Dynamically Identifying Community-level COVID-19 Impact Risks

We build a new database of highly spatially disaggregated indicators related to risk and resilience to the social and economic impacts of the COVID-19 pandemic in Uzbekistan. The outbreak disproportionately affects particular groups – the elderly, the poor, those living in areas under lockdown, and families who rely on remittance income are all examples of groups that are especially vulnerable to effects of the crisis in Uzbekistan. We assemble indicators summarizing concentrations of these and other risk factors at the lowest administrative level in the country, neighborhood-sized units called mahallas. Local official administrative statistics (published for the first time in this study) are combined with monthly panel survey data from the ongoing Listening to the Citizens of Uzbekistan project to produce an overall risk index, which is decomposable by dimension or risk factor to inform targeted and issue-specific responses. We then demonstrate a process for updating key indicators (such as employment or remittance flows) on a monthly basis using linked survey data combined with small area estimation techniques. These neighborhood-level results are intended to improve resource allocation decisions and are particularly relevant in Uzbekistan where local representatives are responsible for implementing key social and economic programs to respond to the outbreak.


I -Introduction
The impacts of the coronavirus pandemic on health and economic wellbeing are unprecedented. As of this writing, the disease has claimed more than 540 thousand lives around the world, World Bank estimates suggest that extreme poverty has increased more than at any other time since the Second World War, and per capita incomes have suffered the largest decline since 1870. 6 It is also the most severe crisis Uzbekistan has faced in generations, with expected annual GDP growth for 2020 reaching its lowest point since independence from the Soviet Union. Although the outbreak in Uzbekistan has remained moderate thus far -the official case count is currently about 14 thousand and the death toll of 68, for a country with a population of more than 34 million -Uzbekistan has not been spared the economic impacts of the crisis. In April 2020, the authorities introduced lockdown measures of all non-essential work and travel to protect public health. As the health situation permitted, restrictions were gradually relaxed in May and June. However, national lockdowns were reintroduced on July 10 th due to a resurgence in the rate of infection. These lockdowns have caused similar collateral economic damage that has been seen elsewhere in the world, leading to sharp declines in employment, income, and other measures of economic wellbeing.
A national monthly household panel survey focused on social and economic wellbeing called Listening to the Citizens of Uzbekistan (L2CU) was in the field leading up to and following the COVID-19 outbreak in Uzbekistan. Data collected in the survey are used in this study to measure the impacts of the crisis and extrapolate lessons to individual communities throughout the country that can be used to guide anti-poverty and recovery efforts. Following the outbreak, the core survey instruments were expanded to cover focus areas relevant to COVID-19 and the economic impacts of the unfolding global recession. Analysis of the ongoing survey modules that monitor employment, migration, and similar themes gained added urgency as a result of the crisis, especially as traditional surveys in the country use in-person interview methods and were partially disrupted during lockdowns. Fortunately, data collection activities conducted in the L2CU project have thus far been unaffected.
Results from L2CU reveal dramatic declines in employment and incomes beginning in April 2020, as well as very high levels of concern about the health and economic impacts of the pandemic among the population. World Bank projections find that between .5 and .8 million additional people will likely fall into poverty in 2020 -with high risks of further deterioration in the event of a more extended emergency. Considering the widespread impact of COVID-19 globally, there are also likely to be both large reductions in remittances and increased domestic unemployment, as the lingering pandemic severely affects both domestic and international businesses. Production and import disruptions increase the risk of inflation. Together, these factors are likely to have a profound and long-lasting impact on the overall wellbeing of the population, increase the poverty levels in the country, and create deep hardship for those who are directly affected.
The aim of the analysis described in this paper is to identify risk factors of COVID-19 impacts at the level of small communities, such that policy makers can prioritize actions and support those in greatest need. The smallest administrative unit in Uzbekistan is called the mahalla (Figure 1). The leaders of these communities collect administrative data on an annual basis, and these statistics are also used in this study. The databases generated by local authorities are conventionally aggregated to the district-level in Uzbekistan (including 200 districts and urban areas in total) and primarily used for implementing local policies. We are aware of no previous instances when these data have been systematically gathered into a single database and harmonized at the national level. These data are published together with the results of this paper for use by policymakers and other partners responding to the impact of COVID-19.
Uzbekistan is subdivided into 9145 mahalla neighborhoods in both rural and urban areas (the precise number frequently changes when small mahallas are merged and large mahallas are split). Mahallas usually range in size from between 500 to 10,000 families. The mahalla is a formal institution, and each has a defined geography, though the cartography of the units is not digitized at this time. All maps presented in this paper report aggregate statistics at the district level. However, the data file including mahalla-specific results is published together with this paper for direct use at the level of mahalla.

Figure 1: Administrative Units in Uzbekistan
The official activities of a mahalla are organized and carried out by an executive committee (Mahalla fuqarolar yig'ini) under the leadership of a chairperson (Raiis). Though mahallas are grounded in local tradition, today, mahalla officials implement many state functions including data collection, implementing public information campaigns, and administrative duties related to the social assistance system. The role of the mahalla has been is a state of flux in recent years due to policy and regulatory developments in Uzbekistan since 2017; however, the core social assistance related activities of mahalla leadership are particularly relevant as policy makers expand the provision of benefits to combat the impact of the COVID-19 pandemic.
The results of this study are intended to support efforts to prioritize local interventions in response to the impact of COVID-19. A growing body of evidence from Uzbekistan and elsewhere in the world finds that individual and community level risk profiles from the effects and aftereffects of the pandemic are highly variable. Membership in particular age groups, employment in particular sectors, and access to sources of resilience all play a role in how the crisis will affect a person, family, or village. For instance, work published by the Furman Center found 7 that the both the direct and indirect effects of the pandemic have been highly localized in particular populations in the city of New York, in the United States. Neighborhoods with higher rates of confirmed COVID-19 cases were shown to have much lower median incomes, higher shares of residents from Black or Hispanic minority groups, and higher shares of residents under the age of 18 relative to less affected neighborhoods. Residents of disproportionately affected neighborhoods were also shown to be less likely to be able to work from home, disproportionately reliant on public transit during the crisis, and less likely to have internet access. Finally, neighborhood level analysis found that areas with higher numbers of confirmed COVID-19 cases had lower population density, yet higher rates of overcrowding at the household level. Chetty, et. al. (2020) similarly show that neighborhood-level impacts are highly specific and variable. Using high frequency private sector data, the authors demonstrate the heterogeneity of outcomes with respect to incomes and local economic factors. Due to data limitations, the full approach adopted in that study is not possible in all countries, and the range of data available for Uzbekistan is more limited. However, the intuition and objective of the analysis that follows are quite similar to those of Chetty, et. al. (2020) and the Furman Center, if considerably less ambitious in terms of spatial, temporal, and topic granularity.
To summarize, we assemble the data from the highly disaggregated survey and administrative sources described above. The resulting database allows us to develop a variety of measures related to the impacts of COVID-19 at the local level. We then use these indicators to construct a community-level COVID-19 risk index for Uzbekistan. The ultimate aim of the study is to prepare a database of relevant indicators to aid in the design of response and recovery programs. In some cases, these indicators can be updated over time using linked survey data and small area estimation techniques. The summary index includes six dimensions -related to age and ability risk factors, economic conditions, access to social assistance, local services infrastructure, reliance on remittances from migrants, and local measures of monetary poverty -and is comprised of a total of 26 individual mahalla-level indicators. The dimensions and indicators can be decomposed as needed for targeted interventions.
The remainder of this introductory section describes the index dimensions in more detail, as well as their relevance to the outbreak and related economic consequences. Section (II) describes the data used, and section (III) the methods of analysis applied (with additional details provided in annexes). Section (IV) reports the results of small area imputations at the mahalla level (for those high-priory indicators that are not observed in administrative statistics). Section (V) describes the results of the index, both overall, and by dimension. Finally, section (VI) provides examples of dynamic updates of key indicators using linked panel survey data and small area imputation techniques.

