Policy Research Working Paper 9858 Estimating Poverty in Kinshasa by Dealing with Sampling and Comparability Issues Yele Maweki Batana Takaaki Masaki Shohei Nakamura Mervy Ever Viboudoulou Vilpoux Poverty and Equity Global Practice November 2021 Policy Research Working Paper 9858 Abstract This paper proposes monetary poverty and inequality esti- has increased in the city by 12 percentage points, from 53 to mates for Kinshasa using a new Kinshasa household survey 65 percent, partly due to the loss of purchasing power fol- implemented in 2018. Given the obsolescence of the sam- lowing the sharp depreciation in 2017. Other explanatory pling frame, the survey was sampled using satellite imagery. factors include demographic factors, human capital, and However, the collection of data in the field was affected by spatial factors. The deterioration in well-being also appears sampling errors that are likely to compromise the represen- to have been exacerbated by the onset of the COVID-19 tativeness of the sample. After addressing these sampling pandemic through decline in labor and nonlabor income issues and dealing with some comparability issues with the and disruptions in goods and services markets and public 2012 survey, the paper shows that poverty and inequality services. increased significantly during 2012–18 in Kinshasa. Poverty This paper is a product of the Poverty and Equity Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at ybatana@worldbank.org or snakamura2@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Estimating Poverty in Kinshasa by Dealing with Sampling and Comparability Issues Yele Maweki Batana, World Bank 1 Takaaki Masaki, World Bank Shohei Nakamura, World Bank Mervy Ever Viboudoulou Vilpoux, World Bank Keywords: Poverty measurement; Inequality; Sampling errors; Comparability issues; Robustness analysis; Poverty maps; Propensity score; Urban; COVID-19 JEL Codes: C18, D31, I14, I24, I31, I32 1 Corresponding Author: ybatana@worldbank.org. We would like to thank Andrew Dabalen, Ruth Hill, Kirsten Hommann, Alexandra Jarotschkin and Clarence Tsimpo for useful comments. The authors also thank the INS (Institut National de la Statistique) of the Democratic Republic of Congo for access to data from the 2018 Kinshasa household survey. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations or those of the Executive Directors of the World Bank or the governments they represent. 1. Introduction The Democratic Republic of Congo (DRC) remains a post-conflict and fragile country where a long period of conflict and mismanagement have compromised its economic progress, despite the rich endowment of natural resources. In fact, the country’s history is marked by continued violence and insecurity as well as the absence of a functioning legal and regulatory system. This has generated high levels of displacement and humanitarian needs with over 1.9 million internally displaced persons. The country is still fragile and vulnerable to shocks. As a result, despite its rich endowment of natural resources, the country has seen its economic progress compromised, explaining why it is among the poorest countries in the world. Despite a decline, from 69.3 percent to 64 percent, in the proportion of the poor between 2005 and 2012, the number of poor people seems to have increased by 7 million during this period. Moreover, the country has the second largest number of extreme poor people in Sub-Saharan Africa, after Nigeria, accounting for about 14 percent of the poor in that region. If, like other African countries, the DRC is urbanizing rapidly, the speed of urbanization seems to vary depending on the economic region of the country. There are five economic regions: (a) the West region of the Kinshasa and Bas-Congo provinces; (b) the South region with the province of Katanga among others; (c) the Central region of the Kasai provinces, (d) the Congo Basin region with the Equateur, Bandundu, and Orientale provinces; and (e) the East region which includes Kivu provinces and Maniema (World Bank 2018a). The West region is highly urbanized with nearly 80 percent of the population living in urban areas, mostly in Kinshasa, with an average population growth of 4.8 percent per year since the last census in 1984 (Figure 1). The city of Kinshasa, whose population grew at an average rate of 5.1 percent per year between 1984 and 2010, is expected to become the largest megacity in Africa by 2030. Figure 1: Demographics by the five economic regions of the DRC Source: World Bank 2018. As a result of the high urban population growth, poverty has become an urban phenomenon, especially in Kinshasa. Without proper management of urbanization, the economic benefits expected from an agglomeration will not take place, especially when migration to cities is mostly induced by push factors (conflicts and risks) rather than pull factors (job opportunities). In fact, ineffective land management has pushed the DRC’s urban poor into unsuitable settlements, making them more vulnerable to economic shocks (World Bank 2018b). Urban poverty is high 2 and is rising in the country because of unplanned urbanization, while Kinshasa might be on its way to becoming the largest slum in Africa. According to the latest Poverty Assessment based on the 2012 household survey, the urban poverty rate is relatively close to that in rural areas, with 62.5 percent and 64.9 percent, respectively (World Bank 2016a). Unlike in other capital cities in Sub- Saharan Africa, more than half of the population (nearly 53 percent in 2012) of Kinshasa is poor, representing about 7 million people in 2017. 2 Even if poverty incidence reduced between 2005 and 2012, poverty in Kinshasa may have worsened in the late 2010s because of the deceleration of the economic expansion and the high urban population growth. What is the current poverty situation in Kinshasa and how has it evolved between 2012 and 2018? The Kinshasa Household Survey, implemented in 2018, will provide evidence to better understand the living standards in this megacity. In addition, the outbreak of the COVID-19 pandemic at the beginning of 2020 also raises questions about the possible adverse effects on poverty and the well-being of the inhabitants of Kinshasa. This paper aims to understand the magnitude and the manifestation of the income poverty, especially why poverty seems so high while cities are expected to provide more opportunities through agglomeration economies. Estimations are based on the Kinshasa household survey which is implemented in 2018. Unlike the previous national household survey, based on the old sampling frame where Kinshasa was considered as a stratum, the 2018 survey used a satellite imagery approach and considered seven strata. Beyond comparability issues that this can create, the data collection under the 2018 survey that may question the representativeness of households within the province. Issues include an incomplete enumeration within some of enumeration areas (EAs) and an overflow in the enumeration of some other EAs. On comparability issues, the bias will be corrected using the reweighting method suggested by Tarozzi (2007). To address the two sampling issues in the 2018 survey, some robustness and sensitivity tests will be undertaken to check, in each stratum, the significance of differences between subsample (including that with adequate enumeration), with the purpose to generate a representative sample. This paper contributes to enrich the literature on urban poverty especially when the situation is characterized by a lack of data and sampling issues. Findings show that poverty and inequality increased significantly during 2012–2018 in Kinshasa. Indeed, poverty has increased in the city by 12 percentage points, from 53 percent to 65 percent, partly due to the loss of purchasing power following the sharp depreciation from 2017. While demographic factors, human capital, and spatial factors appeared to be also important, the deterioration in well-being seems to have been exacerbated by the onset of the COVID-19 pandemic through decline in labor and non-labor income, disruptions in goods and services markets, and in public services. The rest of the paper is structured as follows. Section 2 presents the data and methodological approaches, including the description of the main issues and methods used to deal with that. Section 3 presents results in particular the trend, the distribution, the profile, and main determinants. Section 4 address the effect of the COVID-19 pandemic on the poverty while section 5 concludes the paper. 2 The numbers of the poor in 2017 are simulated based on the projected population from the National Statistics Institute (Institut National de la Statistique). 3 2. Data and methodology 2.1 Sampling and survey description To overcome the unavailability of an updated sampling frame, an innovative approach, based on the use of satellite images, was used. To carry out the survey, a methodological challenge arose concerning the sampling frame. The last population census for the DRC dates from 1984 and the date of the next census is uncertain. To cope with the lack of an up-to-date sampling frame of the city of Kinshasa, a method of collecting sociodemographic data adapted to the specificities of cities of developing countries was used. Comprehensive information on urban morphology (land use) constitutes a data source allowing rapid collection, by random sampling, of data relating to urban populations. This sampling method is based on the use of space remote sensing (high resolution satellite image) to provide a sampling frame. The morphological information of the urban environment provided by these images are used to stratify an area sampling plan applicable to sociodemographic surveys. To do this, very high-resolution images of the urban Kinshasa city were acquired (see Figure 2). It was then a question of counting the buildings and making a classification according to the type of buildings, collecting the geographical coordinates of the buildings as well as the geographical limits of the districts and the enumeration areas (EAs) contained in each district, and evaluating the land use in terms of area occupied by buildings and by blocks (groups of buildings). Figure 2: Kinshasa and Democratic Republic of Congo maps A - Satellite image of Kinshasa B – Democratic Republic of Congo map Source: Kinshasa Household Survey 2018, and One World - Nations Online. The Kinshasa Household Survey was carried out in 2018 according to a stratified, two-stage random sampling plan, with equal allocation in the first stage. In the first stage, the EAs were drawn, while a constant number of households was drawn from these EAs in the second stage. The objective of the sampling plan was to have a representative sample of households in the city of Kinshasa. Two criteria, the density (low density, medium density, and high density) 3 and the 3 The low-density area is defined by fewer than 10 buildings per hectare, the medium density area by the number of buildings per hectare ranging from 10 to 20, and the high-density area by more than 20 buildings per hectare. 