcomes of Technical and Vocational Training Graduates in Saudi Arabia: A study on graduates from the Technical and Vocational Training Corporation (TVTC) by Nayib Rivera, Mehtabul Azam and Mohamed Ihsan Ajwad January 2022 2201 From Protracted Humanitarian Relief to State-led Social Safety Net System: Somalia Baxnaano Program by Afrah Al-Ahmadi and Giuseppe Zampaglione January 2022 2110 Early Lessons from Social Protection and Jobs Response to COVID-19 In Middle East and North Africa (MENA) Countries by Alex Kamurase and Emma Willenborg December 2021 2109 Migration in Bulgaria: Current Challenges and Opportunities by Daniel Garrote-Sanchez, Janis Kreuder, and Mauro Testaverde December 2021 2108 Social Protection and Labor: A Key Enabler for Climate Change Adaptation and Mitigation by Jamele Rigolini December 2021 2107 Intent to Implementation: Tracking India’s Social Protection Response to COVID-19 by Shrayana Bhattacharya and Sutirtha Sinha Roy June 2021 2106 Social Assistance Programs and Household Welfare in Eswatini by Dhushyanth Raju and Stephen D. Younger June 2021 2105 The Coal Transition: Mitigating Social and Labor Impacts by Wendy Cunningham and Achim Schmillen May 2021 2104 Social Protection at the Humanitarian-Development Nexus: Insights from Yemen by Yashodhan Ghorpade and Ali Ammar April 2021 2103 Review of the Evidence on Short-Term Education and Skills Training Programs for Out-of-School Youth with a Focus on the Use of Incentives by Marguerite Clarke, Meghna Sharma, and Pradyumna Bhattacharjee January 2021 2102 Welfare, Shocks, and Government Spending on Social Protection Programs in Lesotho by Joachim Boko, Dhushyanth Raju, and Stephen D. Younger January 2021 2101 Cash in the City: Emerging Lessons from Implementing Cash Transfers in Urban Africa by Ugo Gentilini, Saksham Khosla, and Mohamed Almenfi January 2021 2011 Building the Foundation for Accountability in Ethiopia by Laura Campbell, Fitsum Zewdu Mulugeta, Asmelash Haile Tsegay, and Brian Wampler January 2020 2010 Safety nets, health crises and natural disasters: Lessons from Sierra Leone by Judith Sandford, Sumati Rajput, Sarah Coll-Black, and Abu Kargbo December 2020 2009 A Reforma do Bolsa Família: Avaliação das propostas de reforma debatidas em 2019 by Matteo Morgandi, Liliana D. Sousa, Alison Farias, e Fabio Cereda November 2020 2008 The Role of Social Protection in Building, Protecting, and Deploying Human Capital in the East Asia and Pacific Region by Harry Edmund Moroz October 2020 2007 Boosting the Benefits of Cash Transfer Programs During the Early Years: A Case Study Review of Accompanying Measures by Laura Rawlings, Julieta Trias, and Emma Willenborg October 2020 2006 Expansion of Djibouti’s National Family Solidarity Program: Understanding Targeting Performance of the Updated Proxy Means Test Formula by Vibhuti Mendiratta, Amr Moubarak, Gabriel Lara Ibarra, John van Dyck, and Marco Santacroce August 2020 2005 Assessing the Targeting System in Georgia: Proposed Reform Options by Maddalena Honorati, Roberto Claudio Sormani, and Ludovico Carraro July 2020 2004 Jobs at risk in Turkey: Identifying the impact of COVID-19 by Sirma Demir Şeker, Efşan Nas Özen, and Ayşenur Acar Erdoğan July 2020 2003 Assessing the Vulnerability of Armenian Temporary Labor Migrants during the COVID-19 pandemic by Maddalena Honorati, Soonhwa Yi, and Thelma Choi July 2020 2002 Getting it Right: Strengthening Gender Outcomes in South Sudan by Samantha de Silva, Abir Hasan, Aissatou Ouedraogo, and Eliana Rubiano-Matulevich July 2020 2001 The Science of Adult Literacy by Michael S. C. Thomas, Victoria C. P. Knowland, Cathy Rogers January 2020 To view Social Protection & Jobs Discussion Papers published prior to 2020, please visit www.worldbank.org/sp. ABSTRACT Social safety net programs focus on a subset of the population, usually the poorest and most vulnerable. However, in most developing countries there is no administrative data on relative wealth of the population to support the selection process of the potential beneficiaries of the social safety net programs. Hence, selection into programs is often multi-methodological approached and starts with geographical targeting for the selection of program implementation areas. To facilitate this stage of the targeting process in São Tomé and Príncipe, this working paper develops High Resolution Satellite Imagery (HRSI) poverty maps, providing both estimates of poverty incidence and program eligibility at a highly detailed resolution (110 m x 110 m). Furthermore, the analysis combines poverty incidence and population density to enable the geographical targeting process. This working paper shows that HRSI poverty maps can be used as key operational tools to facilitate the decision- making process of the geographical targeting and efficiently identify entry points for rapidly expanding social safety net programs. Unlike HRSI poverty maps based on census data, poverty maps based on satellite data and machine learning can be updated frequently at low cost supporting more adaptive social protection programs. ABOUT THIS SERIES Social Protection & Jobs Discussion Papers are published to communicate the results of The World Bank’s work to the development community with the least possible delay. This paper therefore has not been prepared in accordance with the procedures appropriate for formally edited texts. For more information, please contact the Social Protection Advisory Service via e-mail: socialprotection@ worldbank.org or visit us on-line at www.worldbank.org/sp