Elderly and Disabled Populations
Uzbekistan has a relatively young population and the elderly represent a comparatively small share of the total -only about 4.8 percent of citizens are age 65 or older (figure 2). However, the wellbeing of this population is of particular concern. Older people are at much higher risk of health compilations and are more reliant on services that may be impacted by the pandemic. Evidence in many counties has highlighted the greater severity of the disease among older people, and especially high rates of mortality have been concentrated in communities of older people, care centers, and in nursing homes. Goldstein and Lee (2020) find that about 75 percent of all US Covid-19 deaths to be among people aged 70 or above, somewhat above the 64 percent for normal mortality. In China (Hubei), South Korea, Italy, France, and Spain, virus-attributed mortality rates rise by about 11 percent per year (a bit slower in Hubei, where the rate is 9.5 percent), close to the 10 percent that would be expected for allcause mortality. A national analysis of comorbidities in China found strong associations with chronic obstructive pulmonary disease, diabetes, hypertension, and malignancy (Guan, Wei-jie, et. al., 2020). These illnesses are also more prevalent among older populations. In the analysis that follows the elderly and disabled are identified as particularly vulnerable groups. Beyond infection and mortality risks, older people are also expected to have more difficulty adapting to lockdowns and other disruptions of normal life. Older people have more limited information communication technology (ICT) skills on average, which may prevent them from accessing internetbased services or for leveraging other communications needs. Many older people also rely on help from relatives and others who may be prevented from visiting during the pandemic. To address this challenge, there is presently some COVID-19 related support for the elderly provided by government beyond standard pensions. This includes eligibility to receive a package of food from local officials (largely targeting single seniors through the Sponsor Coordination Center). However, there are some concerns as to the adequacy of these measures: wait times have been reportedly quite long (often 3-4 days), and some have reported that care packages are insufficient. Elderly people commonly are also more reliant on the health system, which is strained by the extraordinary demands of responding to COVID-19.
Likewise, people with disabilities often require specialized services that are reduced or unavailable during lockdowns and related disruptions. Health facilities and other buildings are not disability friendly (lacking, for example, accessible toilets and handwashing facilities). UNFPA in Uzbekistan has reported concerns that information dissemination on COVID-19 related issues are not always disabled-friendly. Many people with disabilities struggle with accessing markets, especially while navigating lockdowns and quarantines and while personal support networks are reduced in their functioning. The L2CU baseline survey collected information on disability using the standard Washington Group questions regarding vision, hearing, walking, remembering, ability to provide selfcare, and communication, reported in Figure (3). These results highlight that disability and age are not completely separate considerations: there is large overlap of age risks and disability status.

Economic Factors
Data from April 2020 show that the economic impacts of the outbreak on livelihoods -including through reduced employment and income -has been severe. According to L2CU, the share of households with at least one member actively working fell more than 40 percentage points (from 85 to 43 percent) between March and April in 2020 when lockdowns were instituted to prevent the spread of the disease. But while incomes fell for a large share of the population (median per capita income combined from all sources fell by 38 percent in comparison to the previous month), there were also clear concentrations among particular populations. Individuals with stable formal employment in large firms, SOEs, or government, as well as those relying largely on predictable government transfers (e.g. old-age pensions) were relatively more protected than those with less certainty in their employment and activities. In contrast, those working in sectors particularly reliant on in-person interaction, including retail and other services, construction, transportation, and small-scale business were at much greater risk to the economic consequences.
When lockdown measures were phased out in stages during May and June 2020, a labor-market recovery quickly asserted itself (figure 4). The share of households with at least one working member rebounded by 33 pp in May. Reporting that someone "lost a job or stopped work" in the household jumped from 1 to 19 percent in April, before falling back to 3 percent in June. Nearly all respondents to the survey stated that they believe work disruptions are temporary. However, at the time of this writing, employment remains far below both 2019 levels and the pre-COVID trend. In addition, these statistics do not reflect the reintroduction of stricter lockdown measures effective from July 10, which will likely reimpose economic costs and disruptions of the labor market.  The declines in employment and incomes were largest among the self-employed. In April, the share reporting any self-employment income fell by 67 percent in comparison to the previous month and remained down 26 percent in June. In contrast, the share reporting any wage income declined by 16 percent., but on average re-attained its 2019 level in June, crossing that threshold more quickly among men than among women. Urban incomes started higher and fell faster than in rural areas, due in part to the start of the agricultural season and the relatively limited impact of the lockdown measures on the sector. Thus, declines were larger in urban areas -falling 46 percent in a single month -but were also high in rural and semi-urban areas (37 percent).  Data on new online job listings also showed signs of recovery in May and June, after new listings fell by 80 percent following the outbreak (figures 5 and 6). Sectors with particularly large declines compared to the same period in 2019 included tourism, recreation and entertainment (-95 percent), bars and restaurants, (-91 percent), and education (-85 percent). Even the least affected occupations, declined by 50 percent or more compared to the same period last year, though in June there was a quick recovery in medical and construction sectors ( Figure 6). The challenges posed by recovery will exacerbate difficulties in the labor market that were present pre-COVID-19. While the working age population has been increasing over time in Uzbekistan, formal job creation has not kept pace, resulting in high informality, inactivity rates and growing outmigration. The working age population increased by some 50 percent since 2000, from 14 million to 22 million today. Unemployment and inactivity rates are higher especially for youth, women and people in the poorest two quintiles. Job quality and inclusiveness remain a concern, as average wages are low Based on the recent L2CU data (2018), the lack of jobs as well as the low salaries are main concerns especially among the poorest and the beneficiaries of social assistance.