4 typology (precarious and non-precarious neighborhood) 4 of the city, were used to stratify the urban environment of the city of Kinshasa. A total of seven strata were identified, of which six were identified on the basis of the two criteria mentioned earlier, while the seventh stratum is the rural commune of Maluku (see Table 1). A sample of 288 EAs was drawn in the first stage, including 246 in urban areas and 42 in rural areas. In the second stage, a sample of 9 households per EA/village was drawn, which gave a total sample of 2.592 households representative at the level of the strata and the whole province, including Maluku. However, for the purposes of this study, Maluku is excluded and only the six urban strata are selected, which represents 2,214 urban households. Table 1: Definition of strata using density and typology criteria Strata Definition 1 Precarious at low density 2 Precarious at medium density 3 Precarious at high density 4 Non-precarious at low density 5 Non-precarious at medium density 6 Non-precarious at high density 7 Rural Maluku The urban poverty survey is an integrated survey consisting of two nested surveys, covering different statistical populations: individuals and households. The first part of the survey covers, among other things, sociodemographic characteristics, education, health, living environment, employment, urban agriculture, and non-agricultural household entrepreneurship. The second part consists of carrying out a specific survey on household consumption. This part includes a component of the survey of prices in the markets that took place a few months after the main data collection. This note aims to estimate the indicators of households’ standard of living; analyze the determinants of poverty and inequalities; and study the strategies of households to fight against poverty, savings, and social capital. The main objective of the survey is to provide reliable data on the extent of poverty and inequality of households and populations in the city of Kinshasa to understand the dynamics of the living conditions of urban populations and guide all stakeholders in urban development. 2.2 Data collection issues While theoretically the methodological approach adopted for sampling may prove to be appropriate for conducting a household survey in Kinshasa, some problems have been identified in practice, particularly in the collection of data in the field. This issue is likely to lead to sampling errors, which can call into question the representativeness of the surveyed sample. The first problem relates to an incomplete enumeration within some of the 288 EAs retained in the sample. In total, 92 EAs are affected by this issue (see Table 2). The second problem is the overflow in the 4 In contrast to non-precarious neighborhoods, precarious neighborhoods are neighborhoods whose dwellings are built by the occupants of housing on land acquired in undeveloped areas, agricultural and market gardening land, and uneven land (exposure to erosion and floods). The constructions are anarchic with the frequent use of recycled materials (cardboard, plastics, and sheet metal). Most often, they are located in parts of the city neglected by the more affluent categories, such as on steep slopes and near industrial areas, which makes them all the more dangerous and where misery is concentrated. Community facilities (water, electricity, sanitation, and transport) are reduced and availability is low. 5 enumeration of some EAs. In certain cases, the EAs are fully enumerated, but the enumeration goes beyond the EAs to include totally or partially neighboring EAs not selected in the initial sampling. In other cases, even if the enumeration is incomplete in the selected EA, it goes beyond to cover the neighboring EAs. This may call into question the representativeness of households in the selected EA to which it would not belong. The combination of these two problems leads to consideration of four scenarios: (a) the case where the enumeration is adequate (completely enumerated EAs), (b) the case where the enumeration is incomplete without overflowing, (c) the case where the enumeration is complete but has overflowed to include the neighboring EAs, and (d) the last case with incomplete but overflowing enumeration. It is worth noting that, in all cases, households are picked completely from the enumerated parts of EAs and in both original and overflow EAs when applicable. Table 2 shows how the EAs are distributed according to these various scenarios and the six urban strata. Table 2: Distribution of the various sampling scenarios by strata Scenarios Number of EAs (islets) Stratum Stratum Stratum Stratum Stratum Stratum Total 1 2 3 4 5 6 No issue: Adequate enumeration 5 15 15 12 20 25 92 1st issue: Incomplete enumeration 17 17 18 20 13 7 92 Enumeration is completed in the selected EA and 8 6 6 3 4 6 33 enumeration Overflowing goes beyond to include 2nd issue: neighboring EAs Enumeration is incomplete in the selected EA but goes 10 6 3 5 4 1 29 beyond to include neighboring EAs Total 40 44 42 40 41 39 246 Source: Kinshasa Household Survey 2018. 2.2.1 Adequate enumeration In this case, the EAs were fully Figure 3: Case of adequate enumeration enumerated while the nine households were selected from each of them as expected (see Figure 3). In total, 92 of the 288 EAs belong to this category and were adequately enumerated. We would not have had a sampling problem if all the EAs were enumerated correctly. However, there were disparities between strata since stratum 6 was the best covered with nearly two- thirds of the EAs adequately enumerated, while stratum 1 remained the least well treated with Source: Kinshasa Household Survey 2018. 6 only 12.5 percent of the EAs perfectly enumerated. 2.2.2 Incomplete enumeration In this second scenario, the EAs (92) are Figure 4: Case of partial enumeration partially covered (see Figure 4). To make the corrections, the surface of each EA covered by the enumeration is delimited and the ratio of area covered (rate of coverage) is estimated by making abstraction of the empty spaces. Assuming that the distribution of households within the EA is homogeneous, this rate, considered as a correction factor, is applied to the number of households identified at the level of the whole EA. Due to its simplicity, this approach was preferred to that of redrawing the boundaries and creating new EAs. Source: Kinshasa Household Survey 2018. 2.2.3 Overflowing enumeration In many cases, surveyors went beyond Figure 5: Case of overflowing enumeration the EAs assigned to them during the enumeration (see Figure 5). Sometimes, after fully covering an EA, the enumeration was extended to other EAs not selected in the sampling, with 33 selected EAs affected by this scenario. When other wrongly covered EAs were fully covered, it was then relatively straightforward to merge them with the initial EA to make it a single larger EA. This resizing of the initial EA was possible with 24 of the 33 EAs. Source: Kinshasa Household Survey 2018. The remaining cases are made up of the EAs that have not been completely enumerated but have been extended to other unselected and also incompletely enumerated EAs. In this latter scenario, the resizing of the initial EA previously considered is more complicated and leads to these EAs being excluded. It is however worth noting that one-third of the EAs in this scenario come only from the first stratum, likely to be the poorest, which may bias poverty downward. 7 2.3 Consistency and sensitivity analysis To constitute the representative sample for the estimation of poverty in Kinshasa, 208 EAs were selected from the 246 EAs surveyed in the urban part of Kinshasa. To analyze whether the different sampling problems encountered could generate statistical biases, tests of differences in means and proportions were carried out. The first tests make it possible to compare, within each stratum, subsamples with completely enumerated EAs and that where the EAs were incompletely enumerated (see Table 3). Overall, with the exception of some structural differences at the level of per capita consumption in strata 1 and 6, at the level of household size for strata 5 and 6, and at the level of the sex of the head of household for stratum 4, there is no marked structural difference between the two samples. The two samples can thus be merged, while the sampling weights are calculated as described in Annex A. In addition, 24 EAs that were extended to other EAs not selected in the initial sample were fully counted along with the additional EAs. These EAs were resized by integrating these additional EAs to form 24 larger EAs and were then added to the other 184 previous EAs to constitute a total of 208 EAs. Table 3: Means or proportions difference tests between the sample enumerated adequately and the sample of incomplete enumeration T-Statistics Variables (Probability) Stratum 1 Stratum 2 Stratum 3 Stratum 4 Stratum 5 Stratum 6 a 4.27*** 1.27 0.25 1.64 1.61 2.95*** Consumption per adult equivalent (0.00) (0.21) (0.80) (0.10) (0.11) (0.00) –0.67 –1.27 0.85 0.18 1.38 0.45 Age of the household head a (0.50) (0.21) (0.39) (0.39) (0.17) (0.65) –1.46 –1.64 0.52 0.07 –2.83*** –2.09** Household size a (0.15) (0.10) (0.60) (0.97) (0.00) (0.04) b –1.32 0.27 0.31 2.12** –0.52 1.64 Household head is female (0.75) (0.79) (0.76) (0.03) (0.60) (0.10) 1.10 1.40 0.20 –0.04 –0.54 0.29 Household head is literate b (0.27) (0.16) (0.84) (0.97) (0.59) (0.77) Source: Kinshasa Household Survey 2018. Note: (*), (**), and (***) mean that the null hypothesis of equality of means or proportions can be rejected at the significance level of 10, 5, and 1 percent, respectively; (a) and (b) denote means difference tests and proportions difference tests, respectively. The analysis will therefore be based on the representative urban sample of 1,872 households belonging to the 208 EAs previously selected. The other EAs extended but whose surfaces are not well circumscribed to allow resizing (the 29 EAs in scenario 4 and the 9 remaining in scenario 3) were excluded from the sample. Table 4 performs comparison tests of means and proportions between this subsample and the overall sample of the 208 EAs considered by the study. The results show that, even if there are no observable significant structural differences for most sociodemographic variables, this is not the case for consumption per adult equivalent where the differences in means are significant in four out of the six strata. This supports the idea of excluding the subsample from the study sample, since it may bias household well-being when the poor enumeration in these areas is caused by their sociodemographic characteristics. However, if not, this also may suggest a bias in the retained sample by excluding an informative subsample. In total, the analysis will potentially relate to 1,872 urban households in Kinshasa, representing about 85 percent of the initial sample. 8 Table 4: Means or proportions difference tests between the sample enumerated adequately or incompletely and the sample with overflowing enumeration 5 T-Statistics Variables (Probability) Stratum 1 Stratum 2 Stratum 3 Stratum 4 Stratum 5 Stratum 6 Consumption per adult equivalent 2.37** –1.88* 1.41 1.47 –1.71* –4.90*** a (0.02) (0.06) (0.16) (0.14) (0.09) (0.00) 1.35 1.35 -1.10 0.30 -0.87 0.52 Age of the household head a (0.18) (0.18) (0.27) (0.76) (0.39) (0.60) 2.43** 0.72 –1.12 0.96 0.47 0.02 Household size a (0.02) (0.47) (0.26) (0.34) (0.64) (0.98) 1.10 0.26 0.93 –0.36 0.11 2.17** Household head is female b (0.27) (0.80) (0.35) (0.72) (0.91) (0.03) –1.12 –0.52 –1.24 0.53 1.56 1.20 Household head is literate b (0.26) (0.61) (0.29) (0.59) (0.12) (0.23) Source: Kinshasa Household Survey 2018. Note: (*), (**) and (***) mean that the null hypothesis of equality of means or proportions can be rejected at the significance level of 10, 5, and 1 percent, respectively; (a) and (b) denote means difference tests and proportions difference tests, respectively. 