Social Assistance and Transfers
Uzbekistan has several targeted cash assistance programs to support low-income people (Table 1). These include social assistance (noncontributory), social insurance schemes (contributory), and labor market programs. Entitlement to social insurance programs is conditional on contributions that people make when they work and are supposed to protect people during old age or maternity, or in case of accidents and sickness. Social assistance benefits include four types of programs: • cash allowances provided to low-income households (means-tested benefits); • cash allowances provided to the elderly, persons with disabilities (PWD), and survivors (breadwinner loss); • allowances in case of special events or shocks; • allowances, discounts, and in-kind support to vulnerable groups.
The first type of allowances is means tested, i.e., conditional on household income being below a fixed eligibility threshold (expressed in per capita terms and equal to 1.5 times the minimum wage). Social assistance is provided through two distinctive administrative channels: mahalla and khokimiyats (regional governorates) are responsible for the administration of the low-income family allowances.
Means-tested benefits rely on identification processes administered by local community (mahalla) officials. Almost all other social allowances are administered through the national pension fund, which has an office in each district. Employment Services Centers are responsible not only for labor market programs, but also perform a monitoring function for the low-income family allowances.
Existing targeted social assistance programs have modest inclusion error, but substantial exclusion error due to budget-related caps on the number of beneficiaries (Figure 7). A World Bank assessment of the three main targeted cash assistance programs found more than 70 percent of beneficiaries were members of the bottom 40 (modest inclusion error), but that 63 percent of the poor were not reached by low-income allowances (relatively high exclusion error). The assessment further found that one of the main reasons for exclusion errors is the use of caps in budgeting and in the number of beneficiaries at the local level. The cap results in a rationing behavior, whereby limited resources are spread across eligible households, assigning allowances at a lower amount, or trigger a rotating approach, whereby applications are the facto postponed or payments of eligible applications are delayed. The system also struggled with suboptimal transfer amounts. This imbalance means that among the poor receiving support, only one-half are pushed above the poverty line commonly used by the World Bank for lower middle-income countries.

Local Health Services and Density
Uzbekistan has a network of public health centers represented at every regional and district level. The public health centers include virology laboratories, rapid response teams, epidemiological staff, units responsible for infection prevention and control. Uzbekistan also has an extensive network of state health facilities, including primary care facilities, district and regional general and pediatric hospitals, emergency care hospitals, and specialized inpatient care centers. Throughout the healthcare system, there is a relatively large hospital bed capacity, which is likely to be able to absorb initial surge needs in hospital overall, and specifically in intensive care units if repurposed and complemented by the necessary equipment and human resources. There are 334 acute beds per 100,000 population in Uzbekistan, compared to 290 beds in United States and 275 beds in Italy.
However, the Uzbek health system still faces many challenges in mounting effective prevention and control measures against COVID-19. Public health staffing levels have seen significant cuts over the last several years, which will pose challenges in meeting rapidly increasing needs in case detection, contact tracing, and laboratory testing. There are also challenges regarding the availability of resources in public health facilities to carry out essential functions. In May and June, an elevated number of people reported not being able to get medical care according to L2CU results. Since the outbreak, about 6-8 percent of respondents reported a member requiring medical treatment per month. Beginning in May and continuing in June, about 16 percent reported as being unsuccessful in obtaining treatment, though this estimate is based on a small absolute number of cases.
The population of Uzbekistan is also relatively dispersed (officially about half of the population lives in rural areas), simultaneously reducing some risk of transmission while also leading to high average travel times to local service providers (including clinics, hospitals, and pharmacies). As a proxy for local risk factors in the analysis that follows, those locations are identified that lack a local health clinic (within the mahalla) and/or local hospital, as well as the presence of a local pharmacy within the mahalla. In addition, measures of local density (apartments/families) are included at the mahalla to highlight risks specific to many people living in close proximity.

Migration
Remittance income is falling rapidly in Uzbekistan. In April 2020, the share of households receiving any remittances fell by 54 percent over the same period the previous year. Among those that did receive remittances, the value of the median transfer fell by 21 percent (in terms of inflation adjusted So'm). The share of households with members currently abroad fell by 22 percent in comparison the same period in 2019 (from 17 to 13 percent), and among those still abroad, active employment fell 18 percent in a single month (from 88 to about 73 percent of migrants). Future migration expectations have fully collapsed, as the number of respondents with household members considering seasonal migration fell by more than 95 percent over the previous year (Figures 8 and 9).

Source: Small Area Estimates from Listening to the Citizens of Uzbekistan baseline
Related previous analysis from the L2CU study (Seitz, 2019) found that remittances are very well targeted to depressed regions of the country, and transfers from abroad thus represent a crucial driver of poverty reduction in Uzbekistan. Findings show that weak local labor markets drive labor migration. Beginning to consider migration is associated with low life satisfaction, job loss, and unemployment. In contrast, actually migrating is associated with a remarkable improvement in labor market outcomes, alongside strong recovery in subjective and monetary measures of household welfare. The results further show that current migrants are more likely to send remittance payments when household members have deteriorating life satisfaction and/or subjective reports of worsening economic conditions at home.

Source: Listening to the Citizens of Uzbekistan Panel
That study estimated that in the absence of remittance income, the poverty rate in Uzbekistan (measured at $3.2 per day purchasing power parity) would have been expected to rise from 9.6 (at that time) to 16.8 percent, or to about 12.2 percent assuming (implausibly) that all current migrants were to find formal employment at the local prevailing median wage. In the current context and the near absence of the opportunity to migrate for employment abroad for an extended period, it is likely that many dimensions of wellbeing in areas that were already economically struggling pre-COVID-19 will face deteriorating conditions.