2.4 How comparable are the 2018 survey and the Kinshasa subsample of the 2012 survey? The poverty estimate could be more precise in 2018 due to the stratification adopted in the Kinshasa sample. Even if the same methodology for collecting consumption data was used for the two surveys that were carried out elsewhere during the same period of the year, there are nevertheless some points of divergence. First, the 2010 administrative census was used as the frame for the 2012 survey, while the satellite imagery approach was used for the 2018 survey. Second, the stratification method used is not the same for each survey. If, for the 2012 survey, Kinshasa was considered as a stratum where 54 neighborhoods were drawn in the first degree, in the case of the 2018 survey, six strata were considered for the urban part of Kinshasa based on the density and precariousness level. Unlike the previous survey, taking into account the stratification of slums, where urban poverty remains concentrated, allows for a better understanding of the level of poverty. Indeed, if the gain in precision, allowed by the current stratification, allows by definition to gain in precision, there are greater advantages in the analysis of poverty when this stratification is based on precariousness and density, since the two criteria (precariousness and density) make it possible to differentiate several large homogeneous areas in Kinshasa. By classifying the 54 districts of Kinshasa into precarious and non-precarious zones and using the sampling weights, we observe that the proportion of households living in non-precarious zones was one-third in 2012, which is about twice higher than their proportion in the 2018 survey. Part of the difference in poverty between the two periods could therefore come from an underestimation of poverty in the 2012 sample following the absence of stratification. The question on the frequency of purchasing food items and the number of items are also points of divergence between the two surveys. There are several other points of divergence which, without necessarily calling into question the comparison between the two surveys, can conduct to qualify the results in terms of the magnitude of the difference in poverty. The first one relates to the frequency of acquisitions of food items. In 2012, the questionnaire had just one question 5 The 24 overflowed EAs that were resized are excluded from the other overflowing enumeration sample and added to the sample enumerated adequately or incompletely. 9 capturing the frequency of purchases, while in 2018, instead of one question, the questionnaire considered two different questions, the first to capture the number of times the household buy a good, and the second question to capture the period of time considered (daily, weekly, monthly, etc.). The second point of divergence is about the number of consumption items which increased in 2018 with a total of 653 categories included in the questionnaire against only 557 categories in 2012. The results obtained from the final sample of the 2018 survey are compared, in the next section, to those of the Kinshasa subsample of the 2012 survey. The methodology for aggregating consumption data followed the 2012 methodology overall, and the 2018 poverty line is determined by adjusting prices and making consumption data comparable. Details are presented in Annex B. However, only 828 households in the sample come from correctly listed EAs. The fact that the other subsample includes 1,044 households coming from the EAs that are not completely listed or, in certain cases, from resized EAs can generate selection biases. These biases are based on the fact that the probability of belonging to an EA that is not fully listed may depend on certain sociodemographic characteristics of the households. Although the estimates relate to the overall sample without considering these biases, a robustness analysis is proposed to check whether these selection biases could significantly change the poverty results. The bias is corrected by using the reweighting approach (inverse of the propensity score) suggested by Tarozzi (2007) to make the households in the biased subsample comparable to those in the unbiased subsample. See Annex C for more details on this approach and the results. 3. Monetary poverty in Kinshasa 3.1 Poverty trends The poverty incidence in Kinshasa is one of the highest in the capital cities of Africa. Urban poverty in the DRC is exceptionally high when compared to other countries of Sub-Saharan Africa. For example, compared to other large African cities, Kinshasa has a relatively very high poverty rate and at least twice that of other cities. Figure 6 compares the poverty incidence assessed at the national poverty line, of selected cities in Sub-Saharan Africa between 2010 and 2017. The poverty incidence in Kinshasa was estimated at around 53 percent while those of other cities are found, in most cases, to be below 30 percent. Poverty seems to have increased in Kinshasa between 2012 and 2018, mostly because of loss of purchasing power due to inflation. 6 In fact, according to the Part A of Figure 7, poverty incidence has increased in the city by 12 percentage points, with an estimated rate of 64.8 percent in 2018. The same trends are recorded in terms of the poverty gap and the severity of poverty, indicating that the situation of the poorest and most vulnerable populations has also deteriorated considerably. This increase in poverty may be explained mainly by the sharp drop in purchasing power from 6 Despite the dollarization of the country and informal economy, incomes of the poor seemed to have increased less than the price level. In fact, poor people do not handle a lot dollar bills, which are used in transactions in large denominations preferably, in new condition and without alteration. Moreover, poor people in Kinshasa miss opportunities to benefit from price increases through increases in wages and non-labor income. It should also be noted that there are other causes of the increase in poverty including the rural-urban migration which will be discussed later in the paper. 10 Figure 6: Poverty comparison at national poverty line in selected African cities, 2010–2017 7 Sources: World Bank’s Poverty Assessments for various countries. Figure 7: Loss of purchasing power and increase in poverty in Kinshasa, 2012–2018 Sources: DRC Household Survey 2012, Kinshasa Household Survey 2018, INS, and Banque Centrale. The robustness analysis confirms the decline in well-being for the entire population of Kinshasa, especially among the poorest 85 percent between 2012 and 2018. Figure 8 shows poverty curves for 2012 and 2018. For over 85 percent of households whose annual consumption per capita is 7 Even if poverty comparisons are more suitable when national poverty lines are beforehand converted into Purchasing Power Parity (PPPs), current comparisons based on non-converted national lines are just indicative and allow, given the magnitude of the difference, to assume the big difference between Kinshasa and the other cities. 11 lower than CDF 2.6 million, well-being declined during the period as shown by the 2018 poverty curve which is above that of 2012. The situation also seems to be much more unfavorable for the city's poorest people. In addition to inflation, another factor that may explain the increase in poverty is rural-urban migration. 8 Indeed, this kind of migration is mostly triggered by push factors (conflicts and risks) rather than pull factors (economic opportunities) (World Bank 2016b), and may therefore simply translate into a transfer of poverty from rural to urban areas. Figure 8: Poverty curves for 2012 and 2018, valued at 2018 prices Sources: DRC Household Survey 2012 and Kinshasa Household Survey 2018. 3.2 Distribution and inequality analysis Income inequality is high in Kinshasa, but it increased sharply between 2012 and 2018 at the same time as the sharp increase in poverty. Reducing poverty depends not only on economic growth but also on shared prosperity, meaning that part of this growth is allocated to the growth of the poor. The evolution of inequality may therefore itself determine that of poverty. When growth redistribution is pro-poor, it will lead to lower inequality in the country and, therefore, poverty reduction. The Gini index for Kinshasa, which is a common measure of inequality, increased between 2012 and 2018, with a value passing from 32.4 to 40.2 percent during this period. A more comprehensive tool for measuring inequality is the Lorenz curve. Figure 9 shows the Lorenz curves for 2012 and 2018. These curves describe the evolution of the cumulative proportion of income (proxied by consumption) of households or individuals from a population ranked in ascending order of consumption. When the Lorenz curve coincides with the diagonal, then there is no inequality between individuals. Inequalities are present when the curve is below the diagonal and are as high as the differences between this curve and the diagonal. The Lorenz curves for 2012 is above that for 2018, which confirms that inequality significantly increased during this six-year period. In addition to the fact that rising inequality may hamper the city’s economic development and poverty reduction efforts, it may also create favorable conditions for social unrest and conflicts. 8 Due to the lack of proper migration questions, data from the 2018 Kinshasa Household Survey do not allow to test this hypothesis. 12 Figure 9: Lorenz curves of Kinshasa based on per capita consumption in 2012 and 2018 The consumption growth seems to have been negative for households, which is consistent with trends in poverty and inequality. Figure 10 shows the curves of per capita consumption growth among households in Kinshasa for 2012–2018. As shown in Figure 10, except the 10 percent of richest people, the city’s population experienced a decline in their consumption during the period. However, as noted earlier, the situation seems to have been more dire for the poorest, as illustrated by the upward slope of the median spline and the curve for the mean growth rate for the poorest ‘p’ percent. Figure 10: Growth incidence curve in Kinshasa, 2012–2018 Sources: DRC Household Survey 2012 and Kinshasa Household Survey 2018. 3.3 Poverty profile in 2018 Demographics seem to play an important role in the poverty status of households in Kinshasa. In fact, Part A of Figure 11 shows that poverty tends to be high when the household size is large. The poverty incidence ranges from 10 percent in single person households to almost 90 percent in households of 10 or more people. The same trend is observed in the poverty gap which goes from 1.7 percent to almost 48 percent. This result is observed even if household consumption has been 13 evaluated in terms of adult equivalent instead of simply consumption per capita. 9 As shown in Part D of Figure 11, the presence of children and the elderly, regardless of their gender, tends to increase poverty. Indeed, a high dependency 10 ratio means that a large part of the household is made up of children or individuals who are not usually involved in the income-generation process within households. Figure 11: Poverty profile by demographics in Kinshasa, 2018 Source: Kinshasa Household Survey 2018. The household head’s age seems to play a role, while households headed by women appear, on average, poorer than those headed by men. As shown in Part B of Figure 11, the incidence of poverty in female-headed households is around 72 percent, which is almost 10 percentage points higher than in households headed by men. The trend is similar at the level of the poverty gap with a rate of 34.3 percent in the first household group against 25.8 percent in the second group. This result is obtained although the size of these households is slightly smaller than that of the second households, which therefore reflects the existence of real gender inequality in Kinshasa households. Regarding the household head’s age, results suggest that poverty is lower, on average, among the younger age groups, with a poverty incidence varying from 45.5 percent in households in the age group 25 to 29 to more than 75 percent among households whose heads are in the age group 50 to 54 (see Part C of Figure 11). This can be explained in large part by the fact that households headed by men ages 45 and over are, on average, larger. 9 The per capita approach usually divides the total consumption of a household by its total number of members regardless of their age and gender and ignores the economies of scales inherent in family size. By doing this, the positive correlation between household size and poverty may be quite trivial due to the measure of well-being at the household level, which is not the case with the equivalence scale approach. 10 The dependency ratio is measured as the ratio between the population not in the labor force (children ages 0 to 14 and adults over 65) and the population in the labor force (individuals ages 15 to 64). 14 Education seems to offer real opportunities only for the university level, while public administration concentrates relatively on the less poor. Two-thirds of household heads in Kinshasa have completed secondary education or have a university level education. When compared to those who have not reached and completed secondary school, the results show that households whose head has completed secondary education are, on average, slightly poorer. But the difference is only significant for households where the head has a university level education with a 39 percent poverty incidence, which is at least 30 percentage points lower than that of other households (Part B of Figure 12). The majority of public administration workers come from this group of household heads with a university level education, although it is also estimated that more than 10 percent of them are unemployed. According to Part A of Figure 12, the least poor households are in fact those whose heads work in public administration, with a poverty incidence of around 40 percent. The other relatively better sectors are industry (56 percent), commerce (59 percent), and catering/transport/communication (61 percent). On the other hand, the incidence of poverty remains high compared to the city average for households headed by the unemployed (74 percent) and for those whose head is in construction (72 percent), education and health (67 percent), or other sectors (72 percent). Figure 12: Poverty profile by location and opportunities in Kinshasa, 2018 Source: Kinshasa Household Survey 2018. The poorest populations naturally tend to live in strata defined as precarious and relatively far from the core city. The results of the analysis of monetary poverty seem to be consistent with the definition of the strata. Indeed, the three strata defined as precarious have the highest poverty incidences, varying from 64 percent for precarious at medium density stratum to 70 percent for precarious at low density stratum (see Part C of Figure 12). On the other hand, the incidence of poverty is relatively low in the non-precarious strata with 54 percent in the low-density stratum, 40 percent in the medium density stratum, and 35 percent in the high-density stratum. These results are also consistent with those relating to the distance to the core city. Indeed, while households 15 located relatively closer to the core city tend to live in non-precarious strata, the opposite is true for households living far from the core city center, which are found, for the most part, in precarious strata. As shown in Part D of Figure 12, the probability of households being poor increases as they live far from the core city. For example, the incidence of poverty is estimated at 69 percent among households living more than 15 km from the core city, while it is only 18 percent among households living within 5 km. Indeed, living not far from the core city offers more income- generation opportunities, particularly easier access to infrastructure and public utility services. See World Bank (2020) for more details. Figure 13: Poverty maps 11 for Kinshasa, 2018 A - Commune-level estimates B - Quartier-level estimates Sources: Kinshasa Household Survey 2018, remote-sensing data and satellite imagery. 11 Given that the 2018 Kinshasa household survey is typically too small to produce reliable estimates for areas smaller than strata, and in the absence of an updated population census, remote-sensing data and satellite imagery provide useful information to predict poverty estimates for quartiers and communes. However, this should be used with caution and complemented by other relevant maps since the approach used is not the best one to consensually derive poverty maps. 16 3.4 Why are poverty and inequality high and rising in Kinshasa? Beyond inflation, several socioeconomic and demographic factors explain the high poverty recorded in 2018. As mentioned earlier, inflation following the depreciation of the Congolese franc from 2017 is largely the cause of the decline in purchasing power of Kinshasa households between 2012 and 2018. This seems to have not only explained the increase in poverty but also the general decline in the well-being of the population during the period. But, beyond the rise in prices, several other factors explain the high poverty recorded in the city in 2018. A multivariate analysis, based on the regression of the logarithm of consumption per adult equivalent of households on a certain number of socioeconomic and demographic variables, shows that the demographic (household size, dependency ratio, and sex of the head of the household); socioeconomic (education and sector of activity); and spatial (distance to the core city) factors are determining factors in explaining the level of household welfare (see the table in Annex E). Indeed, the important role played by demographic variables on household consumption, and therefore on poverty, is confirmed by multivariate analysis. The size of the household is significantly and negatively correlated with the level of consumption, which indicates that large households have, on average, a relatively low level of well-being. This result is supported by the dependency ratio, which also has a negative and significant effect on household well-being. This raises the question of fertility, the rate of which was estimated in 2018 at 6.2 at the national level, nearly 5 in urban areas, and 3.6 in Kinshasa, according to data from the last Multiple Indicators Cluster Survey (MICS) of 2017–2018. This question remains important with regard to the role of women from the perspective of a demographic transition, especially since the analysis shows the existence of gender inequalities in the city. In fact, households headed by women have, on average, a lower level of well-being than those headed by men. The level of education of the household head seems to be a determining factor in the well-being of households. Unlike the bivariate analysis where only the university level education seemed decisive, the multivariate analysis shows that education plays an important role in the well-being of households even from the primary level. Indeed, households headed by individuals with a primary level education or higher tend to have much greater well-being than households whose heads have no education. However, as expected, the benefit is much greater when heads have a university level education. With some exceptions, there does not seem to be much difference in terms of well-being between households according to the sector of activity of the head of household. Compared to households whose heads work in the construction industry, households whose heads work in the restaurant, transport and communication industry, and in public administration, seem to be less poor. On the other hand, there is not difference with households whose heads are engaged in manufacturing and other service industries. Even the negative sign expected and observed for households whose heads work in agriculture or are unemployed is not significant. Moreover, compared to households which earned income from a single source, households are likely to have better well-being when sources of income are diversified. The distance to the core city matters also for poverty and well-being. Indeed, the regression results show that, compared to households located within 5 km from the core city, being located further is negatively and significantly correlated with well-being. Moreover, the correlation is higher for 17 households living more than 15 km away from the core city. It is one of the channels through which the rural-urban migration may lead to impoverishment in Kinshasa. As mentioned in World Bank (2018), the urban expansion of the city took place at the expense of living conditions (World Bank 2018a). According to the same World Bank Report, Kinshasa’s areas expanded by 30 percent between 2004 and 2015, growing from 363 km2 to 472 km2. If we assume the same growth rate after 2015, the area of the city may be estimated at around 507 km2. Moreover, these new inhabited spaces, which are located far from the core city and mostly along erosion-prone areas and unserviced outskirts, are the main destination for migrants. Demographics, human capital, and geographical location are determinants of inequality in Kinshasa. Poverty reduction can only be effective if it is combined with shared prosperity. This is true in Kinshasa where inequality is very high, with a Gini index of 40 percent in 2018. A regression-based decomposition of the inequality is done using the Field method (Fields 2003) and presented in Figure 14. The most important factors in explaining the inequality in Kinshasa in 2018 are household size (39 percent), university level education of the household head (31 percent), distance of more than 15 km to the core city (15 percent), and household head working in public administration (14 percent). In a lesser extent, having labor and non-labor income sources contributed almost 7 percent to inequality. Beyond its own contribution to inequality, the distance to the core city can contribute more through other factors. For example, households living more than 15 km tend to be larger than those living less than 5 km, with an average size of 7.4 members against 6 members. Furthermore, only 14 percent of household heads in these remote areas have a university level education compared to 41 percent in nearby areas. Source Kinshasa Household Survey 2018. 4. Poverty and the COVID-19 pandemic in Kinshasa With more than 250 cases of COVID-19 officially registered in the DRC in mid-April 2020, mostly registered in Kinshasa, and a total of 20 deaths, the pandemic seems to be growing in the city. 18 Beyond the overflows, which are feared in the short term at the level of health structures, several adverse socioeconomic impacts are also expected. The poverty and distributional impacts of the pandemic in Kinshasa will take place through four main channels: (a) decline in labor income, (b) decline in non-labor income, (c) disruptions in goods and services markets, and (d) disruptions in public services. 4.1 Impact on labor income An adverse effect on labor income is expected due to the contraction of the aggregate demand. In fact, job losses and slowdown of economic activity in the transport, tourism (hotel and catering), and retail commerce industries will cause the decline in labor income. World Bank (2020) shows that 56 percent of the city’s working-age population is employed, mainly in the service sector (80 percent), while about 7 percent is unemployed (World Bank 2020). Among the employed individuals, even if those working in the hotel and tourism industry are relatively very low (less than 1 percent), almost 42 percent work in commerce and 6 percent work in transport and communication. The adverse effects of the COVID-19 pandemic should be felt much more in these three industries, which account for almost half of the city's workers. Moreover, the containment of the commune of Gombe (the city’s core area) by April 6, 2020, 12 will worsen the situation especially since jobs of relatively good quality are concentrated in Gombe and its surrounding areas. The situation risks further disadvantaging women and young people, which will accentuate the inequalities compared to these two vulnerable groups. Already, underemployment appeared to be prevalent among the youth in 2018, while the female working-age population with higher education faced a high rate of unemployment. 