Dimensions of Monetary Wellbeing
As of this writing, the World Bank projects that the poverty rate is quickly increasing. The poverty rate likely rose to between 8.7 and 10 percent following the outbreak, compared to pre-COVID estimates of 7.4 percent -equivalent to between .45 and .88 million additional people in poverty. This is a remarkable reversal for a country that has seen sustained poverty reduction for decades. Government's official national definition 8 of the low-income population, the poverty rate in Uzbekistan fell from nearly 28 percent in 2000 to 11 percent in 2019, though the pace of progress has gradually slowed over time. Both official and L2CU-based consumption measures are consistent with subjective self-classifications of households believing they are "poor." 9

Figure 10: Average Per Capita Daily Consumption in 2011 $PPP for Uzbekistan (2018)
From Seitz (2019a) During lockdowns, household spending diverged between those with higher and lower incomes. In April, about 55 percent of respondents report significant changes in their household spending. Of the reported changes, about 60 percent report spending more than usual (split evenly between "moderately" and "substantially") over the preceding 30 days. About 40 percent report reducing their spending (23 percent moderately, and 77 percent substantially). Respondents with higher incomes were significantly more likely to report increased spending, compared to those with lower incomes, who report reduced spending on average. There were reports of shortages in April and May as well. About 16 percent of those who responded reported that items were out-of-stock in their local area. Of them, food items were most commonly cited by far (90 percent) and particularly flour. About 5 percent of those reporting out-of-stock items listed an inability to buy medicines, and 5 percent an inability to buy masks. However, shortages were large resolved by May, and by June a negligible number of households reported any remarkable local shortages of essential goods.
High shares of people reported that they were unable to afford basic needs and were without savings. Those reporting an inability to afford food rose from less than 9 percent to more than 12 percent of the population in April. Pre-COVID-19, most respondents already reported that they did not have any financial savings, and the measure spiked 21 percent in April (up from 59 to 71 percent). In April, the share of people who "completely agree" that the prices of bread and flour increased also spiked from 6 percent to 19 percent. A rising share reported that they were worse off financially than the previous month, from 2 to 10 percent, with similar expectations for the coming month. The share of respondents reporting that their financial situation "improved over the past month" fell by 60 percent, again with expectations for the next month falling by a similar amount.

Responding to the Crisis
In the absence of a "quick recovery," the COVID-19 health crisis is likely to be most severe for poor and vulnerable households, limiting their ability to abide by directives to contain the spread of disease. Labor market impacts in particular are expected to have knock-on effects on vulnerable households and are very likely to increase the prevalence and depth of poverty.
In this context, the President of Uzbekistan signed a US$1 billion economic relief plan to aid the economy and vulnerable population groups. The plan establishes the Anti-Crisis Fund and National Anti-Crisis Commission headed by the Prime-Minister. The Anti-Crisis Fund will finance COVID-19 prevention and control activities, social support to low-income families, support to strategic economic areas and small businesses. The plan also introduces time-limited tax rate reductions to support individuals and enterprises. As part of the relief plan, the Government announced salary top-ups for healthcare workers involved in the care of COVID-19 patients. Physicians can receive up to US$ 2,500 per month, nurses -up to US$ 1,500, and ancillary staff -up to US$ 500 per month.
The Government is expanding targeted social assistance programs to respond to the outbreak. Components of many of these will be implemented by local officials and there will be local variation in resource needs. Existing national cash allowances to low-income households currently cover (as of 2019) 249,341 families with children under the age of two, 411,422 families with children between the ages of 2 and 14, and 106,696 families received low-income allowances. However, due to cycling and re-application requirements, many of these families only received benefits for a six-month period, limiting the impact of such assistance. In March 2020, officials announced the expansion of social assistance programs to an additional 60 thousand families in response to the COVID-19 outbreak. In addition, as of April 3, 2020, the Government announced that they would waive the re-registration requirements for existing beneficiaries and automatically extend the payment of benefits to families with children, child care benefits and material assistance (currently slated to expire in March-June 2020) from six months to one year without the need for applying and submitting documents.
In addition, the authorities describe several locally administered initiatives to address the impacts of the crisis. The Centers for Coordination of Sponsorships that operate in all regions and in Tashkent City reported distributing food products worth 49.91 bln UZS, 3,201 drugs and medical items worth 137.9 bln UZS, among 413,072 families in need of social assistance. These benefits were provided according to lists compiled by over 5,200 volunteers together with the Mahalla chairs at citizen's assemblies. 10 The food packages distributed to the population included items mostly needed by the families (flour, potatoes, rice, onions, pasta, oil, sugar, carrots, eggs, meat products, poultry meet, etc.) With the aim of provide direct support to the population, the Cabinet of Ministers of Uzbekistan issued an order 11 on the provision of financial assistance from the proceeds of "Sakhovat va Kumak" Funds established under the Mahalla Public Charity Foundation and its regional branches. As of this writing, work is underway to provide an additional 380 billion UZS to support 400,000 needy families through a new initiative called the "Iron Book" system. Lists of 101,980 families to be supported by the regional departments of the Ministry have been developed including 49,961 poor families, 52,019 families that lost incomes during lockdowns, as well as 106,439 families with elderly people aged over 65 that are identified as in-need of social assistance. Starting 14 July, the reinstated Centers for Coordination of Sponsorships and Volunteering began distributing daily in-kind assistance to needy families in accordance with the lists put together by the Ministry for Support to Mahallas and Families.
Finally, a national hotline connecting to the regional call centers operated by support centers, as well as the hot line operated by the Ministry of Mahalla and Family Affairs, are in continuous operation. Authorities report that a total of 151,600 calls were received from 14 July to 12 August 2020.
Grievances are registered and forwarded to the sponsorship centers. As of this writing, the authorities report that a total of 317,282 families have received food packages after contacting call centers. 12 According to the Decree of the Cabinet of Ministers Nr 346 dated 29 July 2020 13 As provided in the Presidential Decree PD-6038 dated 30 July 2020 "On additional measures aimed at supporting population groups in need of social protection and assistance during Coronavirus pandemics"

II -Data
Data reported in this paper are combined from four primary sources: i) mahalla passport data (local administrative statistics), ii) baseline survey data from the L2CU study, iii) data from the monthly household panel survey in L2CU, and iv) regional price statistics from the Central Bank of Uzbekistan. Spatial details and classifications from the national statistical agency of Uzbekistan are also used.

Mahalla Passport Data
Mahalla officials are responsible for maintaining up-to-date administrative details on people registered to their local area, and regarding the programs they administer. These details include information on demographics, community infrastructure, local services, characteristics, and other data required for implementing social assistance programs. Regional authorities assemble these data on an annual basis within their jurisdictions. For this study, a subset 14 of these mahalla-level statistics were combined into a single national mahalla passport data file, covering the full universe of mahallas in Uzbekistan. Table  (2) lists the regions of Uzbekistan along with their number districts, mahallas, population, and families. Tables (3) and (4) present summary statistics of the key variables used in this study from the mahalla passport database. The data include information for 200 districts and cities that consist of 9120 mahallas, though in the interim, the final list of mahallas has been consolidated due to splits and mergers of mahallas. These passport data were treated as the most authoritative source at the mahalla level, and additional information available from other sources was linked to address missingness or dimensions of wellbeing that were not covered in the Mahalla passport database.

Listening to the Citizens of Uzbekistan
The survey component of the L2CU project is conducted by a private firm (Nazar Business and Technology, based in Tashkent) under the supervision of World Bank staff, the Development Strategy Center, the Center for Economic Research and Reforms, and in cooperation with government Ministries and the Statistical Agency. The study included a comprehensive baseline survey that can be matched at the mahalla level with passport data.
The L2CU survey design closely followed that of conventional Living Standards Measurement Study (LSMS) surveys and was conducted using a standard two-stage sampling design, in which 200 mahalla were randomly selected with probability proportionate to (population) size. The national sample was stratified by region and by urban areas. The data were re-weighted based on observed population totals within the each mahalla at the time of the survey fieldwork. The second stage procedure was conducted using simple random selection with equal probability within selected mahalla. A separate stratification level for households that receive social assistance was included, totaling 4 households per mahalla. The final sample included 4,000 households in total (20 households per mahalla), 800 of which were social protection recipients by design.
The baseline survey included a full consumption and expenditure module using a list/recall approach. The resulting estimates are representative for 12 regions, 1 autonomous republic, and 1 independent city (Tashkent), crossed with their urban areas (except for the City of Tashkent, which is entirely urban). The survey was conducted entirely on tablet devices (CAPI), enabling validation using crossreferencing and other techniques to ensure accuracy. The survey was conducted over the course of a 1.5-month period in May/June 2018.

Listening to the Citizens of Uzbekistan Panel
After completion of the face-to-face baseline, interviewers began regularly calling a randomly selected panel of 1,503 households over the phone to conduct short interviews, following a set monthly schedule agreed with the participating household. The questionnaire for these phone interviews was designed to monitor trends in migration, subjective well-being, measures of income, employment, service disruptions, and related indicators. Phone-based interviews began on September 5, 2018, and the first 22 rounds of the survey are used in the analysis that follows, covering the entire period to the end of June 2020. A total of 33,443 unique observations are available for analysis.
Attrition is one potential concern using panel data of this type. To ensure that non-take-up in the first round (and attrition in subsequent rounds) did not seriously affect the required sample size for survey representativeness, households that refused to participate were replaced with other households drawn from the same sample cluster. However, any systematic difference in the household characteristics due to refusal to participate could lead to bias if the replacement households were different on average (with respect to observable characteristics) from the households that refused. Among the random sample of 1,503 households originally drawn from the baseline, about 25% refused to participate in the first round (i.e. initial take-up in the first phone round totaled 1,122 randomly sampled households, and 381 randomly selected replacement households to make up for those that refused or could not be contacted). Comparing those who agreed and those who refused to participate shows that in general, relevant household characteristics such as total household consumption, migration status, and household size do not differ significantly between the two groups. The exception is that rural households are substantially less likely to drop out of the sample and require replacement. However, random selection of replacements from the same PSU results in near perfect balance when comparing to baseline summary statistics.
Attrition rates (or nonresponse rates) have tended to be low and stable across rounds of the L2CU panel survey, ranging from 1 to 3 percent, and about 66 percent of the sample completed every round between September 2018 and June 2020 (and many of those that missed one or more interviews were successfully re-contacted later in the panel). These results are particularly encouraging if compared to similar high-frequency surveys, in which attrition rates are generally higher. For instance, the World Bank project "Listening-to-LAC" registered attrition rates for Peru of around 67 percent for the first follow-up survey, increasing by about 3 percent with each round and reaching 75 percent in round six (Ballivian et al. 2015). Both the initial and final attrition rate for the Listening-to-LAC survey in Honduras was lower than for the survey in Peru (41 and 50 percent, respectively), but still consistently higher than for L2CU. Other examples World Bank high-frequency surveys in Africa have resulted in similar rates of attrition, or higher (Demombynes et al. 2013;Croke et al. 2012). However, a similar study in Tajikistan (Listening to Tajikistan) that began in 2015 met with similarly high rates of compliance.
To take non-take-up and attrition into account, the participating sample is reweighted by developing a model using observable and relatively time-invariant characteristics from the baseline to predict the probability of dropping out for each household. Responses are then weighted to account not only for the sampling design but are also reweighted in each round to partially account for any bias introduced due to households dropping out (if it is unaccounted for by randomly sampling replacement households from the same PSUs).

Regional Price Statistics
The Central Bank of Uzbekistan monitors regional price changes over time for a core basket of goods, including food and a small number of health supplies. These data are aggregated by group, and regions with the highest average price increases are identified. The resulting measure is included in the analysis that follows.

Derived Small Area Estimates
Many relevant indicators to identify important COVID-19 risk factors are not included in the mahalla passport database. This is due in part to measurement challenges (especially for indicators such as poverty rates, consumption, and rates of migration) but also due to the fact that the system was originally intended for other purposes aside from crisis response and recovery.
The data required to estimate poverty rates, average per capita consumption and other welfare indicators is traditionally collected conducted using surveys. To allow for frequent monitoring and to contain the costs of gathering detailed information, such surveys usually visit only a small sample of the population. When this sample of the population is representative, welfare surveys provide reliable estimates of poverty incidence for the entire population at a small fraction of the cost that would be required to survey each person in the country. However, this approach necessarily leads to sampling errors, and consequently, a typical household income or expenditure survey cannot produce statistically reliable welfare estimates for small geographic units. To address this issue, the approach adopted here begins with nationally representative survey data for measures of consumption per capita and other measures of interest from survey sources. The analysis then proceeds to sharpen the reliability of the survey estimates using small area estimation techniques to allow reporting at a level below what is traditionally reported (moving from oblast-level estimates, to either district or mahallalevel estimates). Using statistical models, these approaches provide estimates of indicators for small areas that would not be possible to reliably construct with traditional survey data alone. In such studies, the results are often used to target policies and assign resources to have greater poverty-reducing impact or are intended to address the concerns of specific welfare groups at the local level.
A variety of small area estimation methods have been devised to overcome the increasing imprecision of welfare estimates as they are disaggregated. The standard approach used by the World Bank to small area estimation, provided that the required data are available, is described in Elbers, Lanjouw, and Lanjouw (2003) and is often referred to as the "ELL" poverty mapping method. The assumptions and data employed for ELL maps are further elaborated upon in Bedi, Coudouel, and Simler (2007). However, a pre-requisite for using the ELL approach is access to micro-level census data, and no census has been conducted in Uzbekistan since independence in 1991. In such cases, a common alternative approach is the Fay-Herriot (FH) method (Fay & Herriot, 1979), which is adopted to generate the imputed estimates described in this report.
The FH method allows estimation of indicators and rates using a combination of survey data and mahalla-level indicators from available sources that are less subject to imprecision, such as administrative data or remote sensing. In this report, most of the publicly available sources used are administrative, while a small number are derived from satellite imagery. The FH approach proceeds by matching accurate area-based information with indicators that are aggregated to the level of interest in the survey (the mahalla, in this case). Starting from the relatively imprecise estimates from the survey, a statistical model is developed, which attempts to explain the variation of the welfare indicator at the mahalla level (in this case, focusing on average consumption per capita, per capita income, and rates of migration).
Once the model is estimated, the direct survey estimates also enter into the final area-level results: the final estimated area-level level of consumption is a weighted average of the observed and model-based estimates for cases in which both estimates are present. For areas that do not appear in the survey data (accounting for the large majority of cases as only 200 are directly observed out of more than 9000 mahallas), the results rely entirely on estimates derived from the statistical model.