4.2 Impact on non-labor income Non-labor income will significantly decrease in Kinshasa mainly because of a drop in remittances. Considered to be among the most important diaspora from the African continent, with about 7 million migrants, personal remittances from the Congolese diaspora seem to have significantly increased from the early 2010s. It is worth noting that the volume of remittances received from Congolese migrants is by far underestimated, since most remittances are sent through informal channels due to the inadequacies of the banking sector (Sumata and Cohen 2018). However, the trends in Figure 15 should reflect that of actuals transfers received by the country. According to Figure 15, if remittances received seem relatively negligible during the second half of the 2000s, there has been a significant increase from 2011. The very marginal share of gross domestic product (GDP) (by far less than 1 percent) jumped to over 4 percent in 2011, before experiencing a relative decline until the mid-2010s. There has been a revival in the level of transfers received, which approached that of Official Development Assistance (ODA) in 2018. This trend, which continued in 2019, even if it was at a slower pace, 13 with a share of GDP at around 4 percent, will be compromised by the outbreak of the COVID-19 pandemic in early 2020, which is expected to negatively affect the economies of countries where migrants are employed. 12 The authorities of the city-province of Kinshasa have decided to start, since April 6, 2020, the confinement of the commune of Gombe, considered to be the epicenter of the COVID-19 pandemic. 13 https://blogs.worldbank.org/peoplemove/data-release-remittances-low-and-middle-income-countries-track-reach- 551-billion-2019. 19 Figure 15: Evolution of personal remittances and ODA in the DRC, 2005–2018 Source: World Development Indicators, 2005–2018. Falling remittances following the COVID-19 pandemic could increase poverty and inequality in Kinshasa. According to the data from the study, in 2018, around 35 percent of households received remittances for consumption needs. International remittances are much more concentrated, in terms of amounts received, in high density and medium density non-precarious areas. Internal remittances are also high, on average, in all non-precarious areas and to a lesser extent in high- density precarious areas, while the amounts received are relatively low in other precarious areas (low or medium density). If these remittances were removed, the incidence of poverty in the city could increase by 1.4 percentage points (see Part A in Figure 16). However, the effect on redistribution seems to be relatively stronger, notably with a Gini index which could increase by 1.7 percentage points. This is confirmed by other poverty indicators which are the depth of poverty and the severity of poverty. The relatively higher increase in these two indicators, which take into account the situation of poor households, shows that the removal of these remittances could exacerbate the vulnerability of the people of Kinshasa. Part B of Figure 16 shows that the effect on the increase in poverty is not the same according to the strata. Non-precarious areas with high and medium density should be the most affected with respective increases of around 4 percentage points and 2 percentage points. These areas are also the areas that receive the highest amounts of remittances per household. 20 Figure 16: Increase in poverty and inequality in case of remittances removal Source: Kinshasa Household Survey 2018. 4.3 Disruptions in goods and services markets Figure 17: Inflation in Kinshasa Source: Institut National de la Statistique (INS), DRC. 21 The outbreak of the COVID-19 pandemic has caused inflation in Kinshasa. Figure 17 shows a drastic rise in prices during the last week of March, especially for food products. In fact, for foods, cumulative inflation, since the start of January 2020, has gone from 0.8 percent during most of the first quarter to 3.1 percent in just three weeks. The same trends are observed, but to a lesser extent, in the case of health and goods and services. Cumulative food inflation rose from 1.2 percent to 1.9 percent, and from 1.0 percent to 2.3 percent for all goods and services. Although inflation remains relatively lower in transport, there is, however, a gradual and monotonous increase. Part B of Figure 16 shows that such strong inflations during the same period of the year were observed in 2012, and more exceptionally during 2017–2018 mainly because of the drastic depreciation of the Congolese franc experienced during this period. Regarding the general price index, annual inflation was 50 percent in 2017, 7 percent in 2018, and 4 percent in 2019. The outbreak of the COVID-19 pandemic therefore seems to reverse this trend already with this high inflation recorded in 2020. If the pandemic continues, inflation will rise even more, although with the trend already underway, there is no reason to believe that it will reach the level of 2017. Figure 18: Simulated increase in poverty by inflation rate and strata Source: Kinshasa Household Survey 2018. 22 4.4 Disruptions in public services Disruptions in basic services provisions mainly concern health and education sectors. The country has a fragile health system already undermined by several recent health issues including Ebola. Given the weak capacity of the country to finance social spending programs, the new pandemic will jeopardize the situation in social sectors. If the pandemic is not quickly contained in the country, and especially in Kinshasa, the situation could lead to a saturation of the health system with inadequate care for other diseases. The education system, which is also already weak, could deteriorate as the measures adopted by the authorities, including suspension of classes in schools, may affect student retention and learning and compromise, in the short term, the free primary school program. 14 is just under 3 percent (see Part A of Figure 19). More than half of households (51 percent) have spent more than 5 percent of their consumption budget on education needs, while 32 percent have spent more than 10 percent of this budget. There are, however, some disparities at the level of the strata. Precarious areas with low density and those with medium density make up for the catastrophic expenditure on health by the proportions of households which are 4.6 percent and 3.6 percent, respectively. Figure 19: Household health and education expenses Source: Kinshasa Household Survey 2018. 14 Health expenditure is defined as catastrophic when its share in the total consumption budget exceeds 10 percent. 23 5. Conclusion Estimation of the level of poverty in a megacity such as Kinshasa, which is growing rapidly, faces some challenges, especially since the data are poor while the national sampling frame is outdated. To overcome the unavailability of an updated sampling frame, a Kinshasa household survey was carried out in 2018 using a geospatial sampling approach. Based on this recent survey, this paper provides poverty estimates for Kinshasa after dealing with sampling and comparability issues. Main findings show that poverty and inequality increased significantly during 2012–2018 in Kinshasa. In fact, poverty has increased in the city by 12 percentage points, from 53 to 65 percent, while the inequality index increased from 32.4 to 40.2 percent during this period. The deterioration in living conditions is partly due to the loss of purchasing power following the sharp depreciation from 2017. Other explanatory factors include demographic factors, human capital, and spatial factors. The deterioration in well-being also appears to have been exacerbated by the onset of the COVID-19 pandemic through decline in labor and non-labor income, disruptions in goods and services markets, and in public services. These results provide avenues for thinking about strategies to fight against urban poverty relying on spatial features, the development of human capital and better support for demographic growth. Moreover, there is a need to relieve poorest and most vulnerable Kinshasa households by mitigating the adverse effects of the COVID-19, and for the country to be part of the global dynamic in the fight against this pandemic for the resumption of activities. The reflection should be extended to the objective of reducing poverty at the national level but remains challenged by the lack of data. In fact, it is worth noting that data are still lacking to make a current and more accurate estimate of national poverty in the whole DRC. An immediate need, which is also the first step in terms of filling this data gap, is to update the national sampling frame. The geospatial sampling method used under the Kinshasa survey could be extended to the whole country. 24 References Battese, G. E., R. M. Harter, and W. A. Fuller. "An error-components model for prediction of county crop areas using survey and satellite data." Journal of the American Statistical Association 83.401 (1988): 28-36. Das, S. and S. Haslett. 2019. “A comparison of methods for poverty estimation in developing countries.” International Statistical Review 87(2): 368-392. Deaton, A., and S. Zaidi. 2002. “Guidelines for Constructing Consumption Aggregates for Welfare Analysis.” Living Standards Measurement Study Working Paper 135, World Bank, Washington, DC. Elbers, C., J. O. Lanjouw, and P. Lanjouw. "Micro–level estimation of poverty and inequality." Econometrica 71.1 (2003): 355-364. Engstrom, R., J. Hersh, and D. Newhouse. "Poverty from space: Using high-resolution satellite imagery for estimating economic well-being." (2017). Engstrom, R., D. Newhouse, and V. Soundararajan. Estimating Small Area Population Density Using Survey Data and Satellite Imagery: An Application to Sri Lanka. The World Bank, 2019. Fields, G. 2003. “Accounting for Income Inequality and Its Changes: A New Method with Application to the Distribution of Earnings in the United States.” Research in Labor Economics 22: 1–38. González-Manteiga, W., et al. "Bootstrap mean squared error of a small-area EBLUP." Journal of Statistical Computation and Simulation 78.5 (2008): 443-462. Head, A., et al. "Can human development be measured with satellite imagery?." ICTD. 2017. IMF (International Monetary Fund). 2019. Democratic Republic of the Congo, IMF Country Report No 19/285, September. Washington, DC: IMF. INS (National Statistics Institute). 2019. Enquête par grappe à indicateurs multiples 2017–2018, Rapport de résultats de l’enquête, Kinshasa, République Démocratique du Congo, Décembre. Jean, N., et al. "Combining satellite imagery and machine learning to predict poverty." Science 353.6301 (2016): 790-794. Molina, I., and J. N. K. Rao. "Small area estimation of poverty indicators." Canadian Journal of Statistics 38.3 (2010): 369-385. Pape, U. and W. R. Kamara. 2019. Somali Poverty and Vulnerability Assessment: Findings from Wave 2 of the Somali High Frequency Survey (English). Washington, D.C.: World Bank Group. 25 Pape, U. and P. Wollburg. 2019. “Estimation of poverty in Somalia: Using innovative methodologies.” Policy Research working paper; no. WPS 8735. Washington, D.C.: World Bank Group. http://documents.worldbank.org/curated/en/509221549985694077/Estimation-of-Poverty-in- Somalia-Using-Innovative-Methodologies. Peterson, Ryan A., and Joseph E. Cavanaugh. "Ordered quantile normalization: a semiparametric transformation built for the cross-validation era." Journal of Applied Statistics (2019): 1- 16. Pokhriyal, N. and D. C. Jacques. "Combining disparate data sources for improved poverty prediction and mapping." Proceedings of the National Academy of Sciences 114.46 (2017): E9783-E9792. Sumata, C., and J. H. Cohen.2018. “The Congolese Diaspora and the Politics of Remittances”, Remittances Review.” 