Constructing a Summary Risk Index
The summary index described in this note follows the Alkire and Foster (2011) method to developing multidimensional measures of deprivation. Though usually conducted at the household level, in this case the index is calculated at the community (mahalla) level. There are several properties of this approach that are particularly useful in this case. In the context of the COVID-19 outbreak and recovery planning, officials and development partners will engage at several levels to support vulnerable communities. Many response initiatives will likely target particular at-risk populations, and in such cases, issue-specific information is critical. However, broad resource allocation decisions will also be required for recovery programs and anti-poverty initiatives. In this respect, a summary of the overlapping nature risk factors is also useful in understanding local needs. The Alkire-Foster approach is a strong option in such a situation as it is decomposable by indicator, and both by subgroups and dimensions. With this feature, users can thus make comparisons of need between mahallas, by demographic groups, or by vulnerability status. Moreover, dimensional decomposability allows comparisons within specific dimensions across subgroups of mahalla.
A list of the variables used to create the index is included in Table (5), alongside details of the weight placed on each indicator. The inclusion of indicators from the list available indicators in the database was conducted on the basis of literature review and the direct analysis of risk factors using the L2CU survey. In the first stage, the weights for each indicator and dimension were set to be first equal across dimensions, and then equally across indicators. In a second stage, views on the importance of each indicator (measured on a scale of 1-5) were collected through a stakeholder consultation survey, conducted online in Uzbekistan in June 2020. The summary statistics of the stakeholder survey are included in Annex (C). The indicator weights were then adjusted by the difference from the average value across all indicators for each indicator, such that indicators viewed on average as more important receive greater weight, and those viewed as less important receive less weight. Data on the importance of dimensions was also collected, however, these varied so little on average that the final index did left the weights of dimensions unchanged. In the results and following discussion, we defined a mahalla as being "high need" when it is in the top quintile of the index value, at "substantial need" in the next highest quintile, at "medium need" in the middle quintile, at "modest need" for the next highest quintile, and at "lowest need" in the bottom quintile. However, such an index is not intended to suggest that the lowest need mahalla will not require substantial official support in responding to the impacts of the COVID-19 pandemic. Rather, the index highlights the variation in risk factors to efficiently deploy limited resources to those areas that are expected to have the greatest (often overlapping) needs and the lowest resilience to these impacts.

IV -Small Area Estimate Results
Mahalla-level small area estimates of consumption were derived according to the procedure described in Section (III). This only derived indicator estimated directly in this analysis and used in the index (though additional small-area indicators that change over time are discussed in Section VI). Automated model construction using stepwise variable selection and including dummy variables for regions resulted in a model with an adjusted R 2 of .63, after excluding variables with measured variance inflation factors of 10 or more. Modeling of the mahalla-level average consumption was relatively successful, though for some individual mahallas the estimate can be relatively imprecise. The median coefficient of variation was about 10 (average about 11) with only about 2.7 percent of mahallas with coefficients of variation greater than 20 (a rough threshold used in some cases to assess sufficient precision). Out of 200 districts and urban areas, 129 had no mahalla for which the coefficients of variation greater than 20. However, imprecision was particularly concentrated in the city of Navoi, the urban settlement of Nurabad, and the town of Akkurgan, suggesting estimates for these should be treated with additional caution. Annex D includes additional diagnostics regarding precision. Rather than directly using the relatively imprecise average consumption estimated in this procedure, these estimates are grouped into quintiles, and only the lowest two quintiles of average consumption per capita are included as being "at risk" on this indicator included in the overall index, lessening the risk of overly relying on a single imprecise indicator. Small area estimation in this case highlight significant within district variation that would otherwise go unmeasured. At the district level, the mahalla average income per capita is about $6 in the bottom quintile, and 11 in the top quintile. In addition, the extreme concentration in the city of Tashkent of high per capita consumption at the mahalla level is striking, with only 2 percent of mahallas projected to fall below the top quintile on average. Figure (11) reports these estimates aggregated to the district level and weighted by population.

Figure 11: Map of Average Mahalla-level per capita Consumption
V -Index Results Figure (12) reports the summary index value aggregated to the district level and weighted by population. This analysis of the local risk factors highlights substantial spatial variability, both between and within larger territorial units of Uzbekistan. The results thus enable much more granular and targeted interventions than would be possible using solely the aggregated information available at the district or regional level. Further, the resulting values can be decomposed to suit a variety of purposes, and expressed in either a summary index, by dimension, or by individual indicator. The results further identify clusters of need within regions and at times across provincial borders.

Source: Authors' estimates
The results show that a large share of mahallas in the regions Syrdarya, Karakalpakstan, and Namangan face many overlapping risk factors. These regions have relative more mahallas with low estimated consumption per capita (pre-COVID-19), relatively less stable employment, higher levels of unemployment (pre-COVID-19), and much higher reliance on remittances. These regions have a high share of mahalla categized into the "highest needs" group in the overall summary index as a consequence ( Figure 13). At the other extreme, no mahalla in the city of Tashkent is identified in the "highest need" category, with most mahallas clustered at the bottom of the need scale. The region of Tashkent (which is a distinct administrative unit from the city) is also found to have a relatively small share of mahallas in the "highest need" category (Table 7 and Figure 14). This underscores that while several types of impacts are localized in the largest agglomeration in Uzbekistan, there are fewer overlapping risk factors there than in other areas (such as regional capitals, medium-sized cities, and those areas highly reliant on remittance income). This does not minimize the considerable direct impact of COVID-19 in urban areas, and particularly Tashkent, which is the densest location in the country, and has suffered the highest rates of infection at the time of this writing. Indeed, urban areas also face disproportionate economic impacts due to a greater share of employment in services sectors such as retail and transportation. However, the most prosperous urban areas of the country can also rely on many sources of resilience that are unavailable in less dense areas. Urban areas in Uzbekistan have more formal labor markets, and a higher share of (stable) government and stateowned enterprise-based employment. The densest urban areas also have greater access to health facilities, have modest numbers of vulnerable elderly people, faced low initial levels of poverty and unemployment pre-COVID, had a low reliance on social assistance, and send relatively few migrants. As a result, proxies of these factors measured in mahalla data lead to a lower ranking of need in the summary index.