3: 95–108. Tarozzi, A. 2007. “Calculating Comparable Statistics from Incomparable Surveys, With an Application to Poverty in India.” Journal of Business and Economic Statistics 25: 314–36. Torabi, Mahmoud, and J. N. K. Rao. "On small area estimation under a sub-area level model." Journal of Multivariate Analysis 127 (2014): 36-55. World Bank. 2016a. Democratic Republic of Congo Poverty Assessment, Report No ACS19045. Washington, DC: World Bank. ———2016b. Urbanization Review: Scaling up Localized Development. Washington, DC: World Bank. ———.2018a. “Democratic Republic of Congo Urbanization Review: Productive and Inclusive Cities for an Emerging Democratic Republic of Congo.” Directions in Development. Washington, DC: World Bank. ———.2018b. Democratic Republic of Congo Systematic Country Diagnostic: Policy Priorities for Poverty Reduction and Shared Prosperity in a Post-Conflict Country and Fragile State, Report No 112733-ZR. Washington, DC: World Bank. ———.2020. Overview of the Labor Market: Demographic and Spatial Disparities in Outcomes within the Kinshasa Urban Landscape. Washington, DC: World Bank. 26 Annex A: Sampling weights calculation A.1 Survey period and data collection Data from the Kinshasa Household Survey were collected in two 18-day phases between mid- October and the end of November 2018 on a sample of 2,592 households. A.2 Sampling The last general census of the population of the DRC dates from 1984. The next census whose mapping phase would allow having an updated listing of households and EAs has not yet been executed. As a result, the sampling method was based on high resolution satellite imagery. A two- stage stratified sampling design with equal allocation of first-degree EAs were used for the survey. Kinshasa had been divided into two clusters—urban and rural. The 23 communes/districts of the city were considered urban while some areas of Maluku were considered rural. Two criteria were used to stratify the city (a) the building density of EAs and (b) typology of Kinshasa areas. The stratification of EAs according to the density criterion was carried out as follows: • S/Stratum 1 (low density): fewer than 10 buildings per hectare • S/Stratum 2 (medium density): from 10 to 20 buildings per hectare • S/Stratum 3 (high density): More than 20 buildings per hectare The typology refers to whether the neighborhood is considered precarious or not. Categorization of Kinshasa area is as shown in Table A.1. Table A.1: Categorization of Kinshasa area Definition This is a neighborhood whose dwellings are built by the occupants on land acquired in undeveloped areas, on farmland and market gardens, or on uneven ground (exposure to erosion and flooding). Buildings are anarchic and the use of recycled materials (cardboard, 1 = Precarious plastics, and sheet metal) is common. area Most often, they are located in parts of the city abandoned by the more affluent categories, such as on steep slopes and near industrial areas, which makes them all the more dangerous and where misery is concentrated. Collective facilities (water, electricity, sanitation, and transport) are reduced and availability is low. These are neighborhoods where homes are self-constructed, constructed by state, or 2 = Non- constructed with the intervention of a contractor. They are also neighborhoods built entirely by precarious area the state for social housing in accordance with urban standards (drainage system, access to water and electricity, tracing roads, and so on) and habitat. Combining the two criteria, seven strata were formed, including six strata from the urban part of Kinshasa and one from rural Maluku. The sample of households in urban area of Kinshasa was drawn as follows: Step 1: EAs are drawn independently in each stratum. In each of the strata, the number of primary units is drawn in proportion to the size of the municipalities. 27 Step 2: Listing of households of each EA from step 1 is made. Step 3: A constant number of nine households is randomly selected from the listing of households and surveyed. Sampling weights are derived as follows: : Number of dwellings in stratum : Number of dwellings in commune in stratum : Number of EAs in stratum ′ : Estimated number of households in EA of stratum : Number of households listed in EA of stratum for incompletely enumerated EA ( ) : Number of EAs in stratum in the survey sample : Number of households in EA of stratum in the survey sample : Area occupied by household in EA : Total area of EA : Empty area in EA : Area covered by the household enumeration in EA = , … , = , … , = , … , The sampling weight of a household from EA of stratum is: = ∗ × To overcome coverage of EAs, the following procedure was used: • For the EAs that were fully covered and extended to other EAs not included in the sample but also fully listed, we adjusted the drawing probability at the first stage to account for the change. • For the EAs that were incompletely enumerated, we estimated the total household number of the EA based on spatial distribution of the household in the area. Based on the area covered by the listing and the inhabited area of the EA, we calculated a coverage rate and applied this coverage rate to the inhabited area of the EA. ′ = � * with = − 28 Annex B: Consumption and poverty line determination Following global standards (Deaton and Zaidi 2002), welfare is best defined by (food and non- food) consumption expenditure. This annex documents choices made for the estimation of the household-level welfare aggregate as well as some issues of comparability between the two surveys and whether the issues are major or not. B.1 Household consumption Although household income and expenditure surveys collect information on household income, poverty measures are mainly based on consumption data particularly in developing countries. To measure the total household consumption, one has to take into consideration the following elements: monetary consumption (food and non-food), auto-consumption, transfers received in- kind, and the rent attributed to households who are not tenants in their accommodation. Monetary food consumption relates to all food products purchased in the market in exchange for money. Non-food monetary spending concerns other types of consumption (clothing, accommodation, body care, health care, and so on). Food spending consists of daily expenses covering food consumption requirements and transfers in-kind between households or destined for other sectors of the economy. Food auto-consumption was evaluated using data collected in the daily questionnaire and included in total spending on consumption. Using information concerning the frequency of consumption of these products or services, household consumption was calculated on a yearly basis. In addition, the following items were taken into consideration as everyday non-food spending by households: large-scale or exceptional spending on cereals and other food products, other daily non-food spending, clothing spending, accommodation spending (including estimation of attributed rent), furniture spending, health spending, transport spending, communication spending, leisure time spending, education spending, hotel spending, miscellaneous goods and services spending, and gifts received in-kind. The nominal household consumption aggregate was constructed broadly following the best practice guidelines provided in Deaton and Zaidi (2002) and consists of two main components: food and non-food consumption. It must be noted that there are limitations of household surveys in measuring household consumption expenditure for two reasons: (a) self-reported data rather than the data collected by direct measurements, and (b) impossibility to distinguish between consumption and expenditure; for example, a bulk purchase could cause overvaluation of household welfare. However, some categories of spending are excluded from everyday household consumption for the following reasons:  Some categories may be difficult to assign to a direct household consumption due to the significant presence of people from outside the household—this is the case for spending on festivals or ceremonies during the past 12 months.  Some categories represent investments in accommodation for the household—this is the case of spending on construction goods and services during the past 24 months.  Some categories do not actually represent household consumption—this is the case for gifts given or received in cash and taxes paid during the past 12 months. 29 Using the Deaton and Zaidi approach, the consumption aggregation process has adhered to the following principles:  Household expenditures on consumption are used as the money-metric measure of welfare.  Wherever possible imputation and data manipulation have been minimized.  Rent for owner-occupied housing or housing for which rent has not been reported is estimated using a hedonic rent model.  Lumpy expenditures, for example, hospitalization and purchase of large durables such as vehicles, have been excluded from the aggregation.  Food expenditures are cleaned at the item level and per capita level for outliers using quintiles.  Values per unit are computed from unit values constructed. Unit values are then used to compute total values of food purchases, own-consumption, and food consumed away from home.  All expenditures or consumption are converted to a uniform reference period (yearly expenditures). From the data, recall period for frequently purchased or consumed food and non-food items is 15 days (two weeks). Infrequently consumed/purchased items are recorded monthly or yearly.  The aggregate consumption expenditures are adjusted for household size and composition. This is because most surveys use the household as a unit of observation in the measurement of consumption. The reason being that it is both costly and time consuming to collect consumption data on an individual basis. It also facilitates the treatment of joint household goods, such as housing, where it is not possible to assign consumption to specific individuals. B.2 Asset index, identification, and treatment of outliers To assess the robustness of a consumption-based poverty profile, an assets index, also including education variables, was constructed. The assets index was obtained from factor loadings (first axis) of a standard multiple correspondence factorial analysis. The factor loadings of the principal axis were adjusted for household size and normalized to take a value between zero and one. The normalized index was used as a proxy welfare indicator. The percentiles of the normalized asset index (deciles, quintiles, quartiles, or terciles) have been used throughout the analysis including to impute value to outliers. Outlying records, those which greatly diverge from the rest of the data, can be of concern for consumption-based poverty and inequality analysis. Some methodological choices for dealing with outliers can be detrimental to inequality analysis. For any given household expenditure variable, identification of outliers relies on a non-parametric method using 5 percentile and 95 percentile observations for specific groups of households defined by location characteristics and 30 assets-based poverty status. The outliers are regarded as missing values and imputed with the 5th percentile, the 95th percentile, or median values of their groups when each of these values is used. B.3 Food aggregate The constructed food consumption aggregate must include estimates from all possible sources related to food use by the household as a unit or members as individuals. Total food spending consists of expenses covering food consumption, food auto-consumption, and of transfers in-kind received by households. The food component of the questionnaire was mainly captured through diary recordings. Households’ daily consumptions were recorded using a diary for 15 days. In processing the data and based on assets, checks were done for outliers on quantities consumed and correcting missing unit prices especially for own-consumption items and gifts. This was followed by several best practice correction procedures that were aimed at deriving an appropriate food aggregate. These steps are as follows:  Step 1. First major item categories and subcategories were generated. The diary item groups and subgroups were then combined to household weights, quintiles, household size, adult equivalent, sample weights, and multidimensional poverty files that were already generated. This file was generated earlier and combined at this stage to offer an expedient data cleaning process.  Step 2. Then we clean the number of days the diary was filled to ensure that it is between 1 and 15 days. For consistency, we only kept households whose dairy was filled for seven days or more. We then standardize the frequency of purchase or item renewal frequency.  Step 3. Then the food units were checked for each item. If both a standard unit and nonstandard unit was provided for the same item, the standard unit was deemed more reliable and thus the nonstandard unit was replaced.  Step 4. If acquisition mode is missing for some items, a frequency was done by item, quintile, province, pool, site, and milieu and the most common and preferable mode of acquisition was used. For consistency purpose, in the analysis, we only kept households whose diaries were filled for seven days or more. We then scale the purchase to get an annual value using the item renewal frequency. B.4 Non-food aggregate Different recall periods (diary, 1 month, 3 months, 6 months, and 12 months) were used to derive total non-food expenditures. Constructing the non-food aggregate thus entails converting all the reported values to a uniform reference period—one year for consistency and then aggregating across the various item categories. Different non-food sections are then combined, and relevant expenditures were considered. Special consideration in non-food expenditure aggregation was made for (a) education, (b) health, (c) housing, and (d) durable goods. 31 Education. It was included as it contributed to the household utility. Two modules in the questionnaire recorded education expenditures: education module with all categories of education expenses for the last 12 months and education expenditures module. The first module was used as the primary source of data and the second to impute nonresponse. A frequency was decided upon for total education costs and the observations were perused to see if there were any strange values. The same was done by the different categories. Health. The inclusion of health-related expenditure is somewhat controversial because one cannot distinguish between necessary and non-necessary expenditures and what they are defined as. The best practice recommends that health expenditures are included if higher expenditure is evidence that households want to purchase better quality service and therefore associated with higher utility. Lumpy expenditure items are excluded as it is difficult to distinguish between those paid by households or other relatives/friends. Housing. Expenditure on housing is an important component of non-food consumption aggregate. One subcomponent of housing expenditure is rent. However, including rent is not straightforward. For example, including rent for a renter and not for an owner will underestimate the welfare of owner households. An econometric modelling is used to impute a rent value for those who are not paying for housing, based on the Ordinary Least Squares (OLS) and Heckman two stage estimation methods. The regression analysis used declarations of rent paid by households which are tenants, by considering the following explanatory variables: area of residence (town, district, and sector); distance from the core city; type of housing; materials used (walls, floor, and roof); number of rooms; combustible used for cooking; lighting source; water supply source; level of affluence of the area; and waste disposal method. Finally, both actual and imputed rent were used for uniformity. Durables. Ownership of durable goods (non-production items) is an important component of the welfare of the households. However, these goods last typically for many years and the expenditure on the item is not the proper indicator to consider for welfare. For consumption purposes, the right measure to estimate is the utility (services) that households derive from all durable goods in their possession over a reference period. This flow of utility is unobservable, but it can be assumed to be proportional to the value of the good. In the surveys, this could not be derived due to lack of adequate information. B.5 Adult equivalence scale Due to economies of scales and the fact that calories intakes differ across life span and gender, consumption aggregates are often constructed in terms of consumption per equivalent adult rather than in per capita terms. For adult equivalent we relied on the adult equivalence scale recommended by the Food and Agriculture Organization, which would seem closer to the reality in the DRC. B.6 Poverty lines In practice, determining the poverty line involves several key steps starting with determining a basic minimum calorie requirement, creating a food basket, and evaluating the cost of meeting the calorie requirement using that food basket within the poor population. The cost of this basket is 32 the food poverty line which is used to determine the proportion of the population that is unable to meet the minimum basic food consumption needs (that is, the food poor). A minimum allowance for non-food consumption is then added to the food poverty line to determine the overall poverty line, which is used to determine the proportion of the population that is unable to meet the minimum overall basic consumption needs (that is, the absolute poor). The Kinshasa Household Survey failed to provide quality data on quantities in kilogram of food items purchased by households on a daily basis. The price and nonstandard unit surveys were implemented few months after the survey, causing seasonal products to disappear. The food basket used for the consumption price index consists of a third list of items included in the consumption module and prices and quantities are collected from the three largest markets serving more than half of the city. It was challenging to derive a food basket composition and compute the food poverty line using the standard method. To overcome this issue, the 2012 food poverty line has been updated using the inflation rate of food and soft drinks items between the two periods (2012 and 2018). Having determined the food poverty threshold, the non-food poverty threshold has been estimated. The non-food poverty line is the non-food consumption of a household whose total consumption is equal to the food poverty threshold. In practice, the value of non-food consumption is estimated as the average value of non- food consumption of households whose total consumption is within ±10 percent of the food poverty threshold. We then determine the poverty threshold as the sum of the food threshold and the non-food threshold. 33 Annex C: The reweighting using the propensity score approach The basic idea here is the assumption that, unlike subsample A of households from completely and properly enumerated EAs, there is a bias in subsample B of households from the EAs where enumeration is incomplete or where the EAs are resized because of overflowing. Using the two subsamples as one sample, like what is done in this study, without making adjustments in subsample B may lead to biased poverty estimates. For robustness checking, the reweighting approach (inverse of the propensity score) suggested by Tarozzi (2007) is used to correct the bias in subsample B. The results are then compared with those obtained using just the sample without this kind of adjustment (see Table C.1). The approach is to correct the sampling weights of the households in subsample B so that they constitute a counterfactual group of subsample A. To do this, we assume a dummy variable which takes the value 1 when the household comes from subsample B, and 0 when it is part of subsample A. Subsample B is considered the target subsample while subsample A is the ancillary one. Considering the whole sample, we estimate the propensity score using a probit model where the dummy variable is the dependant variable, while the explanatory variables, which are common to both subsamples, are assumed to affect the probability of a household being in one or other of the subsamples. These explanatory variables include household size and composition (dependency ratio); household head characteristics (age, education, sex, and sector of activity); the characteristics of the dwelling (type of habitat, floor, wall, and roof); the access of basic services (sanitation and drinking water); and geographic variables (stratum and distance to core city). The inverse probability weight (propensity score) derived from this model is used to adjust the sampling weight of subsample B and make unbiased welfare and poverty estimates. Table C.1: Comparing poverty results between the non-adjusted sample and the adjusted sample Incidence of poverty Depth of poverty Severity of poverty Dif. with Dif. with Dif. with P0 the non- P1 the non- P2 the non- adjusted adjusted adjusted Precarious at low density 71.4 0.4 31.3 0.0 17.4 –0.1 Precarious at medium density 63.0 –1.1 29.8 –0.9 16.8 –0.6 Precarious at high density 68.7 0.6 27.2 –0.8 14.2 –0.6 Non-precarious at low density 52.3 –1.8 20.2 –0.6 9.8 –0.3 Non-precarious at medium density 37.0 –0.9 10.7 –1.7 4.4 –1.3 Non-precarious at high density 36.4 2.4 10.0 0.3 4.2 0.3 Urban Kinshasa 63.4 –1.1 26.6 –1.2 14.3 –0.8 Source: Kinshasa Household Survey 2018. 34 Annex D: Technical Note on Kinshasa Small Area Poverty Estimation This technical note outlines a methodology used to estimate poverty rates at the subnational level (quartier and commune) for urban Kinshasa. Obtaining accurate and reliable estimates of local poverty (like at the neighborhood level) is difficult due to the high cost of collecting welfare data from sufficiently large samples to allow for such analysis. Household surveys – from which poverty estimates are derived – are typically too small to produce reliable estimates below a certain geographical level. To overcome this challenge, our estimation relies on the combination of a household survey – the 2018 Kinshasa survey – and remote-sensing and geospatial data. Due to its small sample size, the Kinshasa survey does not allow us to produce precise poverty estimates for quartiers or communes. Remote-sensing data and satellite imagery – which are available at a grid level and covers the entire province of Kinshasa and DRC – provide useful supplemental information on socio- economic conditions of local areas that helps improve the coverage and precision of local poverty estimates. Methodology Traditionally, poverty mapping methods estimate a random effect regression model using survey data containing per capita income or consumption data and use the estimated parameters to simulate welfare in a contemporaneous census. Because there is no recent census in DRC, applying this traditional method is not feasible (Lange, Pape, and Putz 2018). Thus, we combine the household sample data with remote sensing indicators at the Primary Sampling Unit (PSU) level, linking the source and auxiliary data geographically through PSU identifiers. To combine these two data sources, we employ the Empirical Best Prediction (EBP) Method (Molina and Rao 2010, Battese, Harter, and Fuller 1988). The EBP modifies the traditional ELL method (Elbers, Lanjouw, and Lanjouw, 2003) in two main ways. First, random effects are introduced at the level of areas (in our case, quartiers or communes) instead of the enumeration area. Second, these area random effects are conditioned on the sample data. This method therefore efficiently combines household level information on per capita consumption, which are only available in sample PSUs, with an exhaustive set of PSU level prediction based on indicators derived from remote-sensing and satellite data. In the EBP method, the two-level nested error regression model is first estimated: = + + , = 1, … . , = 1, … . , 2 ), 2) ~(0, ~(0, (1) where corresponds to normalized per capita consumption 15 for household i in subnational areas g. and are area-specific and household-specific random errors. The EBP estimates of FGT0 15 A correct transformation of the skewed welfare or consumption variable is needed to render the random errors normally distributed – a critical assumption that needs to be met for EBP (see for more details Das and Haslett 2019). We utilize an Ordered Quantile Normalization to transform the dependent variable using the bestnorm R package, which most consistently ensures normality compared to other transformation methods (e.g., log, log- shift, or BoxCox transformations) (Peterson and Cabanough, 2020). 