Elderly and Disabled
The mahalla data reveal relatively high concentrations of elderly and disabled people in at least some mahallas in all regions of the country. Single seniors are found to be a larger share of the population in Karakalpakstan, Namangan, and Andijan. Karakalpakstan also struggles with the highest shares of the population who are registered as disabled. In terms of population shares, the city of Tashkent is the oldest regional unit, with 72 percent of people there living mahallas in the top two quintiles of that measure. Kashkadarya has an abnormally high number of people over the age of 100. Andijan and Jizzakh have relatively high rates of people who are disabled but did not receive disability benefits in 2019. Across all measures in this dimension, Karakalpakstan have the largest number of mahallas with overlapping risk factors overall, followed by the city of Tashkent, and the region of Andijan.
In contrast, mahallas in Surkhandarya, Khorezm, and Bukhara are found to have relatively fewer single seniors. Mahallas in Jizzakh and Samarkand have relatively few disabled people per capita in comparison to other regions, while Jizzakh and Surkhandarya have relatively few retirees overall.
Kashkadarya and Syrdarya have relatively few mahallas with many disabled people who lack financial support. Across all measures in this dimension, mahallas in Jizzakh, Samarkand, and Surkhandarya have the lowest share of overlapping risks ( Figure 15 and Table 8).

Economic Factors
The passport data show that mahallas in the regions of Tashkent and Bukhara have relatively high shares of entrepreneurs and trade workers (measured in two separate indicators). The regions of Fergana, Surkhandarya, Tashkent city and Navoi all have higher recorded "able bodied people not working." This result should be interpreted with caution, however, as it contrasts with official unemployment rates which find much lower joblessness in the city of Tashkent (and other urban areas) than in rural parts of the country. Rather than simply indicating higher deprivation, this indicator may partially reflect a larger share of the population enrolled in education programs in urban areas. Mahallas in Navoi, Samarkand, Bukhara, and Andijan all have higher than average shares of the population working in family businesses, which is assumed here to correlate strongly with informality in this analysis. Finally, mahallas in Karakalpakstan, Namangan and Khorezm all have higher shares of children than the national average, highlighting the difficulties expected for workers and others during the closure of schools, and related care responsibilities. Across all measures in this dimension, mahalla in Bukhara, Ferghana and Syrdarya had the most overlapping deprivations on average.
In contrast, mahallas in Tashkent region (excluding the city) and Karakalpakstan have relatively few concentrations of entrepreneurs according to the mahalla passport data. Trade sectors are a smaller share of workers on average in predominantly rural areas in Karakalpakstan, Kashkadarya, and Navoi. Mahallas in the region of Tashkent, Karakalpakstan, and Khorezm are able bodied but not working. Relatively few people in Tashkent region work in retail jobs, contrasting with the city of Tashkent. Fewer residents in the city or region of Tashkent work in family businesses, on average. Tashkent city and Navoi have relatively few young children, in comparison to other regions ( Figure 16 and Table  9).

Figure 16: Economic Factors Dimension
Source: Authors' calculations

Social Assistance
Social assistance (SA) provision is concentrated in several clusters in Uzbekistan. In particular, mahallas in Karakalpakstan and Jizzakh on average have many more assistance beneficiaries than mahallas in other regions. This relationship is clear across measures in the relevant indicators in the mahalla passport data, as both regions are above average with respect to loss of breadwinner benefits, other SA benefits, overall need of SA, and unmet need of SA. The region of Navoi, in contrast, has relatively low provision, but relatively high unmet need according to passport data. Across all measures in this dimension, mahalla in Karakalpakstan, Jizzakh, and to a lesser extent in Kashkadarya have a larger number of overlapping deprivations in this dimension. In contrast, the city of Tashkent has relatively few people eligible or receiving social assistance, which is also consistent with estimates of average per capita consumption and income. Mahallas in the region of Tashkent are somewhat more often among those with a large number of recipients of lost breadwinner allowances, but among other types of benefits there are similarly low levels as in the City of Tashkent. Across all measures in this dimension, mahallas in and around Tashkent are substantially less likely to be receiving (or identified as in need of) social assistance benefits, followed by mahalla in Bukhara and Ferghana ( Figure 17 and Table 10).

Services and Local Infrastructure
Very few mahalla in Uzbekistan have immediate (within mahalla) access to hospitals, clinics, pharmacies and other health facilities. Those that do, are concentrated largely in urban areas. Tashkent has high numbers of local hospitals, and fewer mahalla with local clinics. The region of Tashkent is an outlier with respect to a very low number of mahallas registered as having a local pharmacy. Across all measures in this dimension, mahalla in Karakalpakstan, Navoi and Namangan are more likely to have overlapping risk factors, while the mahallas located in the city of and region of Tashkent have disproportionately low overlapping risk factors across the indicators in this dimension ( Figure 18 and Table 11).

Migration
International out-migration is much more common in rural areas and highly associated with low levels of labor income. Mahallas in Khorezm and Namangan all send high numbers of migrants abroad. However, many districts and mahallas struggle to accurately record migration patterns, and survey estimates are relatively rough. An additional proxy indicator is therefore included in this dimension: having a large gender imbalance in the mahalla, as a large majority of out-migrants in Uzbekistan are young men. By this measure, Syrdarya has an abnormally high number of such mahallas. Across all measures in this dimension, mahallas in Syrdarya, Karakalpakstan, and Khorezm are most commonly identified as most reliant on migration and remittances with many overlapping concentrations on both indicators in this dimension. In contrast, Jizzakh, Navoi, and Tashkent region or Tashkent city have relatively few mahallas with concentrations of these indicators and cases on which they overlap ( Figure  19 and Table 12).

Dimensions of Monetary Poverty
The regions of Samarkand, Surkhandarya, and Syrdarya all have an above average number of districts with a poverty rate of over 10 percent. Karakalpakstan has a large number of mahallas at risk by this measure and is also an outlier with respect to mahalla in the bottom 40 percent of average consumption per capita. Food and medicine price increases were found to be highest in Jizzakh, Namangan, Navoi, Surkhandarya, Syrdarya and the Region of Tashkent. Across all measures in this dimension, mahallas in Surkhandarya and Syrdarya had a large share of mahallas with overlapping risk factors across all three. In contrast, the city of Tashkent, and the regions Kashkadarya and Khorezm had relatively few mahallas with overlapping risks in this dimension ( Figure 20 and Table 13).

VI -Dynamic Updates of Critical Indicators
The primary risk index described above is set using data that are collected infrequently, leading to challenges in updating responses in light of a rapidly changing situation. To address this, estimates from higher frequency sources of information can be linked with the database and used to impute small area estimates of critical measures over time. Users should bear in mind however that these estimates come with greater uncertainty than is often the case with official data. Nonetheless, patterns of the magnitude observed during lockdowns in April, and the gradual relaxation of these measures in May and June, are clearly discernable and provide much greater nuance to monitoring of nationallevel trends using the panel survey data.
Figure (21) reports the results of small area estimation performed at the level of the mahalla (and aggregated to the district level for the purposes of mapping). Strict lockdown began with the reinstatement of interregional police posts on March 23 to restrict the movement of cars. On March 25, Uzbekistan made mandatory the wearing of face masks in public. On March 27, the movement of people and personal vehicles was restricted to grocery shopping and pharmacy visits. The impact of these measures on reported employment was very large, with a decline of households reporting "any member working" falling by more than 40 percentage points in April ( Figure 21 -Panel (b)). During this time, employment fell dramatically throughout the country, however areas with more resilient labor markets (in particular those with a greater share of wage workers) saw milder declines than those with higher shares of self-employed workers. This is especially clear with respect to the region and city of Tashkent, where disruptions were severe, but less so than in areas with a larger share of people who were unable to work remotely and saw a full suspension of activity. In May and continuing through June, the labor market recovery is also clearly present, however, progress has proceeded unevenly across the country. As reported in figure (22), migration and remittance income declined rapidly following the outbreak. About 70 percent of labor migrants from Uzbekistan live and work in the Russian Federation, and a rapid decline in the value of the Russian Ruble in April substantially decreased the value of sent remittances, before the So'm weakened in parallel somewhat offsetting this effect. Since May, the ruble has been recovering against a USD benchmark, which means that the value of remittances has started to climb following April's large decline. But this is only relevant for those migrants who remain actively employed and are able to send money: lockdowns in Russia have also been severe, which disrupts the ability of workers to earn any income to send home, and the share sending any remittances is presently a much smaller share than is usually the case for Uzbekistan (see figure 9). In addition, travel restrictions mean that many fewer migrants have been able to leave for Russia (and other places) in comparison to last year. Remittance income is one of the most important drivers of poverty reduction in recent years (Seitz, 2019) and poorer, rural households much more commonly rely on remittance income. Small area estimates of per-capita income from all sources (including income from wages, remittances, pensions, agriculture, social assistance, and other sources) is reported in figure (23). The results highlight the link between income from work and per-capita income (by a large margin the most important component of total income in Uzbekistan, accounting for more than half even among the poorest quintile). With the disruption in employment caused by the pandemic, average incomes fell across the country, though some areas much more deeply than others.

Mahalla list provided by the National Statistical Office (NSO Data):
In addition to the mahalla passport data set, the second data set on mahalla characteristics was provided by the National statistical office (NSO). The main reason for using this data set is that it contains a detailed information of settlement type (urban or rural) of each mahalla. According to the NSO, there are three types of settlements: city (urban/shakhar), small city (urban/shakharcha), and village (rural/kishlak). Table 14 shows distribution of mahalla and population by the settlement types. While over 52% of the population are living in rural areas, 16% and 32% of population reside in regular cities and small cities respectively. It is important to note that in this data set, there are 190 districts consisting of 8933 mahallas. Although we are able to match the two data sets (the mahalla passport list and the NSO list) at district level, there are significant mismatch at mahalla level. Table 5 documents the differences between the two data set in terms of number of mahalla, population, and number of families. A fuzzy matching method based on mahalla names within a district was used to merge the two data set at mahalla level. With this method, we are able to match 8700 mahallas whereas 475 mahallas from the passport data and 255 mahallas from the NSO data are not matched given the information provided in the data set.

Annex B: Technical Description of the Fay-Herriot Model
The basic area-level model setup is as follows. Let be the true average consumption incidence in each mahalla i, and let the sampling model be defined by: where is the observed survey direct estimate of average consumption per capita , and is the sampling error associated with , such that |~(0, ) and are assumed to be known. The linking model is defined as: where denotes a vector of area characteristics, and are independent and identically distributed random errors with ( ) = 0 and ( ) = 2 . The data on are obtained from fully enumerated administrative sources and hence are free of sampling error. Combining the above sampling and linking models, it follows that the observed average consumption level from the survey can be modeled as follows: Given this setup, the best linear unbiased estimator of = + , one that minimizes the mean squared error (̃) = (̃− ) 2 is: where = ( − ), and = 2 + 2 is referred to as a "shrinkage factor". Given that 2 is unknown, the Best Linear Unbiased Predictor (BLUP) is replaced with its empirical counterpart EBLUP: ̂ =̃(̂2), which can be rewritten as: where ̃ is the Feasible GLS estimator for and ̂=̂2 + ̂2 . Thus, ̂ is a weighted average of the direct survey estimate and the synthetic (model-based) estimate ̃, and the weights are given by ̂. For with smaller sampling variances the shrinkage factor gives higher weight to the direct estimate, while for with higher sampling variances a higher weight is assigned to the synthetic estimate. In areas that are not part of the survey sample, the prediction is based on the synthetic estimate ̂, where ̂=̃(̂2 ). The prediction error associated with ̂ takes account of the sampling variance associated with , as well as the uncertainty associated with the estimate of and 2 (see Rao, 2003 for more details).
When calculating the term needed for the Fay-Herriot approach, there are several potential methods for considering the stratified and clustered two stage sample designs of the surveys used in this application. Common practice in the World Bank has been to obtain sampling variances associated with the area-level welfare measure by taking the variance estimate from the survey data source and dividing it by the sample size for each domain to obtain a set of "smoothed" sampling variance estimates. This ignores components of the clustered sample design; however, these smoothed sampling variances are commonly less volatile than alternatives. Another approach is to compute variance and the associated root mean square error of the mean using the linearized variance estimator approach-based on a first-order Taylor series (Wolter 2007). In sensitivity analyses this was the most stable variance measure, and the preferred approach for this application. Final results are quite similar when comparing the "smoothed" and "linearized" options described. For more detail on the tradeoff between approaches for domain variance estimation, see Heeringa et. al., (2017); Molina and Rao (2010);and Wolter (2007).
The results of the SAE estimates are presented graphically in the following section. The model variables that are part of the X vector in the estimation procedure were chosen to maximize the ratio of explained variance to the total variance, as captured by the adjusted R 2 . 15 There is no pre-set group of variables that are guaranteed to achieve that objective. Instead, automated variable selection using the stepwise approach was used.

Annex D: Small Area Estimate Model Diagnostics
Notes: the left-had graph presents coeffici3nts of variation among within-sample mahallas. The right-had graph describes the mahalla-level shrinkage factor. Highest Lowest