35 are obtained by estimating the nested-error model (Equation 1); generating area-level effects ∗ ~(0, 2) � and unit-level effects ∗ ~(0, 2) � ; and then calculating population welfare values through micro-simulation based on the sample and non-sample values of explanatory variables as specified in Equation 1. Importantly, the EBP model conditions the area effect on the mean of the sample residuals, thereby incorporating information from the sample directly into the estimates. To estimate the Means Squared Errors (MSE) of FGT0, we follow Molina and Rao (2010), and use a parametric bootstrap MSE estimator, following the bootstrap method for finite populations of González-Manteiga et al. (2008). 16 Alternative estimates generated using a PSU level model, a variant of the sub-area estimator proposed by Torabi and Rao (2014), yielded similar results. The specification of the consumption model in Equation 1 affects the area-level poverty rates. The set of variables included in each specification was chosen from the list of candidate variables using the stepwise selection process (with a significance level threshold of 0.001 for removal). The stepwise selection process yielded a model with 20 right hand side variables as reported in Table 1. The model achieved a reasonably high R2 of 0.31, despite the fact that the explanatory variables are at the PSU level whereas consumption is measured at the household level. Finally, we apply a rescale factor to ensure that the aggregate poverty rate of EBP estimates is consistent with that of direct estimates based on the Kinshasa survey, which is representative collectively for urban areas of Kinshasa. 17 Data Estimating area-level poverty estimates in the EBP framework involves deriving zonal statistics based on remote-sensing data and satellite imagery for each PSU, which then are matched with welfare data from the Kinshasa. These zonal statistics are then used to predict welfare levels in PSUs not included in the sample survey. The geospatial data used in the prediction model are drawn from a number of different sources, which include nighttime light data from the Visible Infrared Imaging Radiometer Suite (VIIRS); population data from WorldPop; built-up area information from the Global Settlement Layer, measures of flood risks 18; distance to schools, hospitals, and churches derived from OpenStreetMap; the size of EA area; as well as strata information. Next, we use geospatial information in the Kinshasa survey to link with the PSU-level zonal statistics. The computed zonal statistics are then used to model consumption/welfare as defined in Equation (1). Finally, we use the R SAE package to simulate the predictions in a synthetic census of households. Population estimates extracted from WorldPop were used to determine the number of households in each PSU. 16 The R “sae” package has been used to implement this estimation. 17 The scale factor used in our estimation is 1.4, which means that our spatial features overall understate the level of poverty compared to the urban poverty rate of Kinshasa as derived from the Kinshasa survey alone. 18 SSBN 3 arc-second (90 m) Global Flood Hazard Data (World Bank License). 36 Results The consumption model – the model used to predict consumption as a function of spatial covariates (Equation 1) – has a reasonably high 2 of 0.31, which means that the geospatial variables explain 31 percent of variation in household consumption in the first stage. This is impressive given that the geospatial variables are only capable of explaining variation in household-consumption across PSUs. The results from the consumption model are presented in Table D.1. Figure D.1 shows estimated commune-level poverty rates with the confidence intervals. The average coefficient of variation – which is derived as the estimated standard error divided by the estimated poverty rate and averaged across all areas – is 15 percent for communes and 31 percent for quartiers, which mean that our estimates at the quartier level are much less precise than commune-level estimates and thus need to be treated with caution. One thing to note is that our spatial features consistently underestimate poverty rates vis-à-vis direct survey estimates if they are not rescaled to match with the direct survey estimates. For instance, Table D.2 compares our EBP estimates against direct survey estimates at the stratum level for which the survey is representative and there is a significant gap between EBP and direct survey estimates. This gap can be driven by a number of different factors. One possibility is that EBP estimates are weighted accordingly based on the estimated population distribution based on WorldPop instead of sampling weights used in the Kinshasa survey, discrepancies in these weights may introduce some bias in our EBP estimates. Second, strata-level poverty estimates based on the survey may somehow overrepresent areas that are poorer and our spatial features adjust accordingly by entirely covering all those EAs that fall into each stratum (because our spatial features are available for all EAs). Figure D.1: Poverty rates at the commune level (with 95% CIs) 1.2 1 0.8 0.6 0.4 0.2 0 Kimbanseke Kinshasa Selembao Kisenso Bumbu Matete Barumbu Masina Kasavubu Ngaba Kalamu Kintambo Limete Nsele Mont Ngafula Makala Ndjili Gombe Ngaliema Lingwala Lemba Ngiri-Ngiri Bandalungwa 37 Table D.1: Beta Model Results (Variables Selected through Stepwise Process) (1) VARIABLES Medium_FLOOD__Medium_CD-UU-10-3_MAX 1.449*** Low_VIIRS_201703_SUM 0.018*** Medium_FLOOD__Medium_CD-M-3_MAX 10.321*** Medium_FLOOD__Medium_CD-UD-5-3_MAX -1.392*** Medium_FLOOD__Medium_CD-UU-1000- - 3_MEAN 39.047*** Medium_FLOOD__Medium_CD-PD-50-3_MAX -5.068*** Low_VIIRS_201704_SUM -0.015*** strate_2==NP low 0.279*** Medium_GHSL__c_2 -0.003*** Medium_FLOOD__Medium_CD-PU-250- 3_MAX -1.998*** Medium_FLOOD__Medium_CD-UU-250- 3_MEAN 6.041*** Medium_FLOOD__Medium_CD-FU-100- - 3_MEAN 14.106*** - Medium_FLOOD__Medium_CD-M-3_MEAN 10.855*** Area 1.570*** Medium_FLOOD__Medium_CD-PD-75-3_MAX 4.922*** Low_VIIRS_201704_MEAN 0.039*** Medium_FLOOD__Medium_CD-UD-10- 3_MEAN 6.872*** Medium_FLOOD__Medium_CD-FD-250- 3_MAX 7.402*** Medium_FLOOD__Medium_CD-UU-1000- 3_MAX 1.147*** Medium_FLOOD__Medium_CD-PU-1000- 3_MEAN 33.939*** Constant -0.461*** Observations 1,608 R-squared 0.311 F 140.0 Note: Standard errors in parentheses clustered by quartiers. * p < 0.1, ** p < 0.05, *** p < 0.01. Original raster data are available at the pixel level and aggregated by PSU before being used for the consumption model. The minimum, maximum, mean, and total values of those pixels at the PSU are computed and used to explain variation in consumption, as indicated by _mean, _median, _max, _min, and _sum in the variable names. Below is a brief description of each variable included: Medium_FLOOD__Medium_CD_{TP}_{Y}_3_: the maximum expected water depth in metres at different return periods for different areas. {TP} corresponds to data type where T can be coded as F (fluvial; riverine flooding); P (pluvial; surface water flooding from extreme rainfall); U (maximum of fluvial and pluvial flooding); M (urban mask). P can be coded as D (defended; including the effects of estimated flood defenses) and U (undefended). {Y} corresponds to return periods in years. Low_VIIRS_{YYYY}{MM}_: VIIRS nighttime light luminosity. {YYYY}{MM} correspond to year and month in which data are collected. strate_2==NP: A dummy variable for non-precarious low-density strata. Medium_GHSL__c_2: Share of non-built up areas. Area: area of EAs. 38 Table 2: Comparison of EBP estimates vis-à-vis direct survey estimates from the Kinshasa survey at the strata level Std. errors of Std. errors of FGT0 from FGT0 Direct survey Strate EBP direct survey EBP rescaled estimates estimates estimates Low density - 0.4950 0.0185 0.6860 0.7096 0.0849 precarious Medium density - 0.5396 0.0231 0.7479 0.6413 0.0904 precarious High density - 0.5468 0.0201 0.7579 0.6809 0.0449 precarious Low density – 0.2913 0.0153 0.4037 0.5405 0.1166 non-precarious Medium density 0.2575 0.0169 0.3568 0.3785 0.0592 – non-precarious High density – 0.2918 0.0173 0.4044 0.3452 0.0441 non-precarious Notes: FGT0 from EBP: EBP estimates of stratum-level poverty rates FGT0 Std. Error from EBP: Standard errors of EBP estimates FGT0 rescaled: EBP estimates rescaled to ensure that the overall poverty rate matches with that of the Kinshasa survey Direct survey estimates: Direct stratum-level poverty estimates from the Kinshasa survey 39 Annex E: Main correlates of the log of consumption per adult equivalent in Kinshasa, 2018 Standard Coefficient T-statistic Probability Error Demographics Household size -0.131*** 0.019 -6.89 0.000 Squared household size 0.002*** 0.001 2.69 0.007 Dependency ratio -0.062** 0.027 -2.31 0.021 Household head is female -0.169*** 0.060 -2.81 0.005 Education level of household head (the control is no education) Incomplete primary 0.319** 0.151 2.11 0.035 Incomplete secondary 0.140 0.135 1.03 0.302 Secondary 0.300** 0.130 2.31 0.021 University 0.715*** 0.142 5.03 0.000 Industry of household head (the control is construction industry) Unemployed -0.038 0.112 -0.34 0.734 Agriculture -0.220 0.1861 -1.18 0.237 Manufacturing 0.150 0.136 1.1 0.271 Commerce 0.170 0.109 1.56 0.12 Restaurant/Transport/Communication 0.218* 0.118 1.85 0.064 Public administration 0.396*** 0.121 3.27 0.001 Education/Health 0.044 0.127 0.35 0.726 Other activities -0.007 0.104 -0.07 0.945 Sources of income (the control is only labor income) Labor income and Remittances only 0.069 0.054 1.27 0.205 Labor and non-labor income (remittances excluded) 0.194*** 0.064 3.02 0.003 Labor and non-labor income (remittances included) 0.278*** 0.103 2.68 0.007 Non-labor income only 0.129 0.095 1.36 0.175 Distance to the core city (the control is 5 km or less) Between 5 and 10 km -0.408*** 0.065 -6.26 0.000 Between 10 and 15 km -0.527*** 0.063 -8.38 0.000 More than 15 km -0.534*** 0.069 -7.72 0.000 Constant 14.873*** 0.186 79.87 0.000 Number of Observations = 1,585 Adjusted R-squared = 0.459 Probability > Fisher = 0.000 Note: (*), (**) and (***) denote that the coefficients are significant at 10 percent, 5 percent, and 1 percent levels, respectively. Source: Kinshasa Household Survey